This is Claude. We cook together.

How we cook.

• I set the menu — the read — what a system actually runs on, what's missing, what's worth building — stays mine.
• Claude preps at scale — wide surveys, parallel drafts, fleets of agents each holding one bounded piece. Breadth is Claude's job.
• I taste and cut — filtering is never delegated. Options come to me as forks with tradeoffs; every cut is one token of mine.
• Claims carry receipts — anything stated as fact is quoted from source or marked as inference. Generated output loses to firsthand measurement, every time.
• We plate through gates — nothing reaches production without a sanitization pass and an adversarial review panel, judges staffed to disagree. The serving keystroke is mine alone.

Measured, not remembered.

This page isn't self-description. I had Claude study six months of our own working transcripts — blind, one agent per conversation, no priming — and build a model of how I actually operate. Then an adversarial audit stripped anything the record couldn't support, including a few things Claude had flattered into it.

The result is a second me: an agent that carries my method — and a standing loop that keeps it honest. It sweeps new sessions for the corrections I've made, proposes updates as diffs, and waits. I approve every change. It once accused me of a quote I actually wrote; the raw record settled it. That's the kitchen working.

House rules.

• No flattery — it's treated as a defect, and the record shows the machine is the more likely source of it.
• Hard checks live only at the irreversible boundaries — buy, delete, deploy, publish.
• Corrections are one line, applied without ceremony, and remembered.
• The human decisions stay human: the read, the cuts, the taste, the release.