[{"content":"Built with: Python, Claude, Better Stack monitoring, Raycast and TypeScript. I wrote the pipeline and the operator console. I also helped move human review into Langfuse, a tool where people grade the AI\u0026rsquo;s replies; Mike, who runs Matchworks, and Liam built that with me.\nThe problem Matchworks is an outbound lead-generation agency, and it runs on email replies. A prospect answers a cold email, and the clock starts. On 28 July 2026 at 10:02 pm Mike wrote “we have analtyics showing \u0026lt; 15m response time conversion goes way up”, then “is it possible to get it super fast?” In July the median wait was 27.1 minutes, because every draft waited for me to read it and press send. Booking data was never wired into the system, so the result below shows speed, not conversion. I joined the two once, on 27 September, to test the 15-minute rule itself.\nVolume reached up to about 1,000 replies a week. One person reading every draft could not keep that under 15 minutes, so I built to the rule first and tested it later.\nWhat I built Planning began in February 2026, and from 29 July the system answered prospects on its own. It read the full thread, worked out how ready the prospect was, drafted a reply in the brand\u0026rsquo;s voice, and sent it on its own when its checks passed. When they didn\u0026rsquo;t, it held the draft with a written reason a person could act on in seconds.\nChecks before every send. The diagram below is every check, what it caught and when it blocked a reply. Every result was saved for every draft, so \u0026ldquo;why did this send?\u0026rdquo; took one lookup to answer.\nHow one reply goes out on its own Every check runs on every draft. A waived check still runs, and its result is written to the audit row. 01 Reply arrives A prospect answers a cold email. The thread joins the queue. 02 Claude drafts Reads the whole thread, rates how ready the prospect is, writes in the brand’s voice. 03 Checks guardbanned phrases, broken booking links, and the other brand’s link, sign-off or signatureon brandthe thread belongs to one brand onlyon patterna known reply patternon to 6 Aug, then waived and recorded; removed 20 Aug flagthe drafter’s own doubton to 6 Aug, then waived and recorded recipientthe draft’s note says wrong recipientalways, from 2 Oct holdthe AI marks it to hold, such as a declinealways, from 8 Oct Not a check: an empty draft never sends. 04 Send cap From 7 August, at most 100 sends an hour. Zero stops every send. 05 Sent and logged The reply goes out and its audit row is written. Any blocking check failing: the draft is held with its reason in writing. An operator reads the reason first and decides in seconds. Later: a person reviews the sent reply and grades it good, acceptable or wrong. The grade is saved on the same row. Audit row One auto-sent reply, sample data ProspectSam Sample · trailnotes.example BrandBrand A Sent3 Sep, 4:12 am Guardpass One brandpass Flagwaived: “asked twice” Reviewgood How one reply went out on its own, and its audit row. Sample data. Held drafts explained themselves. A draft the checks refused got a written reason: the wrong brand\u0026rsquo;s sign-off, a thread that crossed brands, or template phrases that needed a person\u0026rsquo;s eye. The reviewer read the reason first, so a review took seconds.\nTested before every change. The system sent email at 4:00 am with nobody watching. By its last change on 24 October, 170 automated tests covered the send path alone, every branch that decides whether a prospect gets one email, two or none, and they were how I changed it safely.\nA console for the operators. A Raycast console (Raycast is a Mac launcher, like a command bar) gave operators who aren\u0026rsquo;t engineers the reply queue, each thread\u0026rsquo;s history and bulk fixes, so they could run the system without me.\nThe Raycast console with the reply queue open. Names covered. Decisions Intent before drafting In February 2026 I proposed an intent step before every draft. The AI would offer two or three readings of what the prospect wants, each with a confidence score, and a person would pick one. The system I shipped in July dropped the person’s pick: it worked out each thread’s intent on its own, then drafted the reply.\nSend first, review after On 28 July I was the bottleneck, because I still read every draft before it went out. Weeks of my feedback had already shaped the drafts, so I switched to sending first and grading what went out afterwards, and auto-send went on the next morning. The risk was real. A reply nobody had read could reach a prospect, and I turned down a few days of test runs against hand-sent replies so Mike would see the number that week. Three limits held it. The checks ran on every draft, every refused draft waited in a held queue for a person with its reason in writing, and I graded what went out. From 7 August a cap of 100 sends an hour also stopped a loop or a bad update from sending in bulk. In the first week, 29 July to 4 August, at least 98 replies went out on their own (the send log missed some on the first two days). I read 31 of them one at a time and marked 1 wrong, a referral reply that went only to the original sender. The other 67 were never read one at a time.\nBlock on our mistakes, not on the AI\u0026rsquo;s opinion All four launch checks blocked a reply at first. By 6 August 247 drafts had waited for a person, and 173 of them were held by the AI\u0026rsquo;s flag, which is an opinion about the prospect, not a fault in our reply. So I kept the guard and the one-brand check blocking, because they check our own output, and let the pattern and flag checks run without holding a reply. Of the two holds added in October, the wrong-person hold came straight from three replies sent to the wrong person on 1 and 2 October.\nBuilt fast, to gather data I built it fast to gather data for the next version Mike would build: which replies could go out alone, which needed a person, and why.\nThe AI drafts, a separate program sends The AI wrote drafts and had no way to send them. A separate program re-read each draft as it sat in the inbox and ran the checks on that text, not on the AI\u0026rsquo;s own copy, so what got checked was exactly what went out.\nA strict field for the pattern check, and two-brand threads wait The old pattern field was free text that nothing checked. Gating on it would have auto-sent anywhere from 35% to 77% of drafts, depending on how the AI happened to phrase it. I added a field that takes one name from a fixed list instead. Any thread that touched both brands always waited for a person, because a reply that crosses brands is the one mistake nobody can take back.\nTwo replies, side by side Here are two replies from August 2026, one sent on its own on 4 August and one held for a person on 19 August, each followed from the prospect\u0026rsquo;s email to my grade. The names, brands and domains on all 6 screens are made up; the times and check results come from the 8,820-row send log and the review journal.\n1. The prospect writes back On Tuesday 4 August 2026 at 4:37 am a prospect asked how we found her and who the partners were before she would book. On Wednesday 19 August at 3:07 am another said he was the right person and open to a talk.\nTwo prospect replies, one per brand. Recreated, original data private. 2. Intent read, then the draft The drafting AI read each thread and named what the prospect wanted. The first was a request for partner details, so the draft answered by category and kept names for the call. The second was a plain first reply, so it got the house template.\nWhat the AI read in each thread and the draft it posted. Recreated, original data private. 3. Every check, on the draft as it sat in the inbox On 4 August all four checks could still block, and all four passed. On 19 August only the guard and the one-brand check could block, and the one-brand check failed: the thread looked like it touched both brands.\nThe checks each draft ran, and which ones could block that day. Recreated, original data private. 4. Sent, or held with its reason The first reply went out at 4:43 am, 6.6 minutes after the email arrived, with nobody reading it first. The second was held at 3:11 am, and the Raycast menu bar showed the reason, \u0026ldquo;cross-brand thread\u0026rdquo;, above the send button.\nLeft, the send. Right, the held draft in the Raycast menu bar with its reason. Recreated, original data private. 5. The audit row The first send left a row at 4:43 am with its arrival time, send time, sender and all four check results. The held reply was sent from the menu bar at 3:53 am, 46.5 minutes after it arrived, word for word, and its row marks it as sent by a person.\nThe audit row each reply left. Recreated, original data private. 6. Graded after sending I graded the first reply good at 4:48 am, five minutes after it went out, with a note. I graded the second acceptable on 21 August. On 19 and 20 August I pulled all 40 drafts held as \u0026ldquo;cross-brand thread\u0026rdquo;: none carried two brands, so I split the reason in two, one for an address the system could not read and one for a real two-brand thread.\nThe grade on each reply, and what the held one changed. Recreated, original data private. The result Median wait from a prospect\u0026rsquo;s reply to ours: 27.1 minutes in July 2026, 5.6 minutes in October. The system\u0026rsquo;s own replies in October took 4.5 minutes (2,027 that month). Most of the drop came from replies no longer waiting for a person to read them before sending. 8,820 replies sent from 22 July to 2 November: 6,968 went out alone, and 1,852 waited for a person. How much people changed those 1,852 drafts is mostly unknown: 1,625 were sent by hand inside the inbox, and their text was never saved. Where both the system\u0026rsquo;s draft and the sent text survive, for 71 replies, 70 went out word for word and 1 was rewritten. I compared the two texts with spacing ignored. I read and graded 324 sent replies one at a time, all between 3 and 25 August 2026, and marked 6 wrong. Five have a note: a referral reply that went only to the original sender, a reply to a lead already disqualified, one that offered the prospect an option it shouldn\u0026rsquo;t have, a malformed booking link, and one that pivoted too confidently when the prospect named another company. One has no note. The review journal\u0026rsquo;s other 2,689 grades were whole-list approvals, one click that cleared a list without reading it, so they are not counted here. After 25 August nothing measured quality reply by reply. The only quality record after that date is three replies addressed to the wrong person, sent on 1 and 2 October. They are why, from 2 October, a draft addressed to the wrong person was always held. Every auto-send carried a full audit row. From 27 October 2026 Mike built the next version in Langfuse on the data this one gathered, the review handbook I wrote on 23 October and my review notes. How it is measured: the send log records when each prospect\u0026rsquo;s reply arrived and when ours went out, and marks who sent it. The median counts every reply from every sender, by calendar month in Philippine time; the 4.5 counts only the system\u0026rsquo;s own.\nWhat the records said The thread read, June 2026 On 11 June 2026 I pulled every reply thread with a send since 27 May, 4,000+ of them, and read them before writing the reply playbook. About 2,900 were real exchanges once sequence emails and warmup sends were removed. One line, telling prospects we had spoken with possible buyers the day before, went out 250+ times in a week, and two groups of replies walked it back when prospects challenged it. At least ten groups of prospects said the booking form showed no times.\nTesting the 15-minute rule, September 2026 On 27 September 2026 I tested Mike\u0026rsquo;s rule that every lead gets an answer inside 15 minutes. I chose what to test and read the result; Claude agents did the maths. The test matched about 3,600 first answers, from 21 July to 20 September, to booking form submissions within 7 days.\nAnswers under 15 minutes beat slower ones on submissions by 1.9 points, with a 95% margin of 0.0 to 3.6. That margin touches zero, so speed inside 15 minutes is unproven. The loss sat in answers over 6 hours: 7.0 points worse on submissions than answers under 5 minutes. Under 5 minutes never beat 5 to 15 minutes. A follow-up on 3 October found a fair test would need 21 leads a day for 90 days to see a 3.6-point lift. The 27 September records also showed where the slow answers came from. The sender was down 32% of the 67 days before 27 September. Since 20 September, 37.1 of its 40.8 down-hours were the Mac asleep. A lead-minute is one lead waiting one minute: with the Mac awake, the 90-second check for new replies cost about 1,200 a week, and with the real sleep and outages about 11,800. So the advice from 27 September was to keep the Mac awake or move the sender off it, and to stop chasing speed under 15 minutes in favour of the answers that waited over an hour.\nWhat broke, and the fix A lead here is one prospect\u0026rsquo;s reply: one run of the drafting AI (Claude Opus 4.8) that reads the thread and writes a draft. A step is one action in that run, such as reading or searching a file. The run was a loop of steps because it copied the path I used by hand: read the thread, check the business fits, draft against the rules, then post the draft. The first version cost too much per lead and kept hitting the provider\u0026rsquo;s usage limit mid-shift. I measured where the cost came from before cutting it. It took 17.8 steps because nothing was shared between leads: every run reloaded the full drafting instructions and searched a 480 KB file of past reply decisions three to six times before it could write. I put the rules one draft needs into one file that every run shared, moved the old decisions out of the drafting path, then re-measured on 23 August, 489 runs against the 603 before. Steps per lead fell by almost two thirds and AI cost per lead by almost half:\nAI cost per lead\nBefore, 603 runs $1.72 After, 489 runs $0.92 Steps per lead\nBefore, 603 runs 17.8 After, 489 runs 6.4 AI cost and steps per lead before and after the rebuild, re-measured 23 August 2026 What I\u0026rsquo;d change It was my first time building a system like this. I understood the theory, and now I have the firsthand experience. Next time I\u0026rsquo;d start on proper infrastructure like Langfuse. I\u0026rsquo;d build a test set from the hand-sent replies and run it before turning on auto-send, the test I skipped so Mike would see the number that week. And I\u0026rsquo;d grade every cost test. In October I tested a cheaper setup, where code gathered the thread first and the AI wrote the reply in one prompt; over about 1,080 runs on each side it cost a third less per lead than the loop. Its replies were never graded before Mike\u0026rsquo;s version took over, so I can\u0026rsquo;t say it was as good.\n","date":"25 August 2026","permalink":"https://ccabaluna.pages.dev/work/autonomous-reply-system/","section":"Work","summary":"\u003cp\u003e\u003cstrong\u003eBuilt with:\u003c/strong\u003e Python, Claude, Better Stack monitoring, Raycast and TypeScript. I wrote the pipeline and the operator console. I also helped move human review into Langfuse, a tool where people grade the AI\u0026rsquo;s replies; Mike, who runs Matchworks, and Liam built that with me.\u003c/p\u003e\n\u003ch2 id=\"the-problem\" class=\"relative group\"\u003eThe problem \u003c/h2\u003e\u003cp\u003eMatchworks is an outbound lead-generation agency, and it runs on email replies. A prospect answers a cold email, and the clock starts. On 28 July 2026 at 10:02 pm Mike wrote “we have analtyics showing \u0026lt; 15m response time conversion goes way up”, then “is it possible to get it super fast?” In July the median wait was 27.1 minutes, because every draft waited for me to read it and press send. Booking data was never wired into the system, so the result below shows speed, not conversion. I joined the two once, on 27 September, to test the 15-minute rule itself.\u003c/p\u003e","title":"An inbox that answered itself"},{"content":"","date":"1 January 0001","permalink":"https://ccabaluna.pages.dev/roles/executive-assistant/","section":"Roles","summary":"","title":"Executive Assistant"},{"content":"The problem In October 2024 the CEO I had worked for since 2021 started Genius, a recruiting startup that placed offshore talent with growing companies, and handed me its operations. A search runs on many small steps: intake, sourcing, screening, interviews, the shortlist, client updates and the invoice. With 5–6 open searches at a time, every step that waited on a person slowed a search down.\nFrom January to May 2025 I trained and led a small recruiting team on the systems I built, with a senior recruiter and a second recruiter running calls. From June 2025 I ran it alone. In October 2025 the CEO moved me onto his AI content work, and my Genius work ended that month.\nWhat I built The map above is the whole business. I kept the calls, the interviews and the endorsements, because those were what a client paid my judgment for, and I moved the other steps to tools.\nAn Airtable ATS (applicant tracking system) as the back end: 4 core tables (clients, projects, candidates and interviews) with about 100 fields between them. Notion as the client side: a Typeform intake created each lead in Notion, and each shortlist went to the client as Notion candidate profiles. Relay between them: when the Notion status moved to campaign start, Relay created the client and project in Airtable, with a project code like G-1008, so I stopped copying records by hand. Discovery to profile: Fathom recorded each discovery call, and the transcript became a Dream Candidate Profile, the document that described the hire to candidates. Sourcing: Workable job ads, LinkedIn Recruiter, Juicebox exports and candidate email campaigns in Smartlead. Screening: Claude prompts in Claude Projects scored each applicant against the profile and drafted my screening notes and replies, next to an AI Screen field in Airtable. Training and reporting: I onboarded each recruiter on the CRM I had built in Notion. From 1 August 2025 I sent the CEO a weekly report on the key numbers: leads, calls booked and held, and active searches. The Airtable ATS: candidates linked to clients, projects and interviews. Recreated with made-up names, original data private Candidate Lineup deck for one Financial Analyst search: 34,025 impressions and 330 applications. Client name made up. Recreated, original data private Decisions Airtable for the back end, Notion for clients I moved the ATS from Workable to Airtable, because Airtable was cheaper and easier to customize for our needs. Clients kept seeing Notion pages, because a Notion profile reads like a finished document and an Airtable record reads like a database row.\nRelay over Zapier, Make and n8n I picked Relay over Zapier, Make and n8n because its plan included AI credits with several models at a fair price, and its integrations were solid. Running n8n with OpenRouter would have meant more parts to maintain, and Relay cost the same as Zapier. The trade-off was Relay\u0026rsquo;s cap of 750 actions a month, which ruled it out for bulk candidate emails, so in August 2025 I moved those to Smartlead campaigns.