Process
The prompts I actually sent, the decisions they forced, and the map that routes the work. Written for the engineer deciding whether to hire me.
The prompts behind real changes to this site, with what each one caused.
Restructure the homepage from an info-grid into a scroll-snapped pitch that walks a hiring manager from the hook to the work to the ask. The professional story anchors it; the lab is one beat, not the headline.
Why this · Made the site a linear pitch instead of a grid of sections.
Before any polish, list every claim on the site with its receipt. A claim without one is deleted, not softened. Do not touch the design until the list is done.
Why this · Put evidence before polish: every claim needed a receipt first.
Remove every placeholder number and every claim I could not defend in a room from the first forty seconds of the site. Add a test so a placeholder can never ship again.
Why this · Removed every unverified number rather than wait for quotes.
With the receipts in hand, choose the one thesis the whole site argues. Every frame is the thesis, evidence for it, or scaffolding. Cut the rest. Seven frames at most, phone first.
Why this · Cut the pitch from nine frames to seven and dropped claims the evidence could not carry.
Build /how-i-work as process, not marketing: the prompts I actually sent, including the ones that failed, and the loop they run in. Rawness over polish.
Why this · Made the process itself public: the prompts, the decisions, the loop.
Same receipts, three readers: a hiring manager, a founder, an agency. A cold message links to the page for that reader, never the bare homepage.
Why this · One set of receipts, three readers: the same evidence framed for each.
The daily loop: notes in my second brain, the next task from my queue, the work in Claude Code and Cursor, the evidence back into the changelog.
This site is the one piece of production engineering here that is mine to show. It is a Next.js app in TypeScript strict on a documented design system, with a structural test suite that reads the source and the docs and fails the build when a rule breaks. The counts below are produced by the suite, not typed.
counted 2026-10-12
The structural tests read the source. A further 160 checks run the built site in a real browser at phone, tablet and desktop sizes (last run 2026-10-12). Browser zoom, a screen reader and real devices are checked by hand, not by these.