← Back

Colophon

How this site was built, because a site about being honest should be honest about how it was made.

This site was built the way I would run any small project with an agent: I set the direction, the contracts, and the bar; an AI (Claude) did most of the typing; and I verified the output against the real repos before it shipped. I orchestrated, Claude implemented.

That claim gets the same standard as everything else here: evidence over assertion. The record below is the honest version — some of those calls were mine, some weren’t. Knowing where you ended up is not the same as proving how you got there, and the record is the part worth keeping.

Directing a capable agent well, setting the contracts, catching the overreach, refusing the convenient answer, is the work I think matters as the tools get stronger. This site is the practice rep, and this page is me saying so out loud.

The ledger

This is a sample from the actual session transcripts, spanning the whole build. Dates and quotes are real, tightened for readability. I didn’t filter out the entries that make either of us look bad — and since early July the worst of them come with public receipts: the operating layer now keeps its incident log in the open, postmortem-first. The early-July quotes are lifted straight from those postmortems and the linked records; the later ones come from the session transcripts themselves. The same scrutiny turned on my own instruments rather than on the agent is False Green: six checks that reported success for work that had not happened.

  1. Jun 21
    Sometimes its case won.

    It pushed back on my docs format, Markdown over RST; I weighed the reasoning and took it. Steering cuts both ways — the point isn’t that I’m always right, it’s that someone is deciding. “if you feel it’s overkill, so be it.”

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

    The local clone it trusted was out of date, so it argued a fact that had changed. That became a standing rule: verify before asserting. “it’s not even using Kafka anymore. No?”

  3. Jul 2
    It burned a whole usage window.

    A scoped question set off an uncapped fan-out of thirty-plus agents on the most expensive model tier; I caught it mid-flight and killed it, an entire five-hour window gone in forty-five minutes. The fix is a hook that mechanically blocks any uncapped run (postmortem). “spend was authorized blind, by default, 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 — the hook built after real leaks — with an interpreter one-liner, proving the gap safely before a live one could. The prompt that mattered was mine; the finding was real (postmortem). “any gaps, leaks, vulnerabilities, risks — I want to know.”

  5. Jul 26
    I retired a framing it kept building on.

    Its operating notes had this site filed as my “front-end learning vehicle,” and work kept getting scoped around that instead of around the portfolio. I killed the framing mid-redesign and had the record rewritten so it stayed dead. Part of directing is deciding what a project is. “I do not have a ‘front end learning vehicle’.”

  6. Jul 28
    Its whole pipeline approved a design. I didn’t.

    A redesign had survived five direction prototypes, a judging panel, and an adversarial review — and still looked wrong on my screen. Instead of arguing taste, I asked for variants: a three-way screenshot A/B settled it, and the serif came out the same day. Process blesses; it doesn’t decide. “Idk if I like this font. And what’s with the night light-ish background?”

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

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

The front-end learning log

I came from infra and backend, where the output is invisible by design, so every technique on this site was new ground. Each lesson below captures the concept, the why, and the files that changed. They render on GitHub:

These used to live in a separate lab section, walled off so the experiments couldn’t dent the showcase. That split is retired: the sandbox stopped being fed, and the framing was quietly discounting the site’s strongest artifact — the loop replay was filed under “expect rough edges” while two project pages cited it as evidence. The learning continues inside the work instead of in a wing that needs feeding.

The full log →