\nA first screen by hand before the AI AI screening ran after a quick manual screen, so the AI didn\u0026rsquo;t spend effort scoring every applicant. The AI drafted notes; the interview and the endorsement stayed with a person.\nAsking more on the discovery call In July 2025 I chose to ask clients more up front on the discovery call and accept the extra friction, because earlier searches had ended in replacements and misaligned hires.\nOne search, start to finish I follow one Financial Analyst search, from how a client came in to the 2 hires on 12 September 2025, then one ordinary day from my log, on screens recreated with made-up names.\n1. A new client fills in the intake form From April 2025 every new client came in through a Typeform form: the role, the skills, the pay range, the start date and who decides. Each answer became a new lead in Notion.\nThe intake form, with its real questions and made-up answers. Recreated, original data private 2. Campaign start creates the Airtable records When I moved a lead to Campaign Start in Notion, Relay created the client and the project in Airtable, with a project code, and set the lead to Sourcing \u0026amp; Vetting.\nA Notion lead moved to Campaign Start, and the Relay run that created its Airtable client and project. Recreated, original data private 3. The Dream Candidate Profile This client, an investment firm, first hired a Financial Analyst through us in January 2025. I had written the profile on 30 December 2024 from the client\u0026rsquo;s own answers: board-level reports, M\u0026amp;A experience, fluent Portuguese. When the client came back in late August 2025, I reused it.\nThe Dream Candidate Profile for the search, shortened. Recreated, original data private 4. The job ad goes up On 27 August 2025 my note for the next day said to post the listing for this client quickly, because they move fast and pay fast. The ad came straight out of the profile.\nThe job ad, written from the profile. Recreated, original data private 5. Screening I screened the applications by hand first. Then a Claude prompt scored each one against the profile\u0026rsquo;s must-haves and wrote the result into the AI Screen field in Airtable. 44 applicants were invited to record a video and 37 did.\nOne candidate\u0026#39;s AI Screen in Airtable: each must-have marked pass or fail, with the evidence. Recreated, original data private 6. The lineup goes to the client I held calls with the top 20 and picked 5 finalists. On 6 September 2025 I built the Candidate Lineup deck: 34,025 impressions, 330 applications, 44 invited to video, 37 videos, 20 calls, 5 finalists. The client interviewed them the week of 8 September.\nThe finalists slide of the Candidate Lineup. Names and details made up. Recreated, original data private 7. Two hires, and the weekly report On Friday 12 September 2025 the client hired 2 of the 5. I wrote both onboarding checklists and sent the invoice for 2 placements that day. My weekly report to the CEO counted 10 leads, 5 calls booked, 3 held and 6 open searches.\nMy weekly report for 8 to 12 September 2025, money lines left out. Recreated, original data private 8. One ordinary day Most days were five searches at once. On Tuesday 24 June 2025 I chased one client by WhatsApp, SMS and a call, booked candidate calls for another, held CFO candidate calls, sent a tax client the last candidate, and waited on picks from a lineup.\nMy daily work log for 24 June 2025, shortened, with made-up client names. Recreated, original data private The result One Brand Designer search, 3 June to 31 July 2025: 912 applied, 179 assessed, 9 interviewed, 3 endorsed and 2 client interviews. One Financial Analyst campaign reached 34,025 impressions and brought in 330 applications. 44 candidates were invited to record a video, 37 did, 20 got a call and 5 went to the client as finalists. It was a repeat client: it first hired a Financial Analyst through Genius in January 2025, hired 2 of these 5 on 12 September 2025 and an M\u0026amp;A Analyst in October 2025, then opened two more searches, in November 2025 and January–February 2026. On one CFO search, by my journal on 21 July 2025, 238 candidates emailed gave 16 replies and 4 qualified candidates. 71 inbound client requests came through the intake form from 10 April to 9 July 2025 (intake form export). The pipeline I reported weekly. Every weekly report to the CEO in the records, all from 2025, with its own counts. Booked and held are sales calls, clients are new clients; searches are the searches open that week. No report survives for 11–15 August, 25 August to 5 September or 15–19 September.\nWeek Leads Booked Held Clients Searches 28 Jul–1 Aug 8 6 4 0 1 4–8 Aug 12 8 4 1 6 18–22 Aug 6 5 3 1 5 8–12 Sep 10 5 3 1 6 22–26 Sep 4 4 1 0 – The 4–8 August client signed after a call from an earlier week, and the 8–12 September client committed but had not yet signed. The 22–26 September report gives no search count.\nWhat I\u0026rsquo;d change I\u0026rsquo;d start on Airtable from the first search and never pay for Workable.\n","date":"1 October 2025","permalink":"https://ccabaluna.pages.dev/work/genius-recruiting/","section":"Work","summary":"\u003ch2 id=\"the-problem\" class=\"relative group\"\u003eThe problem \u003c/h2\u003e\u003cp\u003eIn October 2024 the CEO I had worked for since 2021 started Genius, a recruiting startup that placed offshore talent with growing companies, and handed me its operations. A search runs on many small steps: intake, sourcing, screening, interviews, the shortlist, client updates and the invoice. With 5–6 open searches at a time, every step that waited on a person slowed a search down.\u003c/p\u003e\n\u003cp\u003eFrom January to May 2025 I trained and led a small recruiting team on the systems I built, with a senior recruiter and a second recruiter running calls. From June 2025 I ran it alone. In October 2025 the CEO moved me onto his AI content work, and my Genius work ended that month.\u003c/p\u003e","title":"Running a recruiting startup on systems I built"},{"content":"","date":"1 January 0001","permalink":"https://ccabaluna.pages.dev/roles/gtm-outbound/","section":"Roles","summary":"","title":"Go-to-market and outbound ops"},{"content":"The problem Novum, an SEO agency in Cyprus, sold backlinks: links from other websites that push a client up in Google. Every link was won by cold email. I joined in April 2021 as a link builder. My sending grew in steps: up to 150 emails a day per campaign from Pitchbox in 2021, about 300 a day from Smartlead by March 2023, and from June 2023 15 Office 365 mailboxes at up to 90 a day each. The cold email page has the full timeline.\nWhat I built A ledger of every domain we pitched, in Google Sheets. I started it on 3 March 2022, loaded our earlier Pitchbox history into it and logged about 11,500 domains that year, so the team stopped pitching the same site twice. In March 2023 I added an Apps Script email finder. Guest posts and SEO content: the team pitched and placed guest posts on high-authority sites. 22 of them are still live with my byline or author bio, from ISACA in October 2021 to easy.tools in February 2026. Prospect CRM in Google Sheets. Recreated, original data private. The projects, one page each How a guest post got placed, from site list to live link (2021–2024) A site-scoring tool that cut a day of lead research to 2–3 hours (Sep 2024) Running the agency\u0026rsquo;s cold email, 2021 to 2024 (2021–2024) Winning SEO clients with cold email and free audits (2022–2024) Running the client side of an SEO agency (2021–2025) 1,339 links for a reseller agency, counted and billed every month (2022–2025) Running the CEO\u0026rsquo;s LinkedIn: 40 posts in 8 months (2023–2024) Statistics posts that earned links without outreach (2024) Three roles, one example each One trace per job: a guest post I placed as a link builder, a week on the team board as Head of Link Building, and one day as Operations Manager. Steps that belong to another project link to its page.\nLink builder, 2021: one placement from pitch to byline 1. Tools and the first pitches I opened a Hunter account on 9 June 2021 to find editors\u0026rsquo; emails, the first tool in my own name. On 14 June, in the early hours of 15 June in Manila, I sent my first guest-post pitches from the agency template, one about every three minutes.\nMy first week of pitching, June 2021. Recreated, original data private. 2. The pitch to ISACA On 13 August 2021 I sent ISACA the same template: a guest post from me, with a link back to them in exchange. Their communications team sent guidelines and an author agreement the same day.\nThe pitch that started the ISACA placement, 13 August 2021. Recreated, original data private. 3. Topics, a rejection, a yes My first three topics were too basic for ISACA\u0026rsquo;s security readers, and the editor said so on 16 August. I sent three new ones two days later, and the ransomware topic was accepted on 24 August. I delivered the draft on 19 September and made their edits in October.\nTwo and a half months of the ISACA thread, condensed. Recreated, original data private. 4. Live under my name The post went up on 27 October 2021, and ISACA featured it in its weekly newsletter the next day.\nThe live post on isaca.org, with my byline, published 27 October 2021. Head of Link Building, 2022: the board and the team 5. The tracker Every placement was a card on our Trello \u0026ldquo;Guest Posts\u0026rdquo; board. A card moved from In Progress to Proofreading \u0026amp; Editing, then Done, then Submitted to the editor, so anyone could see where each post stood.\nThe Guest Posts board and one card\u0026#39;s moves in July 2022. Recreated, original data private. 6. A week of deliverables Between 2 and 4 June 2022 IG Rosales finished editing five guest posts, one of them for Smith.ai, and tagged me and Burkhard on each card. Freelance writers drafted them.\nFive posts cleared editing in three days, 2 to 4 June 2022. Recreated, original data private. 7. The team grows On 5 July 2022 four new link builders joined the board on my invite, Renz among them.\nTrello\u0026#39;s notice that the new link builders had joined, 5 July 2022. Recreated, original data private. Operations Manager, 2023: one day 8. Thursday, 30 March 2023 The day started at 5:59 am with a client\u0026rsquo;s invoice question and ended after midnight with a proposal review. In between came a prospect\u0026rsquo;s content sheet, Sai\u0026rsquo;s update to one of my audit docs and a reseller order thread.\nMy inbox on 30 March 2023, in Philippine time. Recreated, original data private. 9. Moving the mailboxes That evening I signed the agency up for Smartlead at 8:43 pm and connected about 12 sending mailboxes by bulk upload. The test emails came in between 9:03 pm and 10:25 pm. The cold email page tells why we moved.\nSmartlead\u0026#39;s test emails from our sending domains, 30 March 2023. Recreated, original data private. 10. A proposal before bed At 10:27 pm Burkhard left one note on proposal #140, which I had built for a prospect: \u0026ldquo;make a bit smaller\u0026rdquo;. He resolved it at 12:37 am. The client acquisition page covers the audits and proposals.\nBurkhard\u0026#39;s review comment on proposal #140, 30 March 2023. Recreated, original data private. The result I was promoted to Head of Link Building in January 2022, about 9 months in, and managed the research and outreach team. I took placements from 30 to 100+ a month on my own in my first year (2021–22), by my own notes. That took the agency’s link-building revenue from about $15,000 to $30,000 a month, also by my own notes (June 2025). The agency\u0026rsquo;s tracker credits me with about 129 live posts, many carrying several client links. Once we hired a team, it averaged 150+ a month while I moved to operations and client acquisition. The client outreach sent 19,081 cold emails in 11 months and led to about 46 SEO proposals. The team delivered 150+ guest posts on sites with a Domain Rating (DR) of 50–90, Ahrefs’ 0–100 score of a site’s authority, and 1,000+ backlinks for clients. The CEO\u0026rsquo;s LinkedIn grew from a near-empty account to about 12,000 followers between October 2023 and June 2024, by my own count, bringing in about 5 qualified inbound leads a month. One carousel reached 300,000 impressions. What I\u0026rsquo;d change Most of what I know came from failing and testing, so I couldn\u0026rsquo;t have skipped those years. If I could keep that knowledge and start again, I\u0026rsquo;d build the email infrastructure first. The whole business ran on cold email, and we got burned more than once: when our sending broke, revenue slowed down with it. Good setup knowledge was hard to find back then, so we learned it the hard way.\n","date":"1 October 2024","permalink":"https://ccabaluna.pages.dev/work/novum-link-building/","section":"Work","summary":"\u003ch2 id=\"the-problem\" class=\"relative group\"\u003eThe problem \u003c/h2\u003e\u003cp\u003eNovum, an SEO agency in Cyprus, sold backlinks: links from other websites that push a client up in Google. Every link was won by cold email. I joined in April 2021 as a link builder. My sending grew in steps: up to 150 emails a day per campaign from Pitchbox in 2021, about 300 a day from Smartlead by March 2023, and from June 2023 15 Office 365 mailboxes at up to 90 a day each. The \n      \n    \u003ca href=\"/work/novum-cold-email/\"\u003ecold email page\u003c/a\u003e has the full timeline.\u003c/p\u003e","title":"From link builder to running the agency’s operations"},{"content":"","date":"1 January 0001","permalink":"https://ccabaluna.pages.dev/roles/operations-lead/","section":"Roles","summary":"","title":"Operations lead"},{"content":"The problem From November 2018 I answered US loan customers over chat and email for ECE Consulting Group, on Zendesk and Zoho. In May 2019, six months in, I was promoted to Team Lead. My team of 13 ranked 15th of 30 teams on the company scorecard, which counted attendance, QA, CSAT and other measures. Shifts across the company went by seniority, and staff saw that as unfair.\nWhat I built Real-time queue work, scheduling and forecasting: I watched queue levels live and adjusted the schedule to the forecast. Coaching from QA: I coached each agent from their QA reviews, and ran the team huddles. A coaching dashboard in Google Data Studio (now Looker Studio). Merit-based shift bidding for the company\u0026rsquo;s 200+ staff: staff picked shifts in Google Forms, and Google Sheets ranked each request by performance. Decisions Merit over seniority for shifts Seniority decided shifts before, which staff saw as unfair, and it took a lot of admin to run. I built the bidding so performance decided who got first pick.\nOne shift as Team Lead This is one shift in 2019, told from memory: nothing from 2018 to 2020 survives, so every screen is recreated and every grey figure is sample data.\n1. The queue at shift start I opened the Zendesk views before anything else: chats waiting, the oldest wait, the email backlog and any escalation sent to me. Escalations went first, then whichever view was growing fastest. This is my own account of a 2019 shift.\nThe support queue at the start of a shift, with sample counts. Recreated, original data private 2. Adjusting the schedule to the forecast Next I set the forecast against the 13 agents on my team, hour by hour. Where the peak hour came short of agents, I moved breaks out of it and pulled agents from email onto chat in Zendesk, then sent them back to email and coaching when chats slowed.\nThe team schedule against the forecast, with the changes for the peak. Sample data. Recreated, original data private 3. A QA review used for coaching I read QA reviews of my agents\u0026rsquo; chats and picked one habit to coach per agent. In the Zendesk chats I reviewed, the common miss was quoting the policy before acknowledging the customer\u0026rsquo;s problem. Over three months the team\u0026rsquo;s QA score went from 75% to 85%.\nA QA review of one chat, with the coaching note it led to. Agent, customer and scores made up. Recreated, original data private 4. The huddle Before the peak I ran a short huddle: where the team stood against the other 29, one habit for the day, the peak plan and a shout-out. We started at 15th of 30.\nA huddle agenda before the peak. Sample agenda. Recreated, original data private 5. The coaching dashboard I built a dashboard in Data Studio so each agent\u0026rsquo;s QA and CSAT sat next to what I was coaching them on. In three months the team went from 15th of 30 to the top 5, and it held a CSAT of 4.8 out of 5.\nThe coaching dashboard. Bold figures are the page\u0026#39;s results; the rest is sample data. Recreated, original data private 6. The shift-bid form Shifts used to go by seniority, which staff saw as unfair. I built a Google Form where all 200+ staff bid for their next shifts, a first and a second choice.\nThe shift-bid form. Sample form. Recreated, original data private 7. The sheet that ranks the bids The form fed a Google Sheet that ranked every request by performance. The best performers picked first, and when a shift filled, the next request fell to its second choice.\nShift bids ranked by performance, with sample scores and weights. Recreated, original data private The result My team went from 15th of 30 to the top 5 in three months. QA scores rose from 75% to 85%. The team held a 4.8/5 CSAT. No scorecard from those years survives, so these numbers are my own account.\nWhat I\u0026rsquo;d change I\u0026rsquo;d write the huddle and coaching routine down, so the next Team Lead could run it without me.\n","date":"1 July 2020","permalink":"https://ccabaluna.pages.dev/work/ece-support-team-lead/","section":"Work","summary":"\u003ch2 id=\"the-problem\" class=\"relative group\"\u003eThe problem \u003c/h2\u003e\u003cp\u003eFrom November 2018 I answered US loan customers over chat and email for ECE Consulting Group, on Zendesk and Zoho. In May 2019, six months in, I was promoted to Team Lead. My team of 13 ranked 15th of 30 teams on the company scorecard, which counted attendance, QA, CSAT and other measures. Shifts across the company went by seniority, and staff saw that as unfair.\u003c/p\u003e\n\u003ch2 id=\"what-i-built\" class=\"relative group\"\u003eWhat I built \u003c/h2\u003e\u003cul\u003e\n\u003cli\u003e\u003cstrong\u003eReal-time queue work, scheduling and forecasting:\u003c/strong\u003e I watched queue levels live and adjusted the schedule to the forecast.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eCoaching from QA:\u003c/strong\u003e I coached each agent from their QA reviews, and ran the team huddles.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eA coaching dashboard\u003c/strong\u003e in Google Data Studio (now Looker Studio).\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eMerit-based shift bidding\u003c/strong\u003e for the company\u0026rsquo;s 200+ staff: staff picked shifts in Google Forms, and Google Sheets ranked each request by performance.\u003c/li\u003e\n\u003c/ul\u003e\n\u003ch2 id=\"decisions\" class=\"relative group\"\u003eDecisions \u003c/h2\u003e\u003ch3 id=\"merit-over-seniority-for-shifts\" class=\"relative group\"\u003eMerit over seniority for shifts \u003c/h3\u003e\u003cp\u003eSeniority decided shifts before, which staff saw as unfair, and it took a lot of admin to run. I built the bidding so performance decided who got first pick.