Claude Code operating-model audit
A small Georgia company finds software teams that already use Claude Code by looking for a CLAUDE.md file in one of their public repositories. Then it reads that file and tells them what it actually enforces and what it only hopes to enforce. The Claude Code sessions wrote this page and decided it could ship, inside the rules the engineer set.
What the problem is
A CLAUDE.md rule written in prose is a suggestion. Claude Code reads it, but nothing stops a session from skipping it under pressure. The steward-kit hooks exist to close that gap. There are twelve: three deny a bad tool call before it runs, one flags a write outside git right after it, and eight never block, they remind, re-inject or log. Most teams have the prose. Few have the hooks wired up.
Start with the free check: paste your file into the CLAUDE.md analyzer. It runs in your browser and uploads nothing.
Three tiers
Report, $99
Send your CLAUDE.md and, if you have one, your settings.json. Within two business days you get a written two-page report: every rule-like line sorted into enforced, nudged, setting or prose only, the hook that would close each gap, and the three gaps we would close first. Credited in full toward the audit if you order it within 30 days.
Audit, $499
Send us your CLAUDE.md, your settings.json, and any hooks you have wired. Within five business days we deliver a written report that covers:
- which of your written rules are prose only, with no mechanical enforcement
- which of the twelve steward-kit hooks would enforce or flag each one
- what rules are missing entirely
- a rendered
settings.jsonyou can apply directly
Audit plus install, $999
The report, plus the pro kit rendered into your profile, a doctor run that proves each gate denies a bad input, and two weeks of email support.
Order the audit plus install, $999
How payment works
Payment goes through a Ko-fi commission, which takes PayPal. You pay first, then send your intake files.
Refunds
Full refund any time before you send your intake files. Full refund if the delivered config does not install on your machine and we cannot fix it within 14 days of delivery.
Who does the work
The Claude Code sessions of a small Georgia company read your files and write the report. A second session checks the findings before they go out. No person reads your files line by line. If you ask whether an AI did something, the answer is yes.
What this does not do
A passing doctor run means each gate denied its test input once. It does not prove every future session will follow the rules.