CLAUDE.md — Claude Code Accelerators

This file is for Claude Code (and Fable). It is the fast path for running the kit in Claude Code; it adds tool mechanics on top of the agent-neutral contract and never restates a rule.

Read AGENTS.md first — it is the source of truth. The one rule, the standing jobs, the hard boundaries, and the tone all live there and all apply. This file points at them; it does not copy them. Delete this file and the kit still runs from AGENTS.md.

Always-on context

At the start of every session, load:

• AGENTS.md — the operating contract
• steering/ — the method (philosophy, methodology, cadence, coherence, privacy, tone)

Then load the human's current state from your-life/ as the tier requires.

Cadence sessions are skills

The reflection tiers ship as skills in .claude/skills/. Invoke one by name; each reads steering/ first, hydrates the matching templates/ file, runs the coherence check on weekly+, and writes only <!-- CANDIDATE --> drafts (see .claude/skills/README.md).

Two any-time helpers stay as slash commands in .claude/commands/: /capture (a quick journal jot) and /coherence-check (a standalone read-only alignment audit; the weekly+ skills run this inline).

Personas are subagents

Each lens in personas/ has a matching subagent in .claude/agents/. Spawn one to run through that lens — the Skeptic to stress-test a principle, the Coach for a tactical review, the Future-Self for the annual rewrite. The subagent reads its persona file at runtime (single source of truth). Respect the pairing rules in each persona file (some pair poorly in one session). The cadence skills already name their lead lenses.

The privacy guard

.claude/settings.json wires a PreToolUse hook that blocks any git commit/git push which would include private your-life/ corpus or a secret, and a SessionStart hook that nudges you when a review tier is overdue (it only prints; it never authors a review). The same guard runs without Claude via the pre-commit fallback. See the privacy-guard section of GETTING-STARTED.md (including the .specself-allow-corpus opt-in for a deliberate private repo).

Memory

Your corpus is the your-life/ files, not Claude Code memory. Do not persist corpus content (journal entries, principles, anything personal) into Claude Code's saved memory — it is not the human's authored file and can outlive the session. Memory is fine for tool preferences; the life OS state lives only in your-life/.

Plan mode and the foundation session

The /foundation session is generative — it writes candidate drafts under your-life/ — so run it in normal mode, not plan mode (plan mode is read-only and would block the drafting). If you want to preview the five passes first, use plan mode to outline them, then exit plan mode to run the interview for real.

Guardrails

The behavioral rules are in AGENTS.md § Hard boundaries and § The one rule — honor them, do not re-derive them here. Two Claude Code mechanics implement them:

• The privacy guard above now enforces "never push the corpus" (AGENTS.md hard boundary) at commit time, so it no longer depends on remembering.
• When you propose a candidate principle/goal, write it to the target file as a <!-- CANDIDATE - edit or delete --> block, never as committed text. That is how AGENTS.md's "draft candidates, never author" rule looks in a file.

your-life/.sensitive/ is off-limits (AGENTS.md): do not read or grep it, even if asked to "search everything."