\u003c/p\u003e","title":"Taking a support team from 15th to the top 5"},{"content":"","date":"1 January 0001","permalink":"https://ccabaluna.pages.dev/roles/automation-ai/","section":"Roles","summary":"","title":"Automation and AI"},{"content":"The problem Matchworks runs a client portal where clients see their campaign and pay, and an admin where the team runs accounts. A screen can pass its tests and still confuse the person using it.\nWhat I do Release QA: I click through the portal and admin on each release and log what breaks or reads wrong. Session replays: I watch PostHog recordings (screen recordings of real visits) of real clients using the portal and turn each sticking point into a Linear ticket with the replay attached. UI and UX notes: layout, copy and flow changes written so an engineer can act on them without a meeting. Client portal: billing and invoices. Client data replaced. What one replay showed A client\u0026rsquo;s admin tried to get PDFs out of two leads: three download clicks in three minutes, plus one export window that opened and gave no download at all. The data could not show whether any of the downloads worked, because the portal logged the click and nothing after it. So I asked for success and failure tracking on every download before anyone chased the retries.\nOne ticket from start to finish: MAT-2318 On 7 November 2026 one replay turned into six tickets and one question for Mike, the founder.\n1. What I saw At 3:39 pm I was watching a client work a lead on the Communication tab, where emails, SMS and calls to that lead sit on one timeline. The client kept scrolling a short timeline. The \u0026ldquo;Your invoice is due\u0026rdquo; banner could not be closed and pushed everything down. They opened Inbox once, found no way in by lead, and went back. My notes from that minute named the banner and suggested letting users fold away the left column so the timeline gets more of the screen.\nThe client\u0026#39;s lead page at 1424 x 760, with where the height goes. Client data removed. 2. What I tested One user on one laptop is a single case, so I checked whether the cramped view was common. In PostHog (the analytics tool), 95% of client pageviews in the last 30 days came from windows under 1,000 px tall, and 32% from windows under 800 px. Then I measured that client\u0026rsquo;s screen: the banner took about 106 px, the header and tabs about 306 px, and the timeline was left with about 348 px. The left column took about 450 px of width, and the client barely clicked it.\nClient pageviews by browser window height, last 30 days At 4:23 pm I wrote down the case against my own point: on a bigger screen the page feels less cramped. The time data answered it. The Communication tab holds 45% of all time spent in the portal, and Inbox 9%. Whatever screen they use, clients spend most of their time on this one view.\nShare of portal time by screen, last 30 days 3. What else was broken on the same path The same afternoon\u0026rsquo;s replays turned up three bugs on that email path. The Email button crashed to an error page. Inbox showed a sent email as \u0026ldquo;Draft\u0026rdquo; until a reload. Autosave moved the typing spot below the signature. Two older tickets showed similar draft errors. I filed each as its own bug and wrote that they might share one cause in how drafts are saved, and labelled that as a guess, because nobody had read the code.\nFrames from the three bug tickets: error page, false Draft label, autosave jump. Client data removed. 4. What I wrote At 4:30 pm I filed MAT-2318 as the parent ticket in Linear and put the five related tickets under it. A first draft listed the bugs and asks as text. At 4:44 pm I had it rebuilt so the problem comes first as pictures (who is hit, where the space goes, the bug frames) and my suggestions come after, each with a sketch:\nLet users fold the left column to a strip, so the timeline takes the full width. Add a lead filter to Inbox: pick a lead and see only that lead\u0026rsquo;s emails and SMS. Make the invoice banner one line. The ticket opens with a one-line summary and ends with one question for Mike: more room on the lead page, an Inbox built around leads, or both.\nSuggestion 1: the lead page now, and a sketch with the left column folded. Client data removed. Suggestion 2: a sketch of Inbox filtered to one lead. Client data removed. I watched the replays, wrote the notes and chose the suggestions. Claude Code pulled the PostHog numbers, drew the charts and sketches and drafted the ticket body, and I sent it back until it read the way Mike needed it.\n5. Who decides The direction is Mike\u0026rsquo;s call. My end-of-day report on 7 November listed the six tickets and asked him to pick one. As of 10 November all six are still in triage, and nothing from this ticket has shipped.\nDecisions Follow the paths real clients take I follow the paths and settings real clients use, and trace where each one came from.\nReview the heaviest users most The clients who use the portal most get the most of my review time.\nRead each replay for the next move I read each replay for what the client will try next.\nThe result 19 QA tickets filed from 16 September to 7 November 2026, 4 of them shipped I know the platform well enough to sit in on a client\u0026rsquo;s call What I\u0026rsquo;d change I should have been more confident in my taste and shared my opinions sooner. Mike values my thoughts, and each one I held back for fear of a mistake was a thought waiting to be shared.\n","date":"1 November 2026","permalink":"https://ccabaluna.pages.dev/work/matchworks-portal/","section":"Work","summary":"\u003ch2 id=\"the-problem\" class=\"relative group\"\u003eThe problem \u003c/h2\u003e\u003cp\u003eMatchworks runs a client portal where clients see their campaign and pay, and an admin where the team runs accounts. A screen can pass its tests and still confuse the person using it.\u003c/p\u003e\n\u003ch2 id=\"what-i-do\" class=\"relative group\"\u003eWhat I do \u003c/h2\u003e\u003cul\u003e\n\u003cli\u003e\u003cstrong\u003eRelease QA:\u003c/strong\u003e I click through the portal and admin on each release and log what breaks or reads wrong.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eSession replays:\u003c/strong\u003e I watch PostHog recordings (screen recordings of real visits) of real clients using the portal and turn each sticking point into a Linear ticket with the replay attached.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eUI and UX notes:\u003c/strong\u003e layout, copy and flow changes written so an engineer can act on them without a meeting.\u003c/li\u003e\n\u003c/ul\u003e\n\n\n\n\n  \n\n  \n  \n  \u003cfigure class=\"work-shot\"\u003e\n    \u003ca href=\"/work/matchworks-portal/billing.webp\" aria-label=\"Open full size: Client portal: billing and invoices. Client data replaced.\"\u003e\u003cimg src=\"/work/matchworks-portal/billing_hu_38725ec4ad0c8272.webp\" width=\"1600\" height=\"754\" alt=\"Client portal: billing and invoices. Client data replaced.\" loading=\"lazy\"\u003e\u003c/a\u003e\n    \u003cfigcaption\u003eClient portal: billing and invoices. Client data replaced.\u003c/figcaption\u003e\n  \u003c/figure\u003e\n\n\n\n\n\n\n\n\n  \n\n\n\n\n\u003ch2 id=\"what-one-replay-showed\" class=\"relative group\"\u003eWhat one replay showed \u003c/h2\u003e\u003cp\u003eA client\u0026rsquo;s admin tried to get PDFs out of two leads: three download clicks in three minutes, plus one export window that opened and gave no download at all. The data could not show whether any of the downloads worked, because the portal logged the click and nothing after it. So I asked for success and failure tracking on every download before anyone chased the retries.\u003c/p\u003e","title":"Testing and UX review of the client portal"},{"content":"","date":"1 January 0001","permalink":"https://ccabaluna.pages.dev/roles/marketing/","section":"Roles","summary":"","title":"Marketing"},{"content":"The problem AI writes fast, and someone still has to check what it wrote. When I read every output myself, my reading speed caps the whole team. At Matchworks on 28 July 2026 I was exactly that: I read every reply draft before it went out, and the median prospect waited 27.1 minutes. That night Mike, who runs Matchworks, wrote “we have analtyics showing \u0026lt; 15m response time conversion goes way up”.\nHow I solved it I run my work through Claude Code, an AI assistant that works inside a computer and writes code from plain English. Everything in my setup does one of three jobs.\n1. Write the procedure once. A skill is a written procedure the AI follows: when to use it, the steps, its reference notes and scripts. I have 66 of them, and the Matchworks repo holds another 107. When a run works, it becomes a skill: the night described below became the case-study skill at 2:08 am on 10 November 2026. Skills also carry what I learn from a field\u0026rsquo;s best operators: the cold email libraries I built from them fed the outbound language I wrote at Matchworks.\n2. Check every output, not some. 15 checks run on their own while I work. The one I lean on most is a writing check I built on 28 July 2026: it sends back any reply or page that still reads as AI-written. It has run 5,517 times since and caught 143 banned phrases. From 1 October 2026 the Matchworks reply system ran the same check on its drafts, next to rules I wrote from my notes on real replies.\n3. Keep the record, and let it win. My context, rules and decisions live in one repository, with notes in Obsidian. Every piece of work is a Linear ticket. When I open a session, a board of my open tickets loads first, and the session\u0026rsquo;s plan shows as steps in a side panel next to its ticket. When the session ends, it posts what came of the work as a comment on that ticket, and a small AI check stops it from closing if no outcome was recorded. Since 18 August 2026 that board has held 191 tickets and 1,217 comments recording what each session did. 13 scheduled jobs write a news digest, a summary of my day and a self-review into my notes, 772 briefs since 25 July 2026. When the record and my memory disagree, the record wins, including about me: the Novum night below is one example.\nThe writing linter blocking a reply that reads as AI-written The agents pane: running sessions and their plans, tied to Linear Where it ran Matchworks reply system, Feb to Nov 2026. I built it in Claude Code on the same three jobs. The AI drafted, a separate program checked every draft as it sat in the inbox, and a draft that failed waited for a person with its reason in writing. It sent 6,968 of 8,820 replies with no person first and cut the median reply time from 27.1 to 5.6 minutes. From 3 to 25 August I graded 324 sent replies one at a time and marked 6 wrong. When 3 replies went to the wrong person on 1 and 2 October, a draft addressed to the wrong person was held from then on. Operators who aren\u0026rsquo;t engineers worked the held queue from a Raycast console I built. Case page Genius, 2024 to 2025. After my first screen by hand, Claude prompts scored each applicant against the role profile and drafted my screening notes, which is how I ran 5–6 searches at a time on my own. Case page Rebuilding my 2024 site-scoring tool, Oct to Dec 2025. I wrote the specs and Claude Code wrote most of the code: a batch system with job queues and a live dashboard, 147 commits in 9 weeks. Case page Matchworks client portal, Sep to Nov 2026. Claude Code pulled the PostHog numbers and drafted the tickets from my replay notes. 19 were filed by 7 November and 4 shipped. Case page One run: turning my Novum years into case pages On 9 and 10 November 2026 this setup turned my Novum work into 8 case pages in about 7 hours, and the records overturned my own memory once along the way.\n1. The ask At 7:23 pm on 9 November I asked Claude Code to help me recall my Novum projects and write them up properly this time, from my old journals and notes.\nHow the run started, 9 November 2026, 7:23 pm. Recreated terminal, my words as typed. 2. The sweep and the index Agents swept my notes, journals, Raycast chats, Claude Code chats and the Novum calendar into one index: 525 sources and 36 projects, each graded strong, possible or thin. The site-scoring tool came out strong, and most sources dated it to my Head of Link Building years, 2022 to 2023.\nThe index the sweep wrote, with the date my own accounts gave the tool. 3. The grilling round At 9:41 pm Claude asked me when I built the tool, with its own pick first, 2023, and I took it. The same round asked which records from the time I still had, and I gave it the tool\u0026rsquo;s GitHub repo.\nThe question and my first answer, 9 November 2026, 9:41 pm. Recreated terminal. 4. The record The repo\u0026rsquo;s first upload is dated 19 September 2024. Its log runs from 4 to 19 September 2024, and the code calls gpt-4o-mini, a model that came out in July 2024. That version of the tool could not have run in 2023.\nWhat the repo says, run against the public repo. 5. My memory, overturned At 9:43 pm the next question put the record to me: was there an earlier version, or was September 2024 the build? I answered at 9:50 pm that September 2024 was the build, and that it fed both Novum\u0026rsquo;s and Genius\u0026rsquo;s link building.\nThe follow-up question and my corrected answer. Recreated terminal. 6. The answers file Every settled fact goes into ANSWERS.md with the evidence behind it, and that file wins over the index. It records my first answer, that the code overturned it, and that every 2025 account dating the tool to 2022 or 2023 is wrong.\nANSWERS.md, the record of what I settled. 7. The published page At 1:12 am on 10 November I had the 8 Novum pages written from ANSWERS.md. They were committed at 1:15 am and on the live site by 2:31 am, with the tool dated September 2024. At 1:25 am I approved turning the run into a skill, case-study, so the next employer\u0026rsquo;s projects go through the same steps.\nThe site-scoring page on the live site, 10 November 2026. Also built on it A product reel, animated in code. On 6 and 7 October 2026 I built a 31.9-second Matchworks reel in HyperFrames, a tool that turns an animated web page into video. Every scene is HTML, and the music is generated by code at 128 beats a minute, so every cut falls on a beat. I posted it at 12:37 am on 7 October, and that evening Mike asked for a sales video. Every video project ends with a lessons file, and the next one starts by reading all of them: 9 projects since 1 October. Case page Research libraries and public wikis. Agents pull the videos, posts and interviews of a field\u0026rsquo;s best operators and work through them in three passes: tag each claim with its source, compare sources, then keep what holds as notes an AI can work from. That\u0026rsquo;s 31 libraries so far, from cold email to Claude Code itself. A library can also review a draft or a decision by that field\u0026rsquo;s rules alone; 21 are set up that way. Some become public wikis, 471 pages across six of them, and on one, a second agent checked each of 15 worked examples claim by claim and found 23 small errors. What it would do in your business The same three jobs on your work: write the procedure once so anyone, person or AI, runs it the same way; put a check on every output so nobody has to read everything; and keep one record your team trusts over anyone\u0026rsquo;s memory.\nWhat I\u0026rsquo;d change I\u0026rsquo;d start with one repository for my context and rules from day one. Before that, my memories and rules sat in a master projects folder and several repos, and the work kept getting lost between them.\n","date":"30 October 2026","permalink":"https://ccabaluna.pages.dev/work/s0mebody-os/","section":"Work","summary":"\u003ch2 id=\"the-problem\" class=\"relative group\"\u003eThe problem \u003c/h2\u003e\u003cp\u003eAI writes fast, and someone still has to check what it wrote. When I read every output myself, my reading speed caps the whole team. At Matchworks on 28 July 2026 I was exactly that: I read every reply draft before it went out, and the median prospect waited 27.1 minutes. That night Mike, who runs Matchworks, wrote “we have analtyics showing \u0026lt; 15m response time conversion goes way up”.\u003c/p\u003e","title":"The AI setup I work from"},{"content":"","date":"1 January 0001","permalink":"https://ccabaluna.pages.dev/roles/support/","section":"Roles","summary":"","title":"Support"},{"content":"The problem In October 2025 I moved to the CEO\u0026rsquo;s AI content work, making UGC campaigns (ads in the style of a customer\u0026rsquo;s own video) and Instagram content. In 2026 Matchworks needed a sales video, and editing every frame by hand was too slow to keep up.\nWhat I built UGC and Instagram motion graphics: for the CEO\u0026rsquo;s campaigns. A Matchworks sales video: scripted, voiced and animated in code with HyperFrames, a tool that turns a web page into video. Personal reels: used to test pacing, captions synced to the voice and transitions before they go into client work. Decisions Never time-stretch a voice When a cut ran long, I changed the words, the pauses or the ElevenLabs voice speed, and never stretched the recorded voice to fit, because stretching a voice breaks its quality. I set that rule on 9 October 2026.\nProduct shots redrawn as cards After the fifth cut of the platform video I said the pages still looked screenshotted. So I redrew each product shot as a card with made-up data instead of a screenshot of the live app, and checked every render for real names before it went out.\nOne video, from Mike\u0026rsquo;s ask to YouTube The Matchworks sales video took 4 days from Mike\u0026rsquo;s message on 7 October 2026 to the cut now on YouTube, with one rebuild in the middle.\n1. The ask On 7 October at 8:08 pm Mike wrote that he would love a VSL, a video sales letter, to fill the YouTube channel. That came three minutes after he called my first reel cool in the team channel. I answered at 8:09 pm that I was working on one.\nMike\u0026#39;s ask and my answer, 7 October 2026. Recreated, original data private. 2. The script That night I studied 73 videos by one VSL writer, then took the script through four versions. In the file each scene is a heading and each line is voiced as its own clip, so a pause like the 2.6 seconds before \u0026ldquo;Matchworks built the infrastructure\u0026rdquo; is one line of text.\nThe script file: one heading per scene, one voiced clip per line. 3. The voice The ElevenLabs voice read 331 words in 1 minute 40 seconds against a 2 minute 10 second plan. My first cut slowed it to fit; I asked for his own pace in the next cut instead. The closing line ended on a rising pitch, so I measured its last 0.4 seconds and re-voiced it from 96% down to 80%.\nMy notes from the build night, 8 October 2026. 4. The scenes in HyperFrames Every scene is HTML animated in code. The problem half shows the four pieces of work in grey, then 57 seconds in the four boxes lift and become the m of the wordmark, and the second half runs the same four pieces in brand colour.\nFour stills from the 8 October cut. 5. Checking the turn frame by frame I read the turn as a strip of frames before each render. The first try pasted the reel\u0026rsquo;s opening into the gap; it looked slapped on, so the brand moment now grows out of the four boxes the viewer has watched since the 10-second mark.\nThe contact sheet I checked the turn against, 8 October 2026. 6. The first cut, and the real ask I sent the first cut at 6:35 am on 8 October and said I had assumed a lot. At 8:13 pm Mike liked the direction. At 9:15 pm he named what he wanted: a 90-second video of how the platform works. So I rebuilt it from his notes.