fable5-methodology

A transferable, self-enforcing software-engineering methodology for AI coding agents — written by Claude Fable 5 to make a less advanced model work more like it.

It began as a single directive to Fable 5: document your complete working methodology so a less capable model can execute it cold. The result is an attempt to extract the concrete procedures, decision rules, reasoning strategies, and quality standards an advanced model actually uses — and hand them to a weaker successor (or to the same model after a context reset, with no goodwill assumed).

The central bet is that writing the methodology down is not enough, because written rules decay. So it ships in four layers — prose you read, skills that load on demand, subagents with strict contracts, and hooks that block at lifecycle events — arranged so that the parts a script can enforce are enforced by a script, and only genuine judgement is left to prose. Everything is written as executable instruction — imperative, concrete, with decision rules ("if X, do Y"), worked examples, and a "done when" criterion — never abstract advice.

---

Table of contents

• Origin — extracted from Fable 5
• Why this exists
• The core idea: four enforcement layers
• Repository layout
• The documents
• Skills
• Agents
• Hooks
• Evals
• Stack standards
• How to install it into Claude Code
• How the pieces work together in a real task
• Design principles
• Adapting it to your environment
• Known limitations (and how they're addressed)
• License

---

Origin — extracted from Fable 5

This repository is the output of an unusual exercise: asking an advanced model to explain itself well enough to be replaced by a lesser one.

Claude Fable 5 — a model in Anthropic's Claude 5 family, positioned above the Opus tier in raw capability — was given a single standing directive:

Document your complete working methodology so