How it was built
Two projects in eleven days, built through Claude Code with one loop, four steps, and a person who writes the issues and says merge. The counts from the session transcripts, what the review caught, and what the person did.
Sheet 05 — Field notes
Notes from things I am building. Mostly the unglamorous half of an AI system, the part that decides what gets indexed before anyone types a query.
07 notes on file
Two projects in eleven days, built through Claude Code with one loop, four steps, and a person who writes the issues and says merge. The counts from the session transcripts, what the review caught, and what the person did.
A clip search over 300 public-domain films where every result opens the player at the frame it matched. Pick clips into a set and one page says what you would be paying for. The page's numbers, the ones that are good and the two that are not, and a sampling thesis the eval turned down.
The same clip search answers a person through a page and an agent through three MCP tools. One query run both ways, and a list of what the person gets that the agent cannot reach yet. The tools have a few things the page lacks too.
Ten incidents scripted into a real Honeycomb environment, with the true cause in a file the agent never sees. A grader that scores the answer, and takes more off a wrong answer said with confidence than a wrong one said with doubt. Thirty of thirty right on one model. Six of thirty on another that passes the vendor's own process check.
A trace of one request, stage by stage, is what explained a bad number. A search timeout had moved from eight seconds to fifteen on the strength of one measurement, and within ninety seconds of the deploy two searches were cancelled at fifteen point one.
A clip search app's 264,626 frame vectors, kept in Postgres all project, written a second time to object storage and asked the same hundred questions through the same route. The answers matched and the price of a query moved. Then a commercial video model went in as a third column, over the slice its free tier would hold.
A kelp farmer spent fifty minutes pointing at a map on a screen-share. A regex over the words that point turned the recording into a permit briefing where every claim plays its own evidence.
Every note here is downstream of something on the work sheet.