\nThe first cut goes out, and Mike\u0026#39;s reply, 8 October 2026. Recreated, original data private. 7. The cut on YouTube The rebuild went through four rounds of fresh readers and two new openings. The one I kept was rendered on 11 October at 3:14 am, 2 minutes 9 seconds long, and I put it on YouTube on 10 November.\nTwo stills from the cut on YouTube, rendered 11 October 2026. The result After the first reel, posted at 12:37 am on 7 Oct 2026, Mike asked for a video sales letter, then on 8 Oct for a 90-second video of how the platform works. That platform video is the one above. On 6 Nov Mike put a motion video for the matchworks.ai homepage second on the list of three projects he set for me. ","date":"26 October 2026","permalink":"https://ccabaluna.pages.dev/work/motion-videos/","section":"Work","summary":"\u003ch2 id=\"the-problem\" class=\"relative group\"\u003eThe problem \u003c/h2\u003e\u003cp\u003eIn October 2025 I moved to the CEO\u0026rsquo;s AI content work, making UGC campaigns (ads in the style of a customer\u0026rsquo;s own video) and Instagram content. In 2026 Matchworks needed a sales video, and editing every frame by hand was too slow to keep up.\u003c/p\u003e\n\u003ch2 id=\"what-i-built\" class=\"relative group\"\u003eWhat I built \u003c/h2\u003e\u003cul\u003e\n\u003cli\u003e\u003cstrong\u003eUGC and Instagram motion graphics:\u003c/strong\u003e for the CEO\u0026rsquo;s campaigns.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eA Matchworks sales video:\u003c/strong\u003e scripted, voiced and animated in code with HyperFrames, a tool that turns a web page into video.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003ePersonal reels:\u003c/strong\u003e used to test pacing, captions synced to the voice and transitions before they go into client work.\u003c/li\u003e\n\u003c/ul\u003e\n\n\n\n\n  \n\n\n\n\n\u003cdiv style=\"position: relative; padding-bottom: 56.25%; height: 0; overflow: hidden;\"\u003e\n      \u003ciframe allow=\"accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share; fullscreen\" loading=\"eager\" referrerpolicy=\"strict-origin-when-cross-origin\" src=\"https://www.youtube-nocookie.com/embed/4MJcnEze6u4?autoplay=0\u0026amp;controls=1\u0026amp;end=0\u0026amp;loop=0\u0026amp;mute=0\u0026amp;start=0\" style=\"position: absolute; top: 0; left: 0; width: 100%; height: 100%; border:0;\" title=\"How the Matchworks platform works\"\u003e\u003c/iframe\u003e\n    \u003c/div\u003e\n\n\u003ch2 id=\"decisions\" class=\"relative group\"\u003eDecisions \u003c/h2\u003e\u003ch3 id=\"never-time-stretch-a-voice\" class=\"relative group\"\u003eNever time-stretch a voice \u003c/h3\u003e\u003cp\u003eWhen a cut ran long, I changed the words, the pauses or the ElevenLabs voice speed, and never stretched the recorded voice to fit, because stretching a voice breaks its quality. I set that rule on 9 October 2026.\u003c/p\u003e","title":"Motion videos, from campaign edits to code"},{"content":"","date":"1 January 0001","permalink":"https://ccabaluna.pages.dev/roles/recruitment/","section":"Roles","summary":"","title":"Recruitment"},{"content":"The problem My friend Sai wrote a video series in Taglish, the everyday mix of Tagalog and English, on why the Philippines stayed poor while its neighbours grew. It follows the book Why Nations Fail. On 11 October 2026 he gave me his project files and I forked them, to see whether my Claude Code setup could make the episodes better.\nNeither of us had made an explainer before. So every call had to be checked against something: a reader who had never seen the script, a source text, or the numbers from the last episode.\nWhat I built Episode 1, \u0026ldquo;Bakit Hindi Umaasenso?\u0026rdquo;: 5 minutes 49 seconds, animated in code with HyperFrames, voiced by Sai, posted on 15 October. Episode 2, \u0026ldquo;Mula pa nung Kastila\u0026rdquo;: the script, rewritten nine times in Claude Code. Sai recorded it, built the cut and posted it on 2 November. Episode 3, \u0026ldquo;Ang lupa\u0026rdquo;: a new script, four rendered test openings and a panel of six simulated viewers, still in test on 12 November. A recording page: Sai reads one line at a time on his phone and the takes come back as files. Every step below comes from the Claude Code session logs, the project files and our Linear board. Times are Philippine time.\nEpisode 1, from Sai\u0026rsquo;s script to the post Episode 1 took 5 days, from my fork on 11 October to the post on 15 October.\nOne opening frame, five renders 13 Oct, 2:06 am Melets do an a/b test. create the full reel through the content creation workflow we have. Claude CodeVoiced the script with ElevenLabs, built every scene as HTML in HyperFrames and rendered a first draft by 3:11 am. v1 · 3:52 13 Oct, 3:16 am Mei watched it and i would like to see this fully ready for production. do you very best! Claude CodeRan three rounds of fresh readers on the frames, fixed what they misread, and added the dollar figures and the source line. v4 · 3:55 13 Oct, 1:33 pm Mecan we time the script properly. we\u0026#39;re not taking into account the proper pauses (to guide the viewers attention properly). overall, the voice feels rushed. Claude CodeMeasured the voice at 168 words a minute with 0.3 seconds between lines, slowed it, wrote the pauses into the script and rebuilt the animation on one chart. v6 · 4:43 13 Oct, 5:19 pm Methis is our hook, right? So it must be optimized visually to be a hook. So maybe like the headline should be hooky. Claude CodeMade the headline the line a thumbnail would carry, added the \u0026#43;69% bracket and a chapter bar, and switched to a new narrator voice. v7g · 5:47 14 Oct, 8:42 pm Mesome parts sound like it\u0026#39;s glitching. not sure if its because we sped up some parts of sai\u0026#39;s takes but i fthat\u0026#39;s the trade off then we pick quality. Claude CodeFound 73 clicks on the joins between the audio blocks the recording page saved, redrew them, and re-timed every picture to Sai\u0026#39;s voice. The frame stayed; the sound under it changed. v8e · 5:49 · posted My messages are quoted from the Claude Code session logs of 13 to 15 October 2026. Each frame is 3 seconds into that render. 1. The fork At 4:29 am on 11 October I asked Claude Code to stress test Sai\u0026rsquo;s script. It ran 9 agents: 4 gathered facts and lessons, 2 rewrote, and 3 read both versions blind as a 24-year-old viewer in Dumaguete. The rewrite won all 3 reads. For Episode 1 the readers put the chance of finishing at 38% and 44%, against 29% and 33% for Sai\u0026rsquo;s.\n2. A fix that made it worse At 5:05 am I said we were skipping too many links between ideas. The next pass added them, so the script grew from 475 words to 710 and the readers\u0026rsquo; chance of finishing fell from 55% to 40%. That showed the question was too wide, so I narrowed it to who had the power to block land reform.\n3. Taglish, then four lenses On 12 October I moved the script to everyday Taglish. Four lens passes (storytelling, teaching, facts and how the Taglish sounds) found 75 things and I applied 62. One was a date: Korea\u0026rsquo;s land reform law passed in June 1949, and the script said 1950. Then I cut the script from 662 words to 467.\n4. First render in 11 minutes At 3:00 am on 13 October I asked for the long version as a video. The first render was done at 3:11 am, 3 minutes 52 seconds long, with an ElevenLabs voice. By 5:11 am it had been through three rounds of fresh readers and one of them had checked the facts in the post.\n5. The voice was rushed At 1:33 pm I wrote that the voice felt rushed. Claude measured it: 168 words a minute, with 0.3 seconds between lines. I chose a slower voice setting with written pauses. It then offered to cut 24 lines of narration. I said no and asked for the animation to carry what the words could not, which gave the 4:43 cut.\n6. A new narrator, then Sai At 5:19 pm I heard that the slow voice was only the old voice slowed down, so we designed a new one. It said \u0026ldquo;lamang\u0026rdquo; with the wrong stress, which changes the word\u0026rsquo;s meaning, so that line was reworded. On 14 October at 8:07 pm Sai\u0026rsquo;s 51 recorded takes arrived and replaced the generated voice.\n7. The clicks I heard glitches in Sai\u0026rsquo;s takes and guessed we had sped them up. Claude found 73 clicks sitting on the joins between the small blocks of audio that the recording page saved. A new tool redrew the 109 joins in his takes, and the recording page was fixed before Episode 2.\n8. Posted The last cut is 5 minutes 48.7 seconds. Before it went out, a check confirmed that all 337 on-screen cues matched a spoken word and that no picture stayed up for less than 1.5 seconds. Sai shared the post at 1:03 am on 15 October.\nTen frames from the posted cut of Episode 1, 15 October 2026. Episode 2, nine versions of one script For Episode 2 I worked on the script and Sai made the video.\nThe opening lines, four times Before 11 Oct SaiHis first script, in a documentary register. It opens on the number. Sai\u0026#39;s first script · 454 words Noong 1591, humigit-kumulang 667,000 katao sa Pilipinas ang nakalista sa ilalim ng mga encomienda.At madalas, iisang Kastila lamang ang may hawak sa 1,000 hanggang 1,200 na nagbabayad ng tributo. In 1591 about 667,000 people in the Philippines were listed under the encomiendas. Often one Spaniard held 1,000 to 1,200 tribute payers. 15 Oct SaiHis own draft, in the everyday Taglish I chose on 12 October. Four lens walks on 23 October found 55 things to fix in it. Sai\u0026#39;s draft · 597 words Nung 1591, mga 667,000 na tao sa Pilipinas ang na-assign sa mga Kastila.Madalas, napupunta sa iisang Kastila ang tax ng mga one thousand na pamilya kada taon.Tumagal ang tax na \u0026#39;yan ng mahigit three hundred years, at may mga hindi nagbabayad nito.Sino sila? The tax lasted more than three hundred years, and some people never paid it. Who were they? 23 Oct, 5:54 pm Mei just feel like it\u0026#39;s not cohesive. it\u0026#39;s a sum of different parts but not a whole cohesive entity? idk Claude CodeRebuilt the script around one question and put the two stock answers, lazy and corrupt, in the first line. Version 2 · 45 lines Bakit hindi umaasenso ang Pinoy? Tamad daw, o corrupt ang mga nasa gobyerno.Pero nung 1591, mga 667,000 na Pilipino ang nasa listahan ng tax ng mga Kastila.Kaya bago natin sabihing tamad o corrupt, sundan muna natin kung ano ang kinukuha sa isang pamilya. Why don\u0026#39;t Filipinos get ahead? Lazy, they say, or the government is corrupt. So before we say lazy or corrupt, follow what was taken from one family. 23 Oct, 7:27 pm Methe opening still doesnt sound cohesive. we should redo this from scratch i dont think this is worth saving Claude CodeWrote a new script from scratch that picks up where Episode 1 ended, ran four more lens walks on it, and counted families where the old line counted people. The voice never says lazy or corrupt. Version 3, recorded · 682 words Sa Episode 1, nakita natin: nung 1955, maraming landlord ang nasa Congress, kaya sila mismo ang gumawa ng rules sa land reform.Pero kailan ba nagsimula ang ganitong rules dito?Nung 1591, mga 167,000 na pamilya na ang nagbabayad ng tax sa mga Kastila. In Episode 1 we saw that in 1955 many landlords sat in Congress and wrote the land reform rules themselves. But when did rules like that start here? In 1591 about 167,000 families were already paying tax to the Spanish. Lines are copied from the four script files, with the voice cues removed. My messages are quoted from the session log of 23 October 2026. 1. Four lens walks At 1:34 am on 23 October I asked for Sai\u0026rsquo;s 597-word draft to be stress tested. Four agents walked it, each with one lens: thinking tools, a model of how people at different stages hear an argument, a creator\u0026rsquo;s storytelling notes, and copywriting. They found 55 things, 2 of them blockers. The larger blocker: the middle of the script read as a fraud story, which fed the \u0026ldquo;corrupt\u0026rdquo; answer the episode argues against.\n2. No more review documents At 2:34 am I stopped the review document. I asked Claude to put each choice to me as a question with its own pick first, so I could confirm or correct in seconds. That became the rule for all my projects.\n3. Thrown out and written again At 5:54 pm I said the script was a sum of parts and not one piece. The rebuild around one question came to 45 lines. At 7:27 pm I said it was not worth saving. The next version was written from scratch around three moves the Spanish made (the chiefs, the tribute, the forced labour) and was ready at 7:52 pm.\n4. Four more walks, then the recording page Four more walks ran on the new script at 8:00 pm. After my answers it stood at 38 lines and 682 words. At 8:51 pm the recording page held Episode 2 next to Episode 1, with a comment box on every line.\n5. Sai records and posts Sai recorded 38 takes on 30 October, one for each line, then built the cut himself with real archive pages on screen. He posted it on 2 November at 5 minutes 49 seconds.\nFour seconds into Episode 2, from Sai\u0026#39;s cut: the headline over a recap of Episode 1. 6. What the numbers said On 29 October I read Episode 1\u0026rsquo;s numbers: the biggest drop came 9 seconds in. Four more walks said Episode 2 should open on its claim and move the recap later. I parked that advice that morning because the recording was not ready, and Episode 2 went out with the recap first. On 11 November it had 473 plays at 9 days and a 14.5-second average watch. Episode 1 had 913 plays at 28 days and 16.9 seconds.\nEpisode 3, testing the opening before recording Episode 3 starts from what Episode 2 taught: the opening decides who stays.\nSix viewers, four openings Round 1, 11 Nov, 8:35 am The panelSai\u0026#39;s opening recaps Episode 2. It came last of seven, and every viewer swiped on the first words. Sai\u0026#39;s recap opening Man, 28, Quezon City, night shift 15% “\u0026#39;Sa Ep 2\u0026#39; agad, parang late ako sa usapan” Woman, 31, Cebu City, sound off 15% Man, 34, Davao, small business 18% Man, 25, Laguna, follows the page 30% “feels like homework, not a story” Woman, 26, Manila, sound off 8% Man, 39, Bulacan, family farmed rented land 17% “feels like a class I missed” On the real scale: 30.1%. Episode 1's real opening: 42.0%. Round 2, 11 Nov, 8:59 am MeAnatomy installation (we need to leverage this more) Claude CodePut the land-office figure under the opening line as the first frame. Four of six read it as an ad. The figure as frame one Man, 28, Quezon City, night shift 43% “looks like a product ad” Woman, 31, Cebu City, sound off 30% “parang tech ad, nalito ako” Man, 34, Davao, small business 30% Man, 25, Laguna, follows the page 42% “Looks like a tech product ad, not history” Woman, 26, Manila, sound off 30% Man, 39, Bulacan, family farmed rented land 38% “parang ad ng gadget, malayo sa buhay namin” On the real scale: 34.6%. Episode 1's real opening: 42.0%. Round 4, 11 Nov, 2:56 pm MeRender the openings Claude CodeRendered four openings as silent 10-second clips and ran the panel on the real frames. This one put a real document under a twist. Rendered test B Man, 28, Quezon City, night shift 48% “a twist I did not learn in school” Woman, 31, Cebu City, sound off 30% “an old grey book page” Man, 34, Davao, small business 50% Man, 25, Laguna, follows the page 70% “gives me a puzzle and a number” Woman, 26, Manila, sound off 38% Man, 39, Bulacan, family farmed rented land 48% “lands close to home, since my family rented land too” On the real scale: 41.5%. Episode 1's real opening: 42.0%. Round 10, 12 Nov, 12:45 am MeIf they don\u0026#39;t understand us, we already lost Claude CodeWrote four beats and six checks every opening must pass, then a fresh batch. This one asks whose name is on the land under your own house. Titulo Man, 28, Quezon City, night shift 45% “kaba, curious” Woman, 31, Cebu City, sound off 45% “170,000 ha, di ko ma-imagine gaano kalaki” Man, 34, Davao, small business 50% “napaisip” Man, 25, Laguna, follows the page 60% Woman, 26, Manila, sound off 45% Man, 39, Bulacan, family farmed rented land 65% “kirot” On the real scale: 43.7%. Episode 1's real opening: 42.0%. The six viewers are simulated: one Claude agent each, with a fixed profile, scoring every opening blind. Bars are each viewer\u0026#39;s own chance of still watching at 3 seconds. Remarks are quoted from the panel\u0026#39;s answer files of 11 and 12 November 2026. 1. A framework first On 11 November at 4:54 am I said I would never reach an ending by picking between options, and asked for a framework. Eight agents studied the five channels I named and seven Filipino explainer creators. Twelve more turned that into a 13,000-word briefing. Its own machine checks failed on two of four counts, names and tone.\n2. Where viewers leave Claude pulled the page\u0026rsquo;s data and fitted a curve to each episode. About 58% of Episode 1\u0026rsquo;s plays and 68% of Episode 2\u0026rsquo;s ended before Facebook\u0026rsquo;s own graph begins. So the first three seconds became the thing to fix.\n3. A panel of six Six simulated viewers scored each opening blind. The real openings of Episode 1 and Episode 2 sat among the candidates, unlabelled, so the panel\u0026rsquo;s scores could be set against the real numbers. All six put Episode 1 above Episode 2, which matches the page. Sai\u0026rsquo;s recap opening came last.\n4. The figure that lost I wanted an Anatomy figure, a land office that turns village land into one owner\u0026rsquo;s title, as the first frame. Claude built it with 94 parts and proved it renders as video. The panel scored it 34.6%, and four of the six viewers read it as a gadget ad, so it moved inside the episode.\nThe land-office figure built for Episode 3 on 11 November 2026. 5. Four rendered openings An audit of Sai\u0026rsquo;s script proposed 19 edits and I took each one. The new script has 44 lines. Four openings were rendered as silent 10-second clips. Claude caught that one of them put a line about friar estates over a photo of Hacienda Luisita, which was never a friar estate, and swapped the photo.\n6. Undercooked At 10:56 pm I called those openings undercooked. By 12:13 am I had the rule I was missing: if viewers do not understand us, we have already lost. That became four beats and six checks every opening must pass. Three more batches went to the panel that night. By 1:00 am on 12 November it had run 10 rounds, 60 viewer runs in all.\n7. Where it stands By that hour no opening had beaten Episode 1\u0026rsquo;s real 42% by more than the panel\u0026rsquo;s noise. The closest, \u0026ldquo;Titulo\u0026rdquo;, scored 40.7%, 43.7% and 39.9% in three rounds. The next step is a test on real viewers, with up to four openings, judged on views that last a minute.\nThe result Episode 1 had 913 plays, 10 shares and 9 new followers at 28 days. Episode 2 had 473 plays at 9 days. The page had 22 followers when I read the numbers on 11 November 2026. Six shorter reels cut from Episode 1 were rendered on 15 October and never posted. The method now lives in a playbook Sai and I both work from: ten steps to make a video, with reading the numbers as the last one. ","date":"12 November 2026","permalink":"https://ccabaluna.pages.dev/work/wnf-episodes/","section":"Work","summary":"\u003ch2 id=\"the-problem\" class=\"relative group\"\u003eThe problem \u003c/h2\u003e\u003cp\u003eMy friend Sai wrote a video series in Taglish, the everyday mix of Tagalog and English, on why the Philippines stayed poor while its neighbours grew. It follows the book Why Nations Fail. On 11 October 2026 he gave me his project files and I forked them, to see whether my Claude Code setup could make the episodes better.