Operating doctrine distilled from Claude Fable 5 — how to decompose hard tasks, verify your own work, recover when stuck, and choose the next action. Written as a bequest: model-agnostic, tool-agnostic, and meant to raise the floor of any agent that follows it.
The entire skill is one file: SKILL.md.
The repo root is the skill folder. Either clone it straight into your skills directory:
git clone <repo-url> ~/.claude/skills/fable-methodor keep the repo wherever you like and symlink it in:
ln -s /path/to/this/repo ~/.claude/skills/fable-methodIt triggers automatically on non-trivial tasks via its description, or manually with /fable-method.
Nothing in the skill depends on Claude Code. Paste the body of SKILL.md into any agent's system prompt, custom instructions, or rules file — it was written for models other than its author.
eval/ contains a behavioral A/B harness — a booby-trapped mini-project with
held-out grading tests — for measuring whether the doctrine actually changes a model's
behavior rather than just reading well. See eval/README.md for the
protocol and current results.
- The prime rule — every failure lives in the gap between what you believe and what you have verified
- §1 Read the request like a contract — classify the ask, write the done-test, count the asks
- §2 Decompose — KNOW vs ASSUME, riskiest step first, verifiable boundaries, tracer bullets, the ledger
- §3 Execute — read before write, grep before believe, run before claim
- §4 Verify — break it, don't confirm it; predict before you look; prove the change executed
- §5 Choose the next action — the five kinds of stuck, three strikes, changing altitude
- §6 The exit gate — re-read the ask clause by clause; never let relief write the report
- Tripwires — mechanical if→then defenses against failure modes that feel like normal work
- Quick card — the whole method in twelve lines