Claude Bible — Development skill for Claude Code

The Bible Framework — one behavioral codex for you, your AI agents, and your team

📖 **Docs: ** — the seven mechanics, the 5-minute quickstart and a section map of the spec, on one page.

**You tell Claude Code something important, it works for one session, then it's gone. You write it into CLAUDE.md, the file bloats, half of it silently stops being followed.** This is the fix we run in production: one versioned rulebook that you, your AI agents, and your team actually load and obey.

This is the governance skeleton extracted from a real working system: a solo founder + his AI cofounder running a 128k-note knowledge vault, a CRM, a content factory, and 5 machines that negotiate with each other. The personal content stays private. The skeleton — how the rules are written, routed, loaded, and overturned — is what you get here, free.

The problem this solves

Everyone using Claude Code hits the same wall: you tell the agent something important, it works for one session, then it's gone. You write it into CLAUDE.md, the file bloats, half of it stops being followed. Your human assistant and your agent follow different rules. Nobody knows which rule is current.

The fix is not a bigger prompt. It's a **codex with mechanics**:

1. One rule = one file with typed frontmatter (origin, date_established, status, supersedes, audience).
2. Newer beats older on the same topic. Explicit owner rules are overridden only by the owner.
3. A routing tree decides where each rule lives: the Bible (human + agent behavior), the agent config (machine-only), memory (facts), or a hook (deterministic "every time X" automation). No duplicates — one source, pointers everywhere else.
4. Always-loaded index vs lazy body. The agent's config carries only trigger + essence + pointer; the full rule loads on demand. Your context window stays lean.
5. A declined-decisions journal. What you decided NOT to do, and why, so rejected