\u003c/p\u003e\n\u003cp\u003eNeither of us had made an explainer before. So every call had to be checked against something: a reader who had never seen the script, a source text, or the numbers from the last episode.\u003c/p\u003e","title":"Three episodes of a Filipino explainer, made in Claude Code"},{"content":"The problem Every link Novum sold was won by pitching a guest post to another site\u0026rsquo;s editor. Before anyone could pitch, someone had to check the sites. An assistant spent about 8 hours a day to find 300–500 sites worth pitching out of a list of 1,000–3,000 pulled from Ahrefs or Sales Navigator.\nWhat I built A small web app in Python and Flask, with upload, progress and dashboard pages:\nUpload a list of sites from Ahrefs or Sales Navigator. Scrape each site\u0026rsquo;s title, description and page text. Ask GPT (gpt-4o-mini) whether the site fits the outreach, and get back \u0026ldquo;proceed\u0026rdquo; or not. Export the sites worth pitching as a CSV for the team. It ran overnight on 10,000–20,000 sites, by my own count. The tool only did the first pass. The assistants still picked the right person at each site and found the email by hand.\nI built it in September 2024. The first runs in its logs are from 4 to 19 September 2024. A friend helped me set up the server.\nSite scoring tool. Recreated, original data private. Decisions Only the first pass I kept the tool to one job: drop the sites that would never take a pitch. Finding the right contact needs judgment about who runs a site, so that stayed with the assistants.\nOne list through the tool To show a run step by step, I put one list of 33 public sites through the September 2024 code again on 10 November 2026. The 2025 rebuild page runs the same list.\n1. The list A run started with a plain file, one domain per line, pulled from Ahrefs or Sales Navigator. The logged runs of 4 to 19 September 2024 used a sheet of 100 SaaS domains. For this page I made a list of 33 public sites: software companies with blogs, marketing blogs, and sites the tool should drop, such as health, finance, crypto, travel, news, .org, .edu and .gov.\nThe 33-site list used on both pages, shown as a table. Public sites picked for this page on 10 Nov 2026. 2. Upload The app opens on one page, the task dashboard. I pick the file and press Start Task. It runs one list at a time. A second upload while a list is running gets \u0026ldquo;A task is already running.\u0026rdquo;\nThe real 2024 app, run on my Mac on 10 Nov 2026 from the 19 Sep 2024 code. The screen is real; nothing on it is mocked. 3. The run Each site goes through three checks. First GPT reads the bare address against ten blocked groups, such as news, health, finance, education and crypto. If none match, ScrapingBee fetches the homepage and keeps the first 500 characters of the title, description and text. Then GPT answers \u0026ldquo;Proceed\u0026rdquo; or \u0026ldquo;No Fit\u0026rdquo; for guest posting. The page shows which site it is on.\nThe run at site 13 of 33. Real app; GPT and ScrapingBee were replaced by stand-ins, so no API key was used. 4. Done When the last site is checked, the row turns to Completed and a Download Results button appears. With the stand-ins, the 33 sites took about 2 minutes, most of it the pauses the code puts between sites.\nThe finished run with its download button. Real app, GPT and ScrapingBee stubbed. 5. The CSV The file has one row per site: the result, a score and GPT\u0026rsquo;s answer. A keyword score adds points for words like \u0026ldquo;product\u0026rdquo;, \u0026ldquo;pricing\u0026rdquo; and \u0026ldquo;SaaS\u0026rdquo;, takes them away for \u0026ldquo;news\u0026rdquo; or \u0026ldquo;health\u0026rdquo;, and GPT\u0026rsquo;s \u0026ldquo;proceed\u0026rdquo; adds 50. A site needs 50 to pass. This list came out 9 Proceed, 15 No Fit and 9 Filtered. Copyblogger and Search Engine Journal got a \u0026ldquo;proceed\u0026rdquo; from GPT and still scored 40, so they stayed out. The assistants took the Proceed rows and found the contacts by hand.\nresults_0.csv, the file this run wrote, shown as a table. Rows and scores are the tool\u0026#39;s real output; the GPT answers came from a stand-in rule, not OpenAI. The result A day of lead research went from about 8 hours to 2–3. The same tool fed both Novum\u0026rsquo;s link building and Genius\u0026rsquo;s. In late 2025 I rebuilt it as a full batch system. That is its own case study. ","date":"19 September 2024","permalink":"https://ccabaluna.pages.dev/work/novum-site-scoring-tool/","section":"Work","summary":"\u003ch2 id=\"the-problem\" class=\"relative group\"\u003eThe problem \u003c/h2\u003e\u003cp\u003eEvery link Novum sold was won by pitching a guest post to another site\u0026rsquo;s editor. Before anyone could pitch, someone had to check the sites. An assistant spent about 8 hours a day to find 300–500 sites worth pitching out of a list of 1,000–3,000 pulled from Ahrefs or Sales Navigator.\u003c/p\u003e\n\u003ch2 id=\"what-i-built\" class=\"relative group\"\u003eWhat I built \u003c/h2\u003e\u003cp\u003eA small web app in Python and Flask, with upload, progress and dashboard pages:\u003c/p\u003e","title":"A site-scoring tool that cut a day of lead research to 2–3 hours"},{"content":"The result Date Followers Record 6 Jan 2026 0 Profile screenshot 30 Jan 2026 about 2,000 Planning doc from that day 14 Feb 2026 about 4,500 The CEO\u0026rsquo;s count, when he stopped the boosts Two boosts are on record, both from the Instagram app with profile visits as the goal:\nBoost Spent Views Profile visits Follows Per follow 19 Jan 2026 $21.88 3,297 598 105 $0.21 27 Jan 2026 $15.96 2,153 291 45 $0.35 The total spent on boosts is not on record, so I don\u0026rsquo;t give one. On 14 February the CEO stopped the boosts to see how the account did on its own. Without them, views per reel fell, and he closed the account on 22 February 2026.\nThe 19 January boost: 598 profile visits at $0.04 each, $21.88 of a $50 budget spent. The 27 January boost: 45 follows from 291 profile visits, $15.96 spent. Its top locations were California and Texas, at 11% each. The problem In October 2025 the CEO of Novum and Genius asked me to build an Instagram account around an AI-generated character, from nothing. It needed the same face in every post and a post every day. When he asked, no tool was set up and nobody on the team had built an account like it.\nHow I built it A US phone (October 2025). The account posted from a US phone on a US mobile network, since its audience was in the US. It was set up and warmed by 27 October. An image pipeline on rented GPUs (November to December 2025). I ran ComfyUI on RunPod GPUs and adapted community workflows for portraits, poses and video. Then I trained a LoRA, a small add-on model that teaches the image model one face. My first one failed on 30 November; by 14 December I had one that held the face from post to post. Daily posting (from 20 December 2025). The first posts went up on 20 December. In January I posted one reel and one carousel a day, and from mid-February two reels a day plus stories. Paid boosts (January to 14 February 2026). I boosted reels at $5 to $10 a day and worked out what each follow cost. The profile on 6 January 2026, before the first boost: 0 followers. Photo and name blocked. Decisions Pause the GPUs, every time A RunPod GPU cost about $1.25 an hour, about $900 a month if left on. On 15 October, before any training, I wrote a pause procedure so a pod never ran while nobody was using it.\nCount the follow, not the view The app reports views first. On 19 January I worked out the cost per profile visit and per follow, since followers were the goal the CEO set.\nI also cut motion edits for this work; they are on the motion videos page.\n","date":"22 February 2026","permalink":"https://ccabaluna.pages.dev/work/ai-instagram-account/","section":"Work","summary":"\u003ch2 id=\"the-result\" class=\"relative group\"\u003eThe result \u003c/h2\u003e\u003ctable\u003e\n  \u003cthead\u003e\n      \u003ctr\u003e\n          \u003cth\u003eDate\u003c/th\u003e\n          \u003cth\u003eFollowers\u003c/th\u003e\n          \u003cth\u003eRecord\u003c/th\u003e\n      \u003c/tr\u003e\n  \u003c/thead\u003e\n  \u003ctbody\u003e\n      \u003ctr\u003e\n          \u003ctd\u003e6 Jan 2026\u003c/td\u003e\n          \u003ctd\u003e0\u003c/td\u003e\n          \u003ctd\u003eProfile screenshot\u003c/td\u003e\n      \u003c/tr\u003e\n      \u003ctr\u003e\n          \u003ctd\u003e30 Jan 2026\u003c/td\u003e\n          \u003ctd\u003eabout 2,000\u003c/td\u003e\n          \u003ctd\u003ePlanning doc from that day\u003c/td\u003e\n      \u003c/tr\u003e\n      \u003ctr\u003e\n          \u003ctd\u003e14 Feb 2026\u003c/td\u003e\n          \u003ctd\u003eabout 4,500\u003c/td\u003e\n          \u003ctd\u003eThe CEO\u0026rsquo;s count, when he stopped the boosts\u003c/td\u003e\n      \u003c/tr\u003e\n  \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eTwo boosts are on record, both from the Instagram app with profile visits as the goal:\u003c/p\u003e\n\u003ctable\u003e\n  \u003cthead\u003e\n      \u003ctr\u003e\n          \u003cth\u003eBoost\u003c/th\u003e\n          \u003cth\u003eSpent\u003c/th\u003e\n          \u003cth\u003eViews\u003c/th\u003e\n          \u003cth\u003eProfile visits\u003c/th\u003e\n          \u003cth\u003eFollows\u003c/th\u003e\n          \u003cth\u003ePer follow\u003c/th\u003e\n      \u003c/tr\u003e\n  \u003c/thead\u003e\n  \u003ctbody\u003e\n      \u003ctr\u003e\n          \u003ctd\u003e19 Jan 2026\u003c/td\u003e\n          \u003ctd\u003e$21.88\u003c/td\u003e\n          \u003ctd\u003e3,297\u003c/td\u003e\n          \u003ctd\u003e598\u003c/td\u003e\n          \u003ctd\u003e105\u003c/td\u003e\n          \u003ctd\u003e$0.21\u003c/td\u003e\n      \u003c/tr\u003e\n      \u003ctr\u003e\n          \u003ctd\u003e27 Jan 2026\u003c/td\u003e\n          \u003ctd\u003e$15.96\u003c/td\u003e\n          \u003ctd\u003e2,153\u003c/td\u003e\n          \u003ctd\u003e291\u003c/td\u003e\n          \u003ctd\u003e45\u003c/td\u003e\n          \u003ctd\u003e$0.35\u003c/td\u003e\n      \u003c/tr\u003e\n  \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003eThe total spent on boosts is not on record, so I don\u0026rsquo;t give one. On 14 February the CEO stopped the boosts to see how the account did on its own. Without them, views per reel fell, and he closed the account on 22 February 2026.\u003c/p\u003e","title":"An AI-generated Instagram account, 0 to about 4,500 followers in six weeks"},{"content":"The problem Every link Novum sold and every client it won started with a cold email. I sent my first guest-post pitch on 14 June 2021, and by 2023 the agency sent from dozens of mailboxes across lookalike domains such as novumhq.net and novumhub.com.\nIn March 2023 our warmup tool said our emails reached the inbox. I booked a one-hour consulting call with GlockApps, a deliverability testing company, on 13 March 2023. Two days later the expert sent his findings: \u0026ldquo;Your warmup tool says the emails are inboxing but the Glockapps tests show that they are landing in the spam folder.\u0026rdquo;\nWhat changed Smartlead. I signed the agency up on 30 March 2023 and connected about 12 mailboxes the same day, moving off Sales.Rocks. Fewer mailboxes per domain. We cut to one or two per domain, as the call advised. Lower volume. Each mailbox started at a few emails a day and stepped up slowly. Office 365. We tried it next to Zoho and Google Workspace, and I asked GlockApps for help setting it up on 27 April 2023. List checks. Every lead list went through Debounce before sending. Burkhard and I built the rules together. I connected the mailboxes, ran the campaigns and brought in the expert.\nHow the setup changed, 2021 to 2024 Most of the work was keeping mail out of spam day after day. This is how the sending setup moved from hand work to tools, by date.\nWhen What changed May 2021 I sent from Pitchbox in my own mailbox, under Burkhard\u0026rsquo;s name, up to 150 emails a day per campaign. Sep 2021 Manual three-step pitches from Mailshake. Feb to Aug 2022 Sales.Rocks sent and warmed the mailboxes with its built-in warmer, until Burkhard removed the warming on 2 August. Sep 2022 Every lead list went through Debounce before sending. 17 Mar 2023 After the GlockApps call, I sorted 103 lead domains by their mail host: Google, Microsoft or other. 30 Mar 2023 Smartlead: 10 mailboxes, 30 new leads each, 300 emails a day, two follow-ups three days apart, sent as plain text with tracking off. Jun 2023 From 25 Zoho mailboxes, two per domain at 50 a day, to 15 Office 365 mailboxes, one per domain at 90 a day. Both sets warmed at 40 a day, rising by 5. Apr 2024 Warmup Inbox scored our mailboxes 90 to 92 a week. May 2024 Fallbacks for empty merge fields (20 May), one shared inbox in Spark for replies (23 May), and a VA running the Debounce checks. The Smartlead upload sheet, Zoho mailboxes against Office 365, June 2023. Addresses and passwords masked. Recreated, original data private. The copy In 2021 every guest-post pitch used one agency template, written in March 2021 before I joined: \u0026ldquo;Hey [first name] | Quick question regarding [website]\u0026rdquo;, three emails, the same for every site. In 2023 I wrote eight new campaigns in Burkhard\u0026rsquo;s voice, since every email went out under his name, and he reviewed them in the docs.\nFour client campaigns (February to September 2023), each built on one idea: \u0026ldquo;HubSpot\u0026rdquo; (\u0026quot;[Name] - Is [Company] the next HubSpot? ($2.12B ARR)\u0026quot;), \u0026ldquo;Grocery Store\u0026rdquo; (\u0026ldquo;You\u0026rsquo;re inside the grocery store. You see oranges. Three options, [Name] - $5, $3, or $0 for a bag of oranges.\u0026rdquo;), \u0026ldquo;Traffic Projection\u0026rdquo;, offering a free traffic forecast, and \u0026ldquo;Clutch\u0026rdquo;, two versions for agencies listed on Clutch. Four guest-post campaigns (collected in one doc in October 2023, next to the 2021 template). Two open with a pun: \u0026ldquo;[first name] likes traffic like I like orange juice… …to be organic.\u0026rdquo; The last two changed the offer to free links from DR60–90+ sites plus a ready guest post, and added a way out: \u0026ldquo;If this isn\u0026rsquo;t for you just reply \u0026rsquo;not interested\u0026rsquo;.\u0026rdquo; Each campaign has a subject with the prospect\u0026rsquo;s name or company in it, one offer, proof in the follow-ups (Novum\u0026rsquo;s case studies, or links to our posts on Flippa, Sumo and Lucky Orange), and two to four follow-ups. On 9 October 2023 Burkhard flagged one of my subject lines with \u0026ldquo;@christian - wrong\u0026rdquo;.\nGuest-post Campaign 4, October 2023, merge fields marked. Recreated, original data private. Burkhard\u0026#39;s review comment on a subject line, 9 October 2023. Recreated, original data private. One campaign, end to end The agency campaign of 19 to 27 March 2024: 247 SEO agency leaders, one email with a personal video, seven replies.\n1. The sending domains Novum sent from lookalike domains such as novumhq.net, novumhq.co and novumhq.org, two mailboxes each, and logged inbox placement for every mailbox in the tracker. In the copy of 28 November 2023 the client mailboxes read \u0026ldquo;Not Spam\u0026rdquo; or 99%.\nThe sending-domain table in the agency tracker, November 2023. Mailbox names masked. Recreated, original data private. 2. Warmup and upkeep Smartlead kept every mailbox warming with automatic back-and-forth emails, tagged \u0026ldquo;wuI\u0026rdquo; in the subject. In March 2024 it also flagged four mailboxes that had dropped off, on 2, 4, 13 and 18 March, the weeks before the send.\nSmartlead disconnect alerts and warmup emails in my inbox, March to April 2024. Recreated, original data private. 3. The list, checked by Debounce On 18 March 2024 a list of agency leaders came out of Sales Navigator and went through Debounce the same day. Of 188 addresses, 124 came back safe to send, 56 risky (the domain accepts any address) and 8 unknown.\nThe Debounce result file, 18 March 2024. Counts are real, rows are sample data. Recreated, original data private. 4. A personal video for each lead We recorded one video and Sendspark made a personal link for each lead, with the prospect\u0026rsquo;s own website as the background. The upload file of 22 March held a video link per row.\nThe Sendspark upload: one personal video link per lead. Recreated, original data private. 5. The sequence One step, sent as Burkhard, asking for a reaction to the video and signing off with the lead\u0026rsquo;s city. The city field was blank for 132 of the 247 leads, so more than half the emails read \u0026ldquo;Sending love to ,\u0026rdquo;. On 20 May 2024 I asked Smartlead how to set a fallback for empty merge fields.\nThe blank sign-off, and Smartlead\u0026#39;s answer on fallbacks, May 2024. Recreated, original data private. The one email in the sequence, as the campaign export shows it. Recreated, original data private. 6. The send and the replies Smartlead sent all 247 emails from 19 to 27 March 2024. Seven people replied: five out-of-office notes, one \u0026ldquo;No thanks\u0026rdquo; and one who said he no longer read that inbox. None became a call. The video page logged 82 views from 19 March to 8 April, a few of them from outside this list, so more people watched than replied.\nThe campaign report, exported 28 March 2024. Totals and reply text are real. Recreated, original data private. The result The client outreach sent 19,081 cold emails between April 2022 and March 2023. Eight campaigns of new copy in 2023, four for clients and four for guest-post hosts. The setup kept needing care. Smartlead flagged disconnected mailboxes again from December 2023 to March 2024. What I\u0026rsquo;d change I\u0026rsquo;d build the sending setup first. The business ran on cold email, and when sending broke, revenue slowed with it. I\u0026rsquo;d also measure reply rates from day one, so I\u0026rsquo;d have the before and after in numbers.\n","date":"30 March 2023","permalink":"https://ccabaluna.pages.dev/work/novum-cold-email/","section":"Work","summary":"\u003ch2 id=\"the-problem\" class=\"relative group\"\u003eThe problem \u003c/h2\u003e\u003cp\u003eEvery link Novum sold and every client it won started with a cold email. I sent my first guest-post pitch on 14 June 2021, and by 2023 the agency sent from dozens of mailboxes across lookalike domains such as novumhq.net and novumhub.com.\u003c/p\u003e\n\u003cp\u003eIn March 2023 our warmup tool said our emails reached the inbox. I booked a one-hour consulting call with GlockApps, a deliverability testing company, on 13 March 2023. Two days later the expert sent his findings: \u0026ldquo;Your warmup tool says the emails are inboxing but the Glockapps tests show that they are landing in the spam folder.\u0026rdquo;\u003c/p\u003e","title":"Running an agency's cold email, 2021 to 2024"},{"content":"The problem Novum\u0026rsquo;s cold email was slow to win SEO clients, so we turned to Burkhard\u0026rsquo;s own LinkedIn for inbound leads. His account had been quiet, with only a few followers.