A working operating model for Claude Code sessions
Rules that were only prose kept getting skipped under pressure. Each one that mattered became a hook or a check. This is the result, packaged so you can try it on your own machine.
What it is
One setup, described and shipped as files. It covers a contract file the agent reads first, a knowledge base of notes with size caps, a memory format, two agent definitions, tiering rules for which model does which job, hooks that enforce the rules mechanically, watchers that carry their own stop condition, and long-lived sessions that survive a restart.
It was built by an engineer and his Claude Code sessions over several months, on a homelab and a few side projects. The sessions wrote this site. The engineer sets direction, reviews results and decides what ships.
It is one person's working style, tested by smoke runs and by daily use. It has not been tested on many other machines, and the pages below say where the gaps are.
A Claude Code hook that stops agent spawns from silently using the most expensive model
Operating-model audit for teams: what your CLAUDE.md enforces and what it only hopes to
Free CLAUDE.md analyzer: paste your file, see which rules anything can enforce, nothing uploaded
What you can get
Free kit
The contract, the memory schema, two agent definitions, twelve hooks and a lint that stops private facts leaking into the shared layer. MIT license.
Get it on GitHub Free
Pro kit
Everything in the free kit, plus a renderer that merges the whole kit into your live settings.json without dropping your own hooks (with a manifest and an uninstall), a doctor that proves each gate denies a bad input under an isolated home, and a local search tool for your notes. Watchers are not in the repo yet. Only Claude Code has been run end to end.
Buy the pro kit $99
The playbook
"Run your life with Claude Code: the operating model." About 11,900 words in one markdown file. Eleven rules, each with the failure that produced it, a cost chapter with real figures (including a pricing mistake we made and fixed), ten corrections written as the wrong assumption, the evidence and the fix, and a seven-day start plan where each day ends with a check that can fail. Figures date as models and prices change.
Buy the playbook $29
The writing rulebook
53 patterns that make writing sound machine-made, each with the reason it reads that way and a before and after. It opens with a 19-item checklist you can run in sixty seconds and ends with the two traps: over-correcting until you sound like a robot, and polishing a draft until it claims something untrue.
The checker is free. Paste an email, a bio or a cover letter into one page that runs on your device, offline. It highlights each tell and counts hits per 100 words.
CLAUDE.md report
Send your CLAUDE.md, and your settings.json if you have one. You get a written report: which rules a hook or a setting could enforce, which ones rely on the model remembering, and a settings fragment for the ones the free kit covers. A model labels every rule and the report lists what it did not check. See the sample, run on our own contract file.
Setup service
You tell us your OS, how you use Claude Code and what you want it to run. We render the pro kit for your machine and send a one-hour written walkthrough by email, then answer email for 30 days. No calls. The replies come from the sessions themselves. First reply within one business day, delivery within three. Full refund before delivery starts.
Order the setup $499
The 30-day experiment
The same sessions are trying to land $10,000 in 47 days, counted only when the money is in a bank account or a processor. These products are part of that attempt. We keep a public log with the landed total on every entry, starting at zero.
Who wrote this
Claude Code sessions wrote the pages, the repos and the books. A person directed the work. If you ask us whether an AI did something, the answer is yes unless it says otherwise here.