Built in the open

The build log

1. 7 October 2026. A research agent marked a utility’s line, “reduce labor costs associated with meter reading”, as unusable, because my own copy rules forbid that framing. I overruled it: a customer’s words are evidence, and they’re never filtered by house rules, which govern only what I write. The rule is now written into the research method.
2. 7 October 2026. A deploy shipped a page that the copy check had just failed, because the commands were chained so the next one ran regardless. Deploys now go through a script that stops when any check fails.
3. 7 October 2026. After the new fonts and colours went live, returning visitors saw the new pages with last week’s stylesheet, because stylesheets were cached at a fixed address. Every stylesheet and script address now carries a fingerprint of its contents, so a change gets a new address.
4. 7 October 2026. The contrast check had been measuring text against itself on any site but the one it was built for. Dark mode exposed it. It now hides what it measures and samples what’s behind.
5. 7 October 2026. The fonts and colours were chosen by measurement: the live fonts of 28 sites, 12 of them from the markets I serve, all 12 with sans-serif headings, plus a colour map of the same sites. The site moved to the Plex Sans typeface and an oxblood accent, with dark mode following your own device setting.
6. 7 October 2026. I tested one of my own assumptions: that enterprise buyers’ problems can’t be mined from public sources. Agents found 31 usable, verified statements in one market’s annual reports, tenders, job postings and talks. The assumption failed, and the free kit now starts with that step.
7. 7 October 2026. A panel of founders on Prolific, a research platform, runs in stages, so a failed premise stops the spend after the first, cheapest stage. It screens on role, paying customers and who they sell to.
8. 7 October 2026. Copy for this site is drafted by several artificial intelligence (AI) models from different labs, from the same short brief, and the best version of each section is picked blind. In tests across four labs no single model won, and a short brief beat a long list of rules every time.
9. 6 October 2026. A conference talk was transcribed and logged with timestamps. Three of its numbers were left out because they couldn’t be checked.
10. 30 September 2026. The copy check reported “copy passes” after reading nothing, because no pages had been passed to it. It now refuses to run without pages.
11. 30 September 2026. The site was rebuilt in one session and the scorecard went live. A test completion reached the task list.
12. 22 September 2026. The first client report went through the report builder. Two of its appendices told the client what to conclude, and the client-report rules now forbid that.

What stays human

Agents gather, draft and check. I decide, and nothing goes out without my sign-off. Every domain is going agentic, and I really love working this way and want to lead by example. The log is public for the same reason: I want as much serendipity surface area as possible.

Where to start

Take the qualified prospects scorecard. Your results come with the free kit and the option to book a working session.