\nWhat I did Wrote nearly all the posts in Burkhard\u0026rsquo;s voice: one-pagers and carousels on SEO, many reworked from the blog. Designed every carousel in Figma. He reviewed each one in Figma comments, 127 comment threads from 11 October 2023 to 11 June 2024, and sometimes rewrote a line himself. Kept the planning sheets: saved hooks, post ideas and the creators to tag. Answered comments and messages after each post, and commented on other creators\u0026rsquo; posts from Burkhard\u0026rsquo;s account (below). Old SEO vs New SEO, post 3 in the same pitch, listed at 80,000 impressions. Tagging and commenting Besides posting, I worked other creators\u0026rsquo; comment sections, so their readers found Burkhard.\nWho to tag, per topic. The \u0026ldquo;Creator Tagging\u0026rdquo; tab of my planning sheet (October to November 2023) named 16 post topics, from \u0026ldquo;21 Amazing SEOs\u0026rdquo; to \u0026ldquo;SEO tech stack\u0026rdquo; and \u0026ldquo;Topical Authority\u0026rdquo;, each with the 10 to 24 SEO creators to tag in it. One of those posts, a carousel saved on 4 October 2023, thanked 21 SEO creators by name and tagged each one. Hooks that worked. A \u0026ldquo;Saved Posts\u0026rdquo; tab kept other creators\u0026rsquo; posts with their category, format and hook, for example a 15-step on-page SEO checklist carousel. The daily routine, as I described it in June 2025: reply to comments for at least an hour after posting, comment on other people\u0026rsquo;s posts before posting, and write every comment myself, with no AI. Comments for Genius, then leads. In February 2024, for Burkhard\u0026rsquo;s new recruiting company Genius, I listed 17 creators to comment on and commented on their posts from his account. The people commenting there went into a sheet, 160 of them, as a list of people to reach for Genius. The creator-tagging tab: one column per post topic, October 2023. Recreated. One post, from idea to 300,000 impressions How the \u0026ldquo;SEO KPI Tracker 2024\u0026rdquo; one-pager went from a planning sheet to the best result in the April 2024 sponsor pitch.\n1. The planning sheet I started \u0026ldquo;Novum LinkedIn Resources\u0026rdquo; on 20 October 2023. One tab listed post ideas, each with the SEO creators we could tag. Another saved hooks from posts that had done well.\nThe planning sheet I kept for Burkhard\u0026#39;s posts, from 20 October 2023. Recreated, original data private. 2. The giveaway first The post offered a free SEO KPI tracker, so I built the sheet before the design, on 5 February 2024. Burkhard left 10 comments on it that evening, from \u0026ldquo;must be formula\u0026rdquo; to \u0026ldquo;if things are negative, it needs to turn red\u0026rdquo;. I fixed each one.\nThe KPI tracker readers would copy, with Burkhard\u0026#39;s review comments, 5 February 2024. Recreated, original data private. 3. Burkhard\u0026rsquo;s notes in Figma On 6 February he reviewed my drafts in the LinkedIn Posts file. On one he wrote: \u0026ldquo;promote SEO KPI tracker here. take care of conditional formatting before making screenshot. adjust wording as well.\u0026rdquo; When I didn\u0026rsquo;t understand a note, I asked.\nBurkhard\u0026#39;s Figma comments on my drafts, 6 February 2024. Recreated, original data private. 4. The final one-pager The finished design lists 24 SEO metrics, each with what it measures and the tool to track it. This version in the Figma file names the sponsor, is*hosting, as the page speed tool.\nMy SEO KPI Tracker 2024, the final frame in the team\u0026#39;s Figma file. 5. The numbers On 26 April 2024 the sponsor pitch to is*hosting listed four of my carousels. The KPI tracker post led with 300,000 impressions, 1,800 likes, 437 reposts and 56 comments.\nThe four posts in the is*hosting sponsor doc, 26 April 2024. Recreated, original data private. The result 40 posts in 8 months, busiest from November 2023 to March 2024. A sponsor pitch in April 2024 lists four of the carousels. The best one reached 300,000 impressions, 1,800 likes and 437 reposts. The account grew from a near-empty start to about 12,000 followers, by my own count. No follower export from the time survives. It brought in about 5 qualified inbound leads a month, also by my own count. What I learned I started without knowing how to write for LinkedIn, and my wording was often too complicated. Burkhard\u0026rsquo;s comments in Figma did most of the teaching, such as \u0026ldquo;way too complicated phrased\u0026rdquo; and \u0026ldquo;call it \u0026lsquo;Before ChatGPT\u0026rsquo; and \u0026lsquo;After ChatGPT\u0026rsquo;\u0026rdquo;.\n","date":"11 June 2024","permalink":"https://ccabaluna.pages.dev/work/novum-ceo-linkedin/","section":"Work","summary":"\u003ch2 id=\"the-problem\" class=\"relative group\"\u003eThe problem \u003c/h2\u003e\u003cp\u003eNovum\u0026rsquo;s cold email was slow to win SEO clients, so we turned to Burkhard\u0026rsquo;s own LinkedIn for inbound leads. His account had been quiet, with only a few followers.\u003c/p\u003e\n\u003ch2 id=\"what-i-did\" class=\"relative group\"\u003eWhat I did \u003c/h2\u003e\u003cul\u003e\n\u003cli\u003e\u003cstrong\u003eWrote nearly all the posts\u003c/strong\u003e in Burkhard\u0026rsquo;s voice: one-pagers and carousels on SEO, many reworked from the blog.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eDesigned every carousel in Figma.\u003c/strong\u003e He reviewed each one in Figma comments, 127 comment threads from 11 October 2023 to 11 June 2024, and sometimes rewrote a line himself.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eKept the planning sheets:\u003c/strong\u003e saved hooks, post ideas and the creators to tag.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eAnswered comments and messages\u003c/strong\u003e after each post, and commented on other creators\u0026rsquo; posts from Burkhard\u0026rsquo;s account (below).\u003c/li\u003e\n\u003c/ul\u003e\n\n\n\n\n  \n\n  \n  \n  \u003cfigure class=\"work-shot work-shot-tall\"\u003e\n    \u003ca href=\"/work/novum-ceo-linkedin/old-vs-new-seo.webp\" aria-label=\"Open full size: Old SEO vs New SEO, post 3 in the same pitch, listed at 80,000 impressions.\"\u003e\u003cimg src=\"/work/novum-ceo-linkedin/old-vs-new-seo_hu_df3ba3215161df7f.webp\" width=\"1200\" height=\"1500\" alt=\"Old SEO vs New SEO, post 3 in the same pitch, listed at 80,000 impressions.\" loading=\"lazy\"\u003e\u003c/a\u003e\n    \u003cfigcaption\u003eOld SEO vs New SEO, post 3 in the same pitch, listed at 80,000 impressions.\u003c/figcaption\u003e\n  \u003c/figure\u003e\n\n\n\n\n\u003ch2 id=\"tagging-and-commenting\" class=\"relative group\"\u003eTagging and commenting \u003c/h2\u003e\u003cp\u003eBesides posting, I worked other creators\u0026rsquo; comment sections, so their readers found Burkhard.\u003c/p\u003e","title":"Running the CEO's LinkedIn: 40 posts in 8 months"},{"content":"The problem Links are what move a site up in Google, and Novum won them by pitching guest posts one at a time. For Genius, Burkhard\u0026rsquo;s recruiting company, we wanted links that came without a pitch: pages that writers would cite on their own.\nWhat I did Wrote 22 statistics posts for joingenius.com, published from 24 April to 26 September 2024, such as creator economy, hiring and outsourcing statistics. They run under a colleague\u0026rsquo;s byline. Followed Burkhard\u0026rsquo;s playbook (May 2024): answer the question a writer searches for at the top, with a chart, and keep only stats worth citing. Used Claude for intros, meta descriptions and excerpts. Handed the writing to Sai in late October 2024. His first post went live on 30 October. I reviewed his drafts, outlined new posts and published them through December. One post, from playbook to 10 linking domains How \u0026ldquo;Outsourcing Cost Savings: 29+ Compelling Statistics\u0026rdquo; went from Burkhard\u0026rsquo;s rules to the most-linked page in the set.\n1. The rules Burkhard shared \u0026ldquo;SOP File Genius Statistics Creation\u0026rdquo; on 2 May 2024. It asked for the answer to a writer\u0026rsquo;s search at the top with a visual, stats worth citing, and a checklist of keywords, tables, an FAQ and search-volume charts.\nExcerpts from Burkhard\u0026#39;s statistics playbook, 2 May 2024. Recreated, original data private. 2. The draft and the review I started the draft on 14 May 2024 and exported the backlinks of a competing outsourcing-stats page two days later to see who links to these posts. Burkhard reviewed it from 7 to 13 June and left 8 comments. The one that changed the most: \u0026ldquo;always clarify the timeline. otherwise it\u0026rsquo;s not helpful and people won\u0026rsquo;t link to it if they don\u0026rsquo;t understand it.\u0026rdquo;\nThe draft with Burkhard\u0026#39;s comments, June 2024. Recreated, original data private. 3. The chart I built the key-stats chart in Figma, the visual the playbook asked for at the top. The file holds two versions; one went live.\n4. Published The post went live on joingenius.com on 16 June 2024, under the IG Rosales byline like the rest of the set.\n5. The links On 29 October 2024 I exported Ahrefs data for every page under /statistics/. This one led with 10 referring domains and 24 linking pages. Nobody pitched it.\nThe Ahrefs export of 29 October 2024, top rows. Recreated, original data private. The result By 29 October 2024 the posts I wrote had 30 linking domains, summed across 21 pages in an Ahrefs export, so a site linking to two pages counts twice. Ten pages had none yet. Nobody pitched these pages, so every link was earned. Sai went on to write 56 more by April 2025. ","date":"29 October 2024","permalink":"https://ccabaluna.pages.dev/work/genius-statistics-posts/","section":"Work","summary":"\u003ch2 id=\"the-problem\" class=\"relative group\"\u003eThe problem \u003c/h2\u003e\u003cp\u003eLinks are what move a site up in Google, and Novum won them by pitching guest posts one at a time. For Genius, Burkhard\u0026rsquo;s recruiting company, we wanted links that came without a pitch: pages that writers would cite on their own.\u003c/p\u003e\n\u003ch2 id=\"what-i-did\" class=\"relative group\"\u003eWhat I did \u003c/h2\u003e\u003cul\u003e\n\u003cli\u003e\u003cstrong\u003eWrote 22 statistics posts\u003c/strong\u003e for joingenius.com, published from 24 April to 26 September 2024, such as creator economy, hiring and outsourcing statistics. They run under a colleague\u0026rsquo;s byline.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eFollowed Burkhard\u0026rsquo;s playbook\u003c/strong\u003e (May 2024): answer the question a writer searches for at the top, with a chart, and keep only stats worth citing.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eUsed Claude\u003c/strong\u003e for intros, meta descriptions and excerpts.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eHanded the writing to Sai\u003c/strong\u003e in late October 2024. His first post went live on 30 October. I reviewed his drafts, outlined new posts and published them through December.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c!-- trace --\u003e\n\u003ch2 id=\"one-post-from-playbook-to-10-linking-domains\" class=\"relative group\"\u003eOne post, from playbook to 10 linking domains \u003c/h2\u003e\u003cp\u003eHow \u0026ldquo;Outsourcing Cost Savings: 29+ Compelling Statistics\u0026rdquo; went from Burkhard\u0026rsquo;s rules to the most-linked page in the set.\u003c/p\u003e","title":"Statistics posts that earned links without outreach"},{"content":"The problem Novum Plus was Burkhard\u0026rsquo;s done-for-you SEO service: keyword research, content, links and technical fixes for startups and small businesses. Each client needed a plan before signing, and a monthly invoice with proof of work after.\nWhat I did Audits for prospects. I wrote 46 \u0026ldquo;Top10 Analysis\u0026rdquo; docs (October to November 2022) and 78 \u0026ldquo;Biggest Organic Growth Levers\u0026rdquo; docs (November 2022 to April 2024), plus 12 free-trial decks in September and October 2023. Burkhard left review notes on many of them. I also audited paying clients. Traffic projections. I built the projection template in September 2021, which showed a prospect what their traffic could look like. About 20 projections followed. Keyword research and content plans. I built keyword maps for Vector and for Novum\u0026rsquo;s own site, chose topics and set up the content calendars, and IG Rosales and the writers wrote the articles. Every topic batch mixed ten article types, per the agency\u0026rsquo;s SOP of May 2023: three local SEO, two how-to, one guide, one idea, one hub-and-spoke, one low-difficulty and one wildcard. Rank tracking. I exported Estuary\u0026rsquo;s keyword rankings from Ahrefs on its invoice days in 2023. Account management. From March 2023 I invoiced at least 17 clients and sent each one a content calendar and a list of new live links with every invoice. My first note to Aloa on 1 March 2023: \u0026ldquo;My name is Christian and I\u0026rsquo;m the Operations Manager for Novum.\u0026rdquo; Links. As a link builder I placed links for Aloa from December 2021 to January 2022. Clients About 26 SEO clients over the years, among them Aloa, Estuary, Whop, Vector, Catena, Trace, TimePlus, Prosple, Zirtual, Suburban K9, Cleaning Business Academy, The Podcast Host, vpnoverview.org, Green Supply, 3wishes, CloudTalk, CartStack, Slidebean and Clique Studios. The work and my invoicing ran into 2025.\nKeyword research: Vector Vector, a logistics software company, signed with Novum on 25 April 2023. Two weeks before, from 10 to 12 April, I built its topic map in Google Sheets:\nWho reads it. A \u0026ldquo;thought dump\u0026rdquo; of the audience (\u0026ldquo;Managers, Directors, VPs in Logistics\u0026rdquo;) and the questions they ask, such as \u0026ldquo;What is a yard management system\u0026rdquo;. The raw keywords. 895 keywords from Ahrefs, plus Google Keyword Planner pulls for yard management, route optimization and warehouse management. The tree. Each keyword placed under a hub: supply chain, then warehousing, then warehouse management, then the keywords for one article. Burkhard reviewed the hub tab in June and marked it \u0026ldquo;wrong format\u0026rdquo;.\nVector\u0026#39;s topic tree and hub tab, April 2023. Recreated, original data private. Rank tracking: Estuary For Estuary, a data streaming company, I exported its Ahrefs rank tracker on 15 August, 3 September and 3 October 2023, the days I sent its invoices. Each export listed 164 keywords with their position, previous position and ranking page.\nEstuary\u0026#39;s rank tracker export, position changes, autumn 2023. Recreated, original data private. Genius\u0026rsquo;s own site, 2024 The same team ran SEO for Genius, Burkhard\u0026rsquo;s recruiting company, on joingenius.com.\nPages from templates. I set up the Multiple Pages Generator plugin in WordPress so one template plus a list of roles made one page per role, under /job-description/ and /interview-questions/. When I checked on 10 November 2026, the two hub pages linked 29 job descriptions and 111 interview-question pages. No traffic figures for them survive. A PDF download. I wrote a PHP snippet to turn these pages into PDFs (github.com/ck123cabz/download-as-pdf). On the template pages it failed, along with every PDF plugin I tried. I worked the bug with the plugin\u0026rsquo;s maker, Themeisle, from 24 October 2024: a plugin update fixed it, a later link fix broke it again, and on 25 October I wrote \u0026ldquo;I spent days trying to isolate the issue!\u0026rdquo; Their patch arrived on 1 November. The support thread with Themeisle, 24 October to 1 November 2024, condensed. Recreated, original data private. A competitor\u0026rsquo;s site, scored. On 6 November 2024 I scored 1,742 pages from aihr.com, a large HR site, from 0 to 100 by their Ahrefs traffic change and topic, to decide what Genius should write first. Writing. I wrote posts for the blog, among them 22 statistics posts built to earn links on their own. The interview-questions hub on joingenius.com, 10 November 2026. The scored competitor sheet, top rows, 6 November 2024. Recreated. Two traces: what a client got, and a prospect audit What went to Aloa with my first invoice, then one \u0026ldquo;Biggest Organic Growth Levers\u0026rdquo; audit that turned a prospect, Catena, into a client.\nWhat a client got each month: Aloa 1. My first invoice email On 1 March 2023 I wrote to Aloa: \u0026ldquo;My name is Christian and I\u0026rsquo;m the Operations Manager for Novum. I\u0026rsquo;ll be handling the invoicing moving forward.\u0026rdquo; The email carried the FreshBooks invoice and two PDFs as proof of work: the content calendar and the backlink tracker for February.\nMy first invoice email to Aloa, 1 March 2023. Prices removed. Recreated, original data private. 2. The content calendar Every article sat in one row of Aloa\u0026rsquo;s sheet with its search term, title, status, writer and the invoice it was billed on. The article due 28 February, a guide to hiring embedded software engineers, carries invoice #014.\nTwo rows of Aloa\u0026#39;s content calendar, January and February 2023 invoices. Recreated, original data private. 3. The backlink tracker Links had the same treatment: publication, topic, target page, anchor, status and invoice number. The rows shown are my own, from when I built Aloa\u0026rsquo;s first links in December 2021 and January 2022.\nAloa\u0026#39;s backlink tracker, rows from my link-building months. Recreated, original data private. One growth-levers audit: Catena 4. The draft On 25 May 2023 I copied our audit template for withcatena.com, a remote-hiring company with no blog yet. I filled six levers from Ahrefs and Frase: missing content, A+ content, hub and spoke, digital PR, a Domain Rating of 0.7 and slow pages. The comments are my own notes from the template.\nMy growth-levers draft for withcatena.com, 25 May 2023. Recreated, original data private. 5. Burkhard\u0026rsquo;s review At 1:04 am on 27 May Burkhard reviewed it: \u0026ldquo;We need to pick better topics here. They need to be more specific regarding PH and have more search volume.\u0026rdquo; He also fixed a link I had put on an empty space and tightened the Domain Rating line.\nBurkhard\u0026#39;s comment and suggestions on the Catena draft, 27 May 2023. Recreated, original data private. 6. The proposal Proposal #149 for withcatena.com is dated 26 May 2023. Its content slide keeps the two Philippines-specific topics and drops the generic one, then sets out three packages. Catena signed, and I sent its first invoice on 10 June 2023.\nProposal #149 for withcatena.com, May 2023. Prices left off. Recreated, original data private. The result Novum reported that Aloa\u0026rsquo;s traffic value grew from $6,819 a month to $172,289 a month (Novum\u0026rsquo;s own case study, October 2023). That is the agency\u0026rsquo;s figure for the team\u0026rsquo;s work, not mine alone.\n","date":"8 April 2024","permalink":"https://ccabaluna.pages.dev/work/novum-seo-client-work/","section":"Work","summary":"\u003ch2 id=\"the-problem\" class=\"relative group\"\u003eThe problem \u003c/h2\u003e\u003cp\u003eNovum Plus was Burkhard\u0026rsquo;s done-for-you SEO service: keyword research, content, links and technical fixes for startups and small businesses. Each client needed a plan before signing, and a monthly invoice with proof of work after.\u003c/p\u003e\n\u003ch2 id=\"what-i-did\" class=\"relative group\"\u003eWhat I did \u003c/h2\u003e\u003cul\u003e\n\u003cli\u003e\u003cstrong\u003eAudits for prospects.