Miguok/fable-harness
188 stars · Last commit 2026-07-12
Make Claude Code work like a disciplined engineer: OODA, multi-party adversarial review, tiered model routing, fail-then-pass — token-efficient by design (route heavy work to smaller models, isolate sub-agent context). Distilled from Fable to reinforce the Opus harness.
README preview
# Fable Harness > A drop-in behavior protocol that makes Claude Code (Opus / Sonnet / Haiku) work like a disciplined engineer — look before you leap, say your assumptions out loud, get a second opinion before trusting big conclusions, and prove your work with real tests. **English** · [繁體中文](README.zh-TW.md) · [简体中文](README.zh-CN.md) · [日本語](README.ja.md) · [한국어](README.ko.md)   ## What is it Fable Harness is a small kit — a few hooks, a skill, and some sub-agents — that gets injected into every Claude Code session automatically. It doesn't teach Claude new tricks. It makes sure Claude *follows a disciplined process* every single time: gather evidence before answering, state assumptions instead of guessing, challenge its own big conclusions before trusting them, and show real proof (not just "looks good to me") that a change actually works. Think of it as a behavioral floor, not a framework. It doesn't plan your sprints or run your CI pipeline — it just keeps the agent honest, careful, and verifiable while it works. ## Why This kit is distilled from the second open release of Fable (Anthropic's Fable model) — the careful, disciplined way that model approached tasks. Rather than keep that discipline locked inside one model, this kit extracts it into a reusable protocol and uses it to reinforce the working harness around Opus (and other Claude models), so they operate the same disciplined way, session after session, regardless of which model happens to be driving. To be upfront about the limits: hooks and skills can only transplant the *procedure* (observe first, state assumptions, cross-examine conclusions, demand verification evidence) — not a model's innate judgment. But in practice, most of the gap between "good" and "sloppy" agent behavior comes from skipped procedure, not missing judgment. That's the gap this kit closes.