03 / The system
The brief said "tell us how you used AI." Here's the honest answer: not as an autocomplete, but as an operating system for the work — parallel research teams, a local video-intelligence lab, and a build partner, all directed from one conversation. Every stage below ran inside the four-hour window.
Before touching the demo, three research agents fanned out across the public web: a site & messaging teardown of every kodexglobal.com page (voice, template anatomy, feature inventory, gaps), a buyer and market pass (the PMM job posting decoded, funding history, the Coinbase case study, the fake-EDR fraud wave), and a competitive landscape (four real competitors, the manual status quo, the e-Evidence regulatory clock). Everything source-cited, inferences labeled.
research corpus: 3 documents, every claim linked to a URLI watched the full demo end to end first — the AI pass came after, not instead of, my own viewing. Then the 34-minute raw recording (720MB, unedited, repeat takes and all) was processed entirely on-device: ffmpeg extracted audio and 135 frames; Whisper large-v3-turbo transcribed locally; three vision agents walked the frames and produced a timestamped, screen-by-screen map of the product. Result: a 21-feature inventory cross-referenced against the site teardown — including sales-motion context that ended up deciding the feature pick.
feature inventory + transcript analysis + 3 frame mapsShortlist of three, argued on the evidence: Transparency Reporting won because it's a provable site gap, undersold by the structure of the sales motion itself, backed by a page-ready automation claim, and riding a live regulatory clock. The runner-up (the API layer) and the case against the winner are both preserved in the working doc.
§1 of the thinking doc, counterargument includedBuyer archetype (grounded in years working with compliance teams), the cost-center-to-trust-asset flip, a three-pillar message house, hero line chosen from four candidates, and an assumptions register that says exactly what I'd verify on day one. The page was built on top of this — not the other way around.
the thinking doc, evidence classes on every claimDesign tokens pulled from your production CSS. Your newest page template decoded from live screenshots. Your licensed typeface matched by rendering six candidate fonts against your real headlines and comparing letterforms. Three review rounds of my notes — typography, contrast bugs, label precision — each verified with headless-browser screenshots in both themes before committing. Plus one addition your site doesn't have: a proper light/dark mode.
/page — one file, no framework, both themes, reduced-motion safeA page tells you what I decided. The working doc tells you how. This site exists so one link carries both — plus the part that usually stays invisible.
fletchbuilds.com| Phase | My time | What ran in parallel |
|---|---|---|
| Setup — project scaffold, working rules | ~15 min | — |
| Recon — direction + review of findings | ~15 min | 3 research agents, ~35 min wall |
| Video — drop file, review inventory | ~10 min | transcription + 3 vision agents, ~25 min wall |
| Pick + messaging — the judgment work | ~50 min | — |
| Page — direction, notes, three review rounds | ~45 min | builds + screenshot QA between rounds |
| Hub + deploy | ~20 min | — |
| Total, as of deploy | ~2 h 35 min | within the 4-hour cap |
Times are honest estimates from the session log, rounded. The ledger includes ~15 minutes of project setup that isn't a pipeline stage; stage times cover direction plus review of what came back.
Anyone can ask AI for a landing page. The job is knowing what to build, why this one, and what not to ship — and having the system to do all of it before the clock runs out.