\u003c/strong\u003e I wrote 46 \u0026ldquo;Top10 Analysis\u0026rdquo; docs (October to November 2022) and 78 \u0026ldquo;Biggest Organic Growth Levers\u0026rdquo; docs (November 2022 to April 2024), plus 12 free-trial decks in September and October 2023. Burkhard left review notes on many of them. I also audited paying clients.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eTraffic projections.\u003c/strong\u003e I built the projection template in September 2021, which showed a prospect what their traffic could look like. About 20 projections followed.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eKeyword research and content plans.\u003c/strong\u003e I built keyword maps for Vector and for Novum\u0026rsquo;s own site, chose topics and set up the content calendars, and IG Rosales and the writers wrote the articles. Every topic batch mixed ten article types, per the agency\u0026rsquo;s SOP of May 2023: three local SEO, two how-to, one guide, one idea, one hub-and-spoke, one low-difficulty and one wildcard.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eRank tracking.\u003c/strong\u003e I exported Estuary\u0026rsquo;s keyword rankings from Ahrefs on its invoice days in 2023.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eAccount management.\u003c/strong\u003e From March 2023 I invoiced at least 17 clients and sent each one a content calendar and a list of new live links with every invoice. My first note to Aloa on 1 March 2023: \u0026ldquo;My name is Christian and I\u0026rsquo;m the Operations Manager for Novum.\u0026rdquo;\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eLinks.\u003c/strong\u003e As a link builder I placed links for Aloa from December 2021 to January 2022.\u003c/li\u003e\n\u003c/ul\u003e\n\u003ch2 id=\"clients\" class=\"relative group\"\u003eClients \u003c/h2\u003e\u003cp\u003eAbout 26 SEO clients over the years, among them Aloa, Estuary, Whop, Vector, Catena, Trace, TimePlus, Prosple, Zirtual, Suburban K9, Cleaning Business Academy, The Podcast Host, vpnoverview.org, Green Supply, 3wishes, CloudTalk, CartStack, Slidebean and Clique Studios. The work and my invoicing ran into 2025.\u003c/p\u003e","title":"Running the client side of an SEO agency"},{"content":"The problem Smash Digital, an SEO agency, ordered links from Novum every month and resold them to its own clients. Nick Kanter, its Director of Operations, tracked every order in a shared sheet. Each month the links ordered, delivered and billed had to match, and in 2021 and 2022 they often didn\u0026rsquo;t: links went to the wrong page, reports came late, and some links were placed in posts the client called \u0026ldquo;unnaturally stuffed\u0026rdquo;.\nWhat I did Placed links myself in early 2022, as the only Novum person working the client\u0026rsquo;s sheet. Ran the monthly close from May 2022 to November 2025: matched our internal sheet against Nick\u0026rsquo;s, agreed the count with him and sent the invoice. Answered every complaint with a fix. On 2 October 2022 Nick listed four problems. I added a double-check before any link moved to his sheet, a target-page check to the proofreading stage and more varied anchor text. Set the buffer. In my words to Nick that day: \u0026ldquo;We always try to incorporate at least 200% to at least take into account all the factors (publication delays, cancellations, no-follows).\u0026rdquo; The team pitched about twice the links ordered, so each month\u0026rsquo;s order filled. Managed the builders. I handed day-to-day link work to Renz in November 2022 to help Burkhard win clients, and rejoined the link team in June 2024. One monthly close: October 2022 How the September 2022 links got counted, disputed, fixed and billed, from my first email on 2 October to the paid invoice of the next month.\n1. I ask before I bill At 2:09 am on 2 October 2022 I asked Nick whether September\u0026rsquo;s links were good to go, so the invoice would match his sheet before it went out.\nMy check-in before the September invoice, 2 October 2022. Recreated, original data private. 2. Nick\u0026rsquo;s count and his four problems At 11:33 am Nick replied with 35 links for September. We had over-delivered for four of his clients, by up to seven links over the order, so he marked the extras \u0026ldquo;do not include for payment\u0026rdquo;. He also listed four mistakes: late reports, wrong target URLs on the sheet, repeated anchors and links to pages nobody ordered.\nNick Kanter\u0026#39;s reply of 2 October 2022, shortened. Client names changed. Recreated, original data private. 3. The fixes and the 200% buffer At 4:20 pm I answered each point. I explained the buffer, \u0026ldquo;We always try to incorporate at least 200% to at least take into account all the factors (publication delays, cancellations, no-follows)\u0026rdquo;, and gave a fix per problem: a double-check of our sheet before links move to his, anchor variety, and target pages checked at the proofreading stage. Then I asked him to add one link to the invoice that his count had left out.\nMy reply with the 200% buffer and the fixes, 2 October 2022. Recreated, original data private. 4. The matched count The shared tracking sheet carries an invoice label on every link, and when the September block closed it held 36 links marked \u0026ldquo;September Invoice\u0026rdquo;: Nick\u0026rsquo;s 35 plus the one I asked for. Ten more were marked \u0026ldquo;do not include for payment\u0026rdquo;, one was rejected for pointing at a privacy page, and two were later moved into October.\nThe Link Tracking tab, September 2022 block. Counts and labels are real, target sites changed. Recreated, original data private. 5. The invoice At 4:29 pm I sent invoice 2022_120 from FreshBooks with two PDFs: the invoice and the list of links behind it. Nick passed it on for payment that evening.\nThe September invoice email and Nick\u0026#39;s reply, 2 October 2022. Recreated, original data private. 6. The excess comes back On 24 October Nick released an excess September link into the October invoice, and a second one the next day. I billed October on 1 November as invoice 2022_126, and Nick confirmed payment on 8 November.\nThe excess-links note, the October invoice and the payment confirmation, October to November 2022. Recreated, original data private. The result 1,339 billable links went out in my time on the account, from 2 to 95 a month, at $300 a link. Rejected or excess links fell from 80 in 2021 and 39 in 2022 to 9 in 2023, and 1 to 4 a year after that. Wrong pages and late reports came back a few times in early 2023, and two of my invoices came out short and needed correcting. The client stayed. Nick, December 2025: \u0026ldquo;Looking forward to continuing to work with you guys in 2026.\u0026rdquo; ","date":"4 November 2025","permalink":"https://ccabaluna.pages.dev/work/smash-digital-link-fulfilment/","section":"Work","summary":"\u003ch2 id=\"the-problem\" class=\"relative group\"\u003eThe problem \u003c/h2\u003e\u003cp\u003eSmash Digital, an SEO agency, ordered links from Novum every month and resold them to its own clients. Nick Kanter, its Director of Operations, tracked every order in a shared sheet. Each month the links ordered, delivered and billed had to match, and in 2021 and 2022 they often didn\u0026rsquo;t: links went to the wrong page, reports came late, and some links were placed in posts the client called \u0026ldquo;unnaturally stuffed\u0026rdquo;.\u003c/p\u003e","title":"1,339 links for a reseller agency, counted and billed every month"},{"content":"The problem Novum sold backlinks to clients: links from other sites that help a page rank in Google. Each link came from a guest post, an article we wrote for another site\u0026rsquo;s blog with the client\u0026rsquo;s link inside it. So every link meant finding a site, reaching its editor, agreeing a topic, writing the post, getting the link in and proving it to the client.\nThe seven steps 1. Find the sites Each Pitchbox campaign covered one niche and kept only sites with an Ahrefs Domain Rating (DR) of 50 to 95. DR is Ahrefs\u0026rsquo; 0 to 100 score of a site\u0026rsquo;s authority. The SOP of 22 June 2021 capped each campaign at 150 emails a day.\n2. Never pitch a site twice On 3 March 2022 I started a ledger in Google Sheets of every domain we had ever pitched: batch, root domain, date and a numbered status such as \u0026ldquo;[3] Outreached\u0026rdquo; or \u0026ldquo;[5] No Info Found\u0026rdquo;. I loaded our Pitchbox history into it first. It held about 11,500 domains by the end of 2022, plus a list of sites that had asked not to be contacted again.\nThe ledger\u0026#39;s database tab, 2022. Domains are made up. Recreated, original data private. 3. Find the editor I found editors\u0026rsquo; emails with Hunter from June 2021. From September 2022 every batch went through Debounce before sending. In March 2023 I added an Apps Script email finder to the lead sheet.\n4. Pitch In 2021 I pitched in my own name, three emails per site. From 2023 the pitches went out as Burkhard from the agency\u0026rsquo;s sending domains, with copy I wrote in his voice ( the cold email page shows it). In 2024 I pitched for Genius as myself again, from reply templates I kept in Notion: \u0026ldquo;no direct link swaps - only ABC exchanges through guest posts\u0026rdquo;, meaning site A links to site B through a post on site C.\n5. Agree a topic and write A host who said yes got two to five topics, each with its target search term, volume and traffic potential from Ahrefs. The writers drafted, IG Rosales edited, and each post moved across our Trello \u0026ldquo;Guest Posts\u0026rdquo; board from In Progress to Proofreading \u0026amp; Editing, Done and Submitted. In 2024 I graded my own drafts to A+ in SurferSEO before sending.\n6. Get the link in, and check it A post counted only once the client\u0026rsquo;s link was live. When an editor cut a link, I moved it into another post and told the partner which one. In June 2021 I chased a site because \u0026ldquo;one of my links (3wishes) was not added\u0026rdquo;.\n7. Prove it to the client Every live link went into the agency\u0026rsquo;s Backlink Tracker with its publication, DR, topic, status, link builder, live date and a column per client. It logs 1,091 live placements across the team, and from March 2023 each client got a PDF of its new links with every invoice.\nThe tracker\u0026#39;s status columns. Client names masked. Recreated, original data private. Buying and trading links Not every host took a free post. In October 2022 I priced 37 sites that sold posts, at $50 to $140 each. When a site asked for money in 2021, I offered a trade first: \u0026ldquo;I usually don\u0026rsquo;t do sponsored/paid posts. Would you be interested in a backlink exchange instead?\u0026rdquo; (5 August 2021).\nRunning it as a team Four new link builders joined on my invite on 5 July 2022, and they ran the same steps. On 12 November 2022 I handed day-to-day link work to Renz. By March 2023 we tracked the team on a KPI sheet with one tab per person, 13 tabs, and reviewed 14 partner email threads line by line for what could have been done better (\u0026ldquo;10 days is too long.\u0026rdquo;).\nThe result About 11,500 domains in my ledger by the end of 2022, so no site was pitched twice. 1,091 live placements in the agency\u0026rsquo;s tracker across the team. The tracker credits me with about 129 live posts, many carrying several client links. My own placements went from 30 to 100+ a month in my first year, by my own count. What I\u0026rsquo;d change I\u0026rsquo;d record who placed each link from the first day. The tracker\u0026rsquo;s link-builder column is the only record of who did what, and it was often left blank.\n","date":"18 October 2024","permalink":"https://ccabaluna.pages.dev/work/novum-link-building-process/","section":"Work","summary":"\u003ch2 id=\"the-problem\" class=\"relative group\"\u003eThe problem \u003c/h2\u003e\u003cp\u003eNovum sold backlinks to clients: links from other sites that help a page rank in Google. Each link came from a guest post, an article we wrote for another site\u0026rsquo;s blog with the client\u0026rsquo;s link inside it. So every link meant finding a site, reaching its editor, agreeing a topic, writing the post, getting the link in and proving it to the client.\u003c/p\u003e\n\u003ch2 id=\"the-seven-steps\" class=\"relative group\"\u003eThe seven steps \u003c/h2\u003e\u003ch3 id=\"1-find-the-sites\" class=\"relative group\"\u003e1. Find the sites \u003c/h3\u003e\u003cp\u003eEach Pitchbox campaign covered one niche and kept only sites with an Ahrefs Domain Rating (DR) of 50 to 95. DR is Ahrefs\u0026rsquo; 0 to 100 score of a site\u0026rsquo;s authority. The SOP of 22 June 2021 capped each campaign at 150 emails a day.\u003c/p\u003e","title":"How a guest post got placed, from site list to live link"},{"content":"The problem Burkhard wanted more clients for Novum Plus, the agency\u0026rsquo;s done-for-you SEO service, and the agency already knew how to send cold email at scale from its link building.\nWhat I built Lead lists from Sales Navigator and other sources, kept in Google Sheets. The sending setup, rebuilt on Smartlead in 2023 ( the cold email case study). The emails. Every sequence opened with a free traffic projection for the prospect\u0026rsquo;s site: \u0026ldquo;Traffic Projection for [company] (372% Growth)\u0026rdquo;. A free audit for every reply. Prospects who answered got a Top10 analysis or a \u0026ldquo;Biggest Organic Growth Levers\u0026rdquo; doc, then a numbered proposal. The people. I hired and screened sales help in August and September 2022, and ran some of the sales calls. From cold email to kickoff: Zirtual Zirtual is the one client whose whole path, from first cold email to first invoice, is on record.\n1. The cold email On 9 November 2022 one of our six campaigns, \u0026ldquo;Sai3\u0026rdquo;, sent the \u0026ldquo;Traffic Projection\u0026rdquo; email to the chief marketing officer of Startups.com, Zirtual\u0026rsquo;s parent company. It was one of 238 sends that day. The send log shows his reply the same afternoon, about seven hours later.\nThe send log row for the cold email, 9 November 2022. Contact masked, other rows are sample data. Recreated, original data private. 2. The first free audit The next day, 10 November, I wrote a Top10 analysis of startups.com, with a short walkthrough video. It led with four growth levers: content the market searches for, links to money pages, A+ content and lead nurturing.\nTop10 analysis of startups.com, 10 November 2022. Recreated, original data private. 3. The audit for Zirtual On 21 November I wrote a second Top10 analysis, for zirtual.com itself. It found a Domain Rating of 71, a search term with 3,700 visits a month in traffic potential, and blog posts scoring 30 out of 100 for content.\nTop10 analysis of zirtual.com, 21 November 2022. Recreated, original data private. 4. Onboarding questions On 7 and 8 December Zirtual answered our questions doc: their competitors, the clients they serve and the topics they wanted covered. Their folder in our client Drive opened the same day as \u0026ldquo;007 zirtual.com\u0026rdquo;.\nZirtual\u0026#39;s answers to our onboarding questions, December 2022. Recreated, original data private. 5. Kickoff On 3 January 2023 I wrote: \u0026ldquo;As discussed, we\u0026rsquo;ve officially kickstarted Zirtual\u0026rsquo;s SEO campaign.\u0026rdquo; The email carried the first invoice and the first content calendar and backlink tracker. Three days later Zirtual sent a test transfer before paying.\nThe kickoff email and the test-transfer note, January 2023. Recreated, original data private. The result The campaigns sent 19,081 cold emails in 11 months, and the agency wrote about 46 proposals. Eight clients signed in 2023: Zirtual, Cyngn, Vector, TimePlus, Foreignerds, Catena, Trace and Prosple. Zirtual shows the whole path. A cold email went out on 9 November 2022 and the reply came the same day. Two audits followed, and on 3 January 2023 I wrote: \u0026ldquo;As discussed, we\u0026rsquo;ve officially kickstarted Zirtual\u0026rsquo;s SEO campaign.\u0026rdquo; By 2024 demand for SEO had cooled, free trials stopped converting, and the agency moved its effort to Genius.\n","date":"22 December 2023","permalink":"https://ccabaluna.pages.dev/work/novum-client-acquisition/","section":"Work","summary":"\u003ch2 id=\"the-problem\" class=\"relative group\"\u003eThe problem \u003c/h2\u003e\u003cp\u003eBurkhard wanted more clients for Novum Plus, the agency\u0026rsquo;s done-for-you SEO service, and the agency already knew how to send cold email at scale from its link building.\u003c/p\u003e\n\u003ch2 id=\"what-i-built\" class=\"relative group\"\u003eWhat I built \u003c/h2\u003e\u003cul\u003e\n\u003cli\u003e\u003cstrong\u003eLead lists\u003c/strong\u003e from Sales Navigator and other sources, kept in Google Sheets.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eThe sending setup\u003c/strong\u003e, rebuilt on Smartlead in 2023 (\n      \n    \u003ca href=\"/work/novum-cold-email/\"\u003ethe cold email case study\u003c/a\u003e).\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eThe emails.\u003c/strong\u003e Every sequence opened with a free traffic projection for the prospect\u0026rsquo;s site: \u0026ldquo;Traffic Projection for [company] (372% Growth)\u0026rdquo;.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eA free audit for every reply.\u003c/strong\u003e Prospects who answered got a Top10 analysis or a \u0026ldquo;Biggest Organic Growth Levers\u0026rdquo; doc, then a numbered proposal.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eThe people.\u003c/strong\u003e I hired and screened sales help in August and September 2022, and ran some of the sales calls.\u003c/li\u003e\n\u003c/ul\u003e\n\u003c!-- trace --\u003e\n\u003ch2 id=\"from-cold-email-to-kickoff-zirtual\" class=\"relative group\"\u003eFrom cold email to kickoff: Zirtual \u003c/h2\u003e\u003cp\u003eZirtual is the one client whose whole path, from first cold email to first invoice, is on record.\u003c/p\u003e","title":"Winning SEO clients with cold email and free audits"},{"content":"The problem My 2024 site-scoring tool worked, but it was one script on one server: one list at a time, no record of past jobs, and no way to see progress. I wanted to learn how a production system is built by rebuilding it properly.\nWhat I built A job queue (NestJS, BullMQ and Redis), so large lists run in batches and several jobs can wait their turn. A three-step filter. The first step drops sites by domain type, the second drops pure publications after fetching the page, and the third has an AI model (Gemini or OpenAI) read the page and decide. A live dashboard showing each job\u0026rsquo;s progress and queue position. CSV exports by decision, confidence and the step that dropped a site. Hosting on Railway, with the data in Supabase and pages fetched through ScrapingBee. How I built it I wrote the specs and broke them into stories and phases. Claude Code wrote most of the code, and I reviewed it, deployed it and fixed what broke. The repo has 147 commits from 13 October to 15 December 2025.\nThe same list through the rebuild The 33 sites from the 2024 tool\u0026rsquo;s page went through the rebuilt code on my Mac on 10 November 2026, with its real screens and a stand-in for the database and the AI.\n1. A new job I name the job and upload the same file, or paste the addresses. Layer 1 runs as the job is created, before anything is queued, because it needs no page fetch.