Foundry · The Operating Guide
How to run a business on an operating system.
For founders who just got an OS installed — or instantiated the open-source template themselves. No AI background assumed. Ten minutes to read; thirty minutes a week to operate.
What you actually have
Most companies use AI as a chatbot: blank window, blank context, generic output. You have something different — a repository that teaches the AI who you are. Five contract files do the teaching:
- CLAUDE.md — the doctrine. The company handbook, read automatically by every AI session.
- AGENTS.md — the same rules on one page, readable by any AI tool.
- SKILL.md — what to load before each kind of work, and when the AI should refuse.
- design.md — every color, font, and spacing value your site is allowed to use.
- taste.md — the judgment tokens can't capture: references, refusals, the polish pass.
Edit these files and every future AI session changes behavior. That's the whole trick — the configuration is the company knowledge, and it compounds.
Day 1 — one evening
- Install Claude Code (or use Cursor or Codex — the contract files cover them too).
- Open a terminal in your repo and just talk. Try: “What are the rules of this repo?” The agent recites your doctrine back — that's how you know the harness is live.
- Read your own CLAUDE.md. It's the best summary of how everything fits.
The weekly rhythm — the whole operation
| When | Command | What happens | Time |
|---|---|---|---|
| Monday | /weekly-content | Plans the week: one post, one distribution action, one ops task. Three items max — sized for founders with day jobs. | 10 min |
| Midweek | /blog-post | Drafts, polishes, gates, and build-verifies a piece of content. | 1–2 h |
| Friday | /weekly-review | Asks you three questions, closes the week, updates the business memory. | 15 min |
The Friday step is the one people skip and shouldn't. The plan is disposable; the review is the asset — it writes to docs/intelligence/, the memory that makes week 30 smarter than week 1. The operating discipline in one line: check memory first, update memory after.
The gates — why your output won't look AI-generated
Nothing publishes without passing:
- @claims-guard — blocks regulated claim language, uncited assertions, banned phrases, and AI-tone. A FAIL is final until a human rewrites. This is the agent that keeps you out of regulatory trouble and out of the “obviously a chatbot wrote this” zone.
- /ship — typecheck, build, claims, and SEO before any deploy. A human always presses the deploy button.
- The polish pass — seven manual checks in taste.md before any visual ships.
The standing rule: agents draft, gate, and commit; humans deploy, post, and send.
When the harness improves upstream
Your repo descends from agentic-business-os. When the upstream harness improves, you receive a pull request with a plain-language changelog. You read the diff — it's markdown, readable in the GitHub interface — and merge or decline per file. Nothing ever auto-merges, and your brand files are never touched. The full contract is one page: HARNESS.md. The principle it encodes: brand is yours, machinery is shared.
Where to get help
- The template and skill packs are MIT-licensed — community help via GitHub issues.
- Installed by the Foundry? You have a direct channel — use it.
- Considering an install? Apply here.