← Back

Colophon

How I built this site. The method, and the session record that is its evidence.

Plate 01 I set the aim. The agent types. I check the output. I set direction It types the agent I check then it ships

I set the direction. Claude typed. The ledger is the proof.

I built this site the way I would run any small project with an agent. I set the direction, the contracts, and the bar. Claude did most of the typing. I checked the output against the real repos before it shipped. The evals and postmortems are the proof.

That claim gets the same standard as everything else here: evidence over assertion. Some of those decisions were mine. Some were not. The ledger below is the record of which.

The ledger

This is a sample from the actual session transcripts. It covers the whole build. The dates are real. I tightened the quotes to the site’s writing standard. I did not remove the entries that record my own errors or the agent’s errors.

Since early July the worst entries carry public records. The operating layer keeps its incident log in the open, postmortem-first. I lifted the early-July quotes straight from those postmortems and the linked records. The later quotes come from the session transcripts themselves. False Green turns the same scrutiny on my own instruments rather than on the agent. Six checks reported success for work that never ran.

  1. Jun 21
    Sometimes its case won.

    It pushed back on my docs format, Markdown over RST. I weighed the reasoning and I took it. Steering works in both directions. The point is not that I am always right. The point is that someone decides. “If you decide the format is more than we need, keep that decision.”

  2. Jun 27
    It insisted the system still used Kafka.

    The local clone that it trusted was out of date, so it argued a fact that changed. That became a standing rule: verify before asserting. “The system does not use Kafka now. Confirm this.”

  3. Jul 2
    It burned a whole usage window.

    A scoped question started an uncapped fan-out of thirty-plus agents on the most expensive model tier. I found it mid-flight and I killed it. An entire five-hour window went in forty-five minutes. I built a hook that mechanically blocked any uncapped run (postmortem). I retired that hook on 2026-08-30, after two weeks of usage data showed that the runs never reached the weekly limit. “Spend was authorized with no review, by default, and with no gate.”

  4. Jul 3
    I asked for holes; it found one in its own guard.

    On an open audit request it bypassed its own credential guard with an interpreter one-liner. I built that hook after real leaks. The bypass proved the gap safely, before a live one could. The prompt that mattered was mine. The finding was real (postmortem). “Tell me every gap, leak, vulnerability, and risk.”

  5. Jul 26
    I retired a framing that it built on again and again.

    Its operating notes filed this site as my “front-end learning vehicle”. That framing scoped the work again and again, instead of the portfolio. I retired the framing mid-redesign. I had the record rewritten, so the framing stayed dead. The direction includes the decision about what a project is. “I do not have a ‘front end learning vehicle’.”

  6. Jul 28
    Its whole pipeline approved a design. I did not.

    A redesign survived five direction prototypes, a judging panel, and an adversarial review. It still looked wrong on my screen. I asked for variants instead of an argument about taste. A three-way screenshot A/B settled it, and the serif came out the same day. A process approves a design. It does not decide. “I do not know if I accept this font. Explain the night-light background.”

  7. Aug 2
    I made it read this site cold.

    I had it read the site as a hiring manager who knows nothing about me. It got no context and no credit for effort. The verdict was blunt and correct. The homepage explained how I work before it said who I am. I re-cut the positioning the same day. “Assume you know nothing about me.”

Site decisions

The site keeps its decisions as dated records in the public repo. These six shape what a reader sees.