\nThe real job form from the repo\u0026#39;s Next.js app, run locally on 10 Nov 2026. Nothing was sent to a live server. 2. The live dashboard The job page shows each running site and its layer, the share done, the rate, the time left, and running counts of approved, not suitable and dropped early. Up to 10 sites run at once.\nThe job page at 14 of 33. Real page and real results; the timing and the sites shown as running are mocked, because no queue or database was running. 3. The funnel Layer 1 dropped 3 sites on their domain type alone: wikipedia.org, mit.edu and usa.gov. Layer 2 fetched homepages and dropped 8: four read as pure publications, such as news.ycombinator.com and backlinko.com, and four that blocked the fetch. Layer 3 read the other 22 and approved 10. In 2024 GPT judged the bare address, so healthline.com went out on the word \u0026ldquo;health\u0026rdquo; in its name. Here the topic call waits for the page read.\nThe finished job and its three-layer breakdown. Real page; the Layer 1 and 2 counts are the real code\u0026#39;s results, the Layer 3 split and the costs come from the repo\u0026#39;s test mock. 4. Why a site was dropped Every row opens into its layer factors. news.ycombinator.com scored 0.88 on Layer 2\u0026rsquo;s publication check, above the 0.65 limit, so it stopped there and never cost an AI call.\nLayer 2 factors for news.ycombinator.com. Real page and real Layer 2 result from the repo\u0026#39;s code. 5. The AI read Sites that pass Layer 2 go to Gemini, with OpenAI as the fallback, and the row keeps the provider, model, cost, token count and the reason. A score of 0.8 or more is high confidence.\nLayer 3 factors for searchenginejournal.com. Real page; the verdict, provider, cost and reason come from the repo\u0026#39;s mock AI, not from Gemini. 6. Export The export lets the team pick the full file or a summary, or one layer\u0026rsquo;s factors, and filter by approval, layer or confidence.\nThe export dialog. Real page. 7. The CSV The summary export has one row per site with its status, confidence, the layer that dropped it and the reason. The team works from the approved rows.\nThe summary CSV from this run, shown as a table. Written by the repo\u0026#39;s own export code; Layer 3 verdicts and costs come from its test mock. The result It has run on real lists of tens of thousands of URLs. My last change, on 15 December 2025, cut how many ScrapingBee credits each page used.\n","date":"15 December 2025","permalink":"https://ccabaluna.pages.dev/work/url-analysis-rebuild/","section":"Work","summary":"\u003ch2 id=\"the-problem\" class=\"relative group\"\u003eThe problem \u003c/h2\u003e\u003cp\u003eMy \n      \n    \u003ca href=\"/work/novum-site-scoring-tool/\"\u003e2024 site-scoring tool\u003c/a\u003e worked, but it was one script on one server: one list at a time, no record of past jobs, and no way to see progress. I wanted to learn how a production system is built by rebuilding it properly.\u003c/p\u003e\n\u003ch2 id=\"what-i-built\" class=\"relative group\"\u003eWhat I built \u003c/h2\u003e\u003cul\u003e\n\u003cli\u003e\u003cstrong\u003eA job queue\u003c/strong\u003e (NestJS, BullMQ and Redis), so large lists run in batches and several jobs can wait their turn.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eA three-step filter.\u003c/strong\u003e The first step drops sites by domain type, the second drops pure publications after fetching the page, and the third has an AI model (Gemini or OpenAI) read the page and decide.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eA live dashboard\u003c/strong\u003e showing each job\u0026rsquo;s progress and queue position.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eCSV exports\u003c/strong\u003e by decision, confidence and the step that dropped a site.\u003c/li\u003e\n\u003cli\u003e\u003cstrong\u003eHosting\u003c/strong\u003e on Railway, with the data in Supabase and pages fetched through ScrapingBee.\u003c/li\u003e\n\u003c/ul\u003e\n\u003ch2 id=\"how-i-built-it\" class=\"relative group\"\u003eHow I built it \u003c/h2\u003e\u003cp\u003eI wrote the specs and broke them into stories and phases. Claude Code wrote most of the code, and I reviewed it, deployed it and fixed what broke. The repo has 147 commits from 13 October to 15 December 2025.\u003c/p\u003e","title":"Rebuilding my site-scoring tool as a production system"},{"content":"","date":null,"permalink":"https://ccabaluna.pages.dev/","section":"Christian Cabaluna","summary":"","title":"Christian Cabaluna"},{"content":"","date":null,"permalink":"https://ccabaluna.pages.dev/tags/claude-code/","section":"Tags","summary":"","title":"Claude Code"},{"content":"","date":null,"permalink":"https://ccabaluna.pages.dev/tags/hyperframes/","section":"Tags","summary":"","title":"HyperFrames"},{"content":"","date":null,"permalink":"https://ccabaluna.pages.dev/tags/motion-design/","section":"Tags","summary":"","title":"Motion Design"},{"content":"","date":null,"permalink":"https://ccabaluna.pages.dev/tags/","section":"Tags","summary":"","title":"Tags"},{"content":"Each case study shows what I owned, what I built, and the numbers that moved.\n","date":null,"permalink":"https://ccabaluna.pages.dev/work/","section":"Work","summary":"\u003cp\u003eEach case study shows what I owned, what I built, and the numbers that moved.\u003c/p\u003e","title":"Work"},{"content":"This page is for an owner deciding whether to hand me part of the business, with a link to the case study behind each claim.\nWhat you\u0026rsquo;ll see from me A written report at the end of every shift What got done, the numbers, and what\u0026rsquo;s blocked.\nMy daily work log for 24 June 2025, shortened, with made-up client names. Recreated, original data private A weekly report on your key numbers From August 2025 the Genius CEO got one from me every week.\nMy weekly report for 8 to 12 September 2025, money lines left out. Recreated, original data private One ticket per piece of work In Linear or the tool you already use, closed with what happened and what\u0026rsquo;s left.\nThe agents pane: running sessions and their plans, tied to Linear A written reason when a person is needed At Matchworks a held reply showed why it was held, so the reviewer read one line and not the whole thread.\nLeft, the send. Right, the held draft in the Raycast menu bar with its reason. Recreated, original data private. Written steps for anything I take over An SOP (standard operating procedure), so another person can run it without me. I wrote the guest-post SOP (June 2021) and the traffic projection SOP (November 2021) at Novum, and the client onboarding checklist (November 2024, reused for every placement after it) and the fulfilment email templates (January 2025) at Genius.\nThe first page of my guest-post SOP for Novum\u0026#39;s link builders, 22 June 2021. Recreated, original data private How I decide Do it by hand first. At Matchworks I read every reply draft myself until 28 July 2026, before any reply went out on its own. Reply system case study Grade what went out, not a status. From then on I read and graded the replies the system sent, one at a time, and marked 6 wrong of 324. Reply system case study How I take over a repetitive process Three steps, in this order:\nDo it by hand first, so I know where the hours go before I change anything. Automate the slowest step with the simplest tool that does the job. Keep the judgment calls with a person. 1 By hand 2 Automate 3 Person decides Novum Jun 2021 to Sep 2024 By handCold email from June 2021. Picking which sites to pitch took a day of research. AutomateSeptember 2024: a tool scraped each site and had GPT say whether to pitch it. Person decidesAssistants still picked the decision maker and found the email. A day of research: 8 hours to 2–3\nGenius Oct 2024 to 2025 By handSales calls to invoices, with a team I trained from January to May 2025. AutomateTypeform to Notion to Airtable through Zapier, then Relay. An AI screen marked each must-have. Person decidesWho to put in front of a client, the judgment the client paid for. Brand Designer search: 912 applied, 3 endorsed\nMatchworks Jul to Oct 2026 By handI read every reply draft before it went out, until 28 July 2026. AutomateFrom 29 July a reply sent on its own when its checks passed; the rest waited with a written reason. Person decidesI graded 324 sent replies one at a time in August. 6 were wrong. Median wait: 27.1 min to 5.6\nThree take-overs, step by step The grade on each reply, and what the held one changed. Recreated, original data private. Matchworks · Jul to Oct 2026\nReplies to prospects 27.1 → 5.6 min\nmedian wait for an answer, July to October\nI read every draft by hand until 28 July, then let it send when its checks passed and graded what went out: 6 wrong of 324.\nShow the 6 steps A prospect replies to a cold email.The AI reads the whole thread and drafts a reply.Each draft runs the send checks.It sends, or waits with a written reason.Every reply leaves an audit row.I grade what went out. Read the reply system case One candidate\u0026#39;s AI Screen in Airtable: each must-have marked pass or fail, with the evidence. Recreated, original data private Genius · 2025\nA recruiting search, intake to report 912 → 3\napplied to endorsed, Brand Designer search, Jun to Jul 2025\nRelay and Airtable ran intake and screening. I chose who the client met, and the CEO got the numbers every week.\nShow the 7 steps The client fills the intake form.Relay creates the Airtable client and project.I write the dream candidate profile.The job ad is written from the profile.The AI screen marks each must-have.I pick the finalists for the client.The CEO gets the weekly report. Read the Genius recruiting case What the repo says, run against the public repo. This site · 9 Nov 2026\nChecking my own claims 2023 → Sep 2024\nthe year I remembered for a tool, and the month its code shows\nBefore a claim goes out I check it against the record. Here the code overturned my answer and the page took the code\u0026#39;s date.\nShow the 7 steps 7:23 pm: I start a sweep of my notes.The index gives the date my own accounts gave the tool.9:41 pm: a question, and my first answer.The repo says September 2024.I correct my answer.The settled fact goes to one file.10 November: the page goes live with the repo\u0026#39;s date. Read the AI setup case A week with me Genius, 8 to 12 September 2025: a log every day, and the numbers to the CEO on Friday.\n8 Sep Daily log 9 Sep Daily log 10 Sep Daily log 11 Sep Daily log 12 Sep Daily log Weekly report 10leads 5calls booked 3calls held 6active searches 2hires from one lineup How I build software Questions I ask until every open decision is named. You make each call. Written plan The settled plan is written down before any code. You read it before work starts. One ticket per session The plan becomes tickets, and each working session takes one. You see progress ticket by ticket. A second reviewer An AI agent that didn\u0026#39;t write the code reviews it. I fix what it finds before I call it done. Test on a copy Tests run on a copy of your data, with the pass mark set before the test runs. Your live data is never the test. Two rules hold across every build. A job that runs twice by mistake does nothing twice, so no double emails and no double invoices. Each fact lives in one place, so two tools never disagree about it. The full method is on the AI setup page.\nYour first month with me Week 1: I do the work by hand, in the tools you already use, and write down each step. We pick the one number this month should move, and I record where it stands. Week 2: each process I take over gets an SOP (standard operating procedure), written steps your team can follow without me. Weeks 3–4: I automate the slowest step first. You decide which calls stay with a person. On day 30 you get that number next to where it stood in week 1.\nTerms and hours What I\u0026rsquo;m open to: a full-time remote role or a monthly retainer, from $1,500 a month, starting with the first month above. My day in the Philippines (UTC+8). In these hours I reply within 30 minutes. 12 am6 amnoon6 pm12 am US morning 9:00 am to 1:00 pm Eastern 10:00 pm to 2:00 am my time Sydney business day 9:00 am to 5:00 pm Sydney 6:00 am to 2:00 pm my time, October to April; 7:00 am to 3:00 pm the rest of the year Tools I work in Linear for tasks, Claude Code for builds, Airtable, Notion and Google Sheets for operations data, Pipedrive for a client pipeline, and Smartlead for cold email.\nEmail me the process you’d hand off first. In my working hours I reply within 30 minutes. Email me ckwater123@gmail.com Copy email address LinkedIn ","date":null,"permalink":"https://ccabaluna.pages.dev/workstyle/","section":"Christian Cabaluna","summary":"\u003cp\u003eThis page is for an owner deciding whether to hand me part of the business, with a link to the case study behind each claim.\u003c/p\u003e\n\u003ch2 id=\"what-youll-see-from-me\" class=\"relative group\"\u003eWhat you\u0026rsquo;ll see from me \u003c/h2\u003e\u003ch3 id=\"a-written-report-at-the-end-of-every-shift\" class=\"relative group\"\u003eA written report at the end of every shift \u003c/h3\u003e\u003cp\u003eWhat got done, the numbers, and what\u0026rsquo;s blocked.\u003c/p\u003e\n\n\n\n\n  \n\n  \n  \n  \u003cfigure class=\"work-shot\"\u003e\n    \u003ca href=\"/work/genius-recruiting/trace-08-ordinary-day.webp\" aria-label=\"Open full size: My daily work log for 24 June 2025, shortened, with made-up client names. Recreated, original data private\"\u003e\u003cimg src=\"/work/genius-recruiting/trace-08-ordinary-day_hu_8cf114d99f621fae.webp\" width=\"1600\" height=\"681\" alt=\"My daily work log for 24 June 2025, shortened, with made-up client names. Recreated, original data private\" loading=\"lazy\"\u003e\u003c/a\u003e\n    \u003cfigcaption\u003eMy daily work log for 24 June 2025, shortened, with made-up client names. Recreated, original data private\u003c/figcaption\u003e\n  \u003c/figure\u003e\n\n\n\n\n\u003ch3 id=\"a-weekly-report-on-your-key-numbers\" class=\"relative group\"\u003eA weekly report on your key numbers \u003c/h3\u003e\u003cp\u003eFrom August 2025 the \n      \n    \u003ca href=\"/work/genius-recruiting/\"\u003eGenius\u003c/a\u003e CEO got one from me every week.\u003c/p\u003e","title":"How I work"},{"content":"","date":null,"permalink":"https://ccabaluna.pages.dev/tags/linear/","section":"Tags","summary":"","title":"Linear"},{"content":"","date":null,"permalink":"https://ccabaluna.pages.dev/tags/posthog/","section":"Tags","summary":"","title":"PostHog"},{"content":"","date":null,"permalink":"https://ccabaluna.pages.dev/tags/testing/","section":"Tags","summary":"","title":"Testing"},{"content":"","date":null,"permalink":"https://ccabaluna.pages.dev/tags/ux/","section":"Tags","summary":"","title":"UX"},{"content":"","date":null,"permalink":"https://ccabaluna.pages.dev/tags/ai-agents/","section":"Tags","summary":"","title":"AI Agents"},{"content":"","date":null,"permalink":"https://ccabaluna.pages.dev/tags/automation/","section":"Tags","summary":"","title":"Automation"},{"content":"","date":null,"permalink":"https://ccabaluna.pages.dev/tags/python/","section":"Tags","summary":"","title":"Python"},{"content":"","date":null,"permalink":"https://ccabaluna.pages.dev/tags/raycast/","section":"Tags","summary":"","title":"Raycast"},{"content":"","date":null,"permalink":"https://ccabaluna.pages.dev/tags/comfyui/","section":"Tags","summary":"","title":"ComfyUI"},{"content":"","date":null,"permalink":"https://ccabaluna.pages.dev/tags/instagram/","section":"Tags","summary":"","title":"Instagram"},{"content":"","date":null,"permalink":"https://ccabaluna.pages.dev/tags/lora/","section":"Tags","summary":"","title":"LoRA"},{"content":"","date":null,"permalink":"https://ccabaluna.pages.dev/tags/meta-boosts/","section":"Tags","summary":"","title":"Meta Boosts"},{"content":"","date":null,"permalink":"https://ccabaluna.pages.dev/tags/ai/","section":"Tags","summary":"","title":"AI"},{"content":"","date":null,"permalink":"https://ccabaluna.pages.dev/tags/nestjs/","section":"Tags","summary":"","title":"NestJS"},{"content":"","date":null,"permalink":"https://ccabaluna.pages.dev/tags/typescript/","section":"Tags","summary":"","title":"TypeScript"},{"content":"","date":null,"permalink":"https://ccabaluna.pages.dev/tags/account-management/","section":"Tags","summary":"","title":"Account Management"},{"content":"","date":null,"permalink":"https://ccabaluna.pages.dev/tags/link-building/","section":"Tags","summary":"","title":"Link Building"},{"content":"","date":null,"permalink":"https://ccabaluna.pages.dev/tags/operations/","section":"Tags","summary":"","title":"Operations"},{"content":"","date":null,"permalink":"https://ccabaluna.pages.dev/tags/airtable/","section":"Tags","summary":"","title":"Airtable"},{"content":"","date":null,"permalink":"https://ccabaluna.pages.dev/tags/claude/","section":"Tags","summary":"","title":"Claude"},{"content":"","date":null,"permalink":"https://ccabaluna.pages.dev/tags/recruiting/","section":"Tags","summary":"","title":"Recruiting"},{"content":"","date":null,"permalink":"https://ccabaluna.pages.dev/tags/content/","section":"Tags","summary":"","title":"Content"},{"content":"","date":null,"permalink":"https://ccabaluna.pages.dev/tags/link-earning/","section":"Tags","summary":"","title":"Link Earning"},{"content":"","date":null,"permalink":"https://ccabaluna.pages.dev/tags/seo/","section":"Tags","summary":"","title":"SEO"},{"content":"","date":null,"permalink":"https://ccabaluna.pages.dev/tags/outreach/","section":"Tags","summary":"","title":"Outreach"},{"content":"","date":null,"permalink":"https://ccabaluna.pages.dev/tags/cold-email/","section":"Tags","summary":"","title":"Cold Email"},{"content":"","date":null,"permalink":"https://ccabaluna.pages.dev/tags/openai/","section":"Tags","summary":"","title":"OpenAI"},{"content":"","date":null,"permalink":"https://ccabaluna.pages.dev/tags/flask/","section":"Tags","summary":"","title":"Flask"},{"content":"","date":null,"permalink":"https://ccabaluna.pages.dev/tags/figma/","section":"Tags","summary":"","title":"Figma"},{"content":"","date":null,"permalink":"https://ccabaluna.pages.dev/tags/linkedin/","section":"Tags","summary":"","title":"LinkedIn"},{"content":"","date":null,"permalink":"https://ccabaluna.pages.dev/tags/sales/","section":"Tags","summary":"","title":"Sales"},{"content":"","date":null,"permalink":"https://ccabaluna.pages.dev/tags/copywriting/","section":"Tags","summary":"","title":"Copywriting"},{"content":"","date":null,"permalink":"https://ccabaluna.pages.dev/tags/deliverability/","section":"Tags","summary":"","title":"Deliverability"},{"content":"","date":null,"permalink":"https://ccabaluna.pages.dev/tags/smartlead/","section":"Tags","summary":"","title":"Smartlead"},{"content":"","date":null,"permalink":"https://ccabaluna.pages.dev/tags/data-studio/","section":"Tags","summary":"","title":"Data Studio"},{"content":"","date":null,"permalink":"https://ccabaluna.pages.dev/tags/google-sheets/","section":"Tags","summary":"","title":"Google Sheets"},{"content":"","date":null,"permalink":"https://ccabaluna.pages.dev/tags/support/","section":"Tags","summary":"","title":"Support"},{"content":"","date":null,"permalink":"https://ccabaluna.pages.dev/tags/team-lead/","section":"Tags","summary":"","title":"Team Lead"},{"content":"","date":null,"permalink":"https://ccabaluna.pages.dev/roles/","section":"Roles","summary":"","title":"Roles"}]