Mise en Place: AI Agent Setup Audit

Everything in its place before service.

A kitchen only scales when the station is set up so the easy thing is the right thing. This skill treats your repository plus your agent setup as a kitchen, inspects every station, and tells you the one upgrade that lets you run more coding agents with less supervision.

The method is distilled from how engineers shipping thousands of agent-written PRs a month actually work: the engineer's job is now the environment. Trust in agents is earned through verification and constraints, not by hoping the model is good.

What you get

• A read-only audit script (kitchen_audit.py, stdlib Python, no dependencies) that scans five stations:
  1. Hands and eyes: can an agent run, drive, and observe your product without you?
  2. Knives: do deterministic helpers exist, or does every agent rebuild them?
  3. Guard rails: linters, strict types, hooks, god files.
  4. Recipe cards: your AGENTS.md / CLAUDE.md, with plugin-injected blocks separated from your own lines.
  5. Tasting notes: reverts, hotfixes, untracked work, and user corrections mined from your Claude Code transcripts.
• The trust ladder: four rungs (home cook, sous chefs, executive chef, restaurant chain), each defined by what the human no longer has to do, with honest placement rules.
• A one-screen Kitchen Report naming the current rung, the single station pinning it there, the one upgrade that moves it up, and two or three quick sharpenings.
• Seven upgrade recipes: hands and eyes, deterministic extraction, failure-to-constraint, gardening routine, outer loop, sampling review, autopilot gate. Each says what to build, why it works, and how to know it is done.
• Thirteen principles with the reasoning behind them, so the agent can handle situations the recipes never anticipated.

Example prompts

• "I'm the only dev on this repo and I'm still babysitting every Claude session. Audit my setup and tell me what to fix first."
• "My agents keep writing the same seed-database helper. Fix that properly."
• "Can I let agents merge their own PRs on this project? What has to be true first?"
• "Three different sessions appended to the same god file. Make that impossible."

Sample output

Kitchen Report: auracheck
Rung: 1 Home cook
Pinned by: Hands and eyes. There is no way for an agent to run an audit
and look at the result; the test bench is production.

The one upgrade: an `npm run verify` an agent can call that starts the
app, posts a fixture, drains the stream, and prints a JSON summary with a
non-zero exit on failure. Moves the kitchen to rung 2. Cost: one agent
session.

Works with

Claude Code first; the skill is plain Markdown plus one Python script, so it runs in Codex, pi, OpenClaw, or any harness that reads SKILL.md. Python 3.8+ for the audit script.

What it is not

It is not a linter or a test framework. It is the judgment layer that tells you which constraint, script, or verification loop to build next, and then walks you through building it.