Before Claude Fable 5 died, it wrote its own instruction manual

What it does

The repo captures Claude Fable 5’s approach as three executable skills—fable-method, fable-loop, and fable-judge—that replace vague prompting with hard decision points. Each skill forces the model to classify the ask, name a verification before acting, gather evidence in parallel, commit to one recommendation, and verify by observation rather than trusting its own report. It ships with domain adapters for marketing, research, legal, and other non-coding tasks, plus an adversarial judge that diffs and executes to catch false completion claims.

The interesting bit

The author treats the instruction file as executable pseudocode: weak models follow rules at decision points, not prose in lists, so every rule exists because a test failed without it. The eval log is unusually honest, reporting null results—ordinary tasks on capable models show no lift—alongside the wins, because a log that only contains wins would not be worth trusting.

Key highlights

• Three skills cover the full lifecycle: classify and plan (fable-method), orchestrate execution (fable-loop), and adversarially verify (fable-judge).
• The judge re-runs claimed checks, diffs actual changes, and hunts weakened tests or planted frauds rather than reading completion reports.
• Eight eval rounds across 159 agent runs tested the method against traps like spec-vs-test conflicts and false “work complete” claims; raw transcripts are committed.
• Domain adapters define binding minimum evidence sets for marketing, research, finance, legal, and design, while deliberately omitting medical work.
• The method is packaged as a Claude Code plugin with versioned, namespaced commands, or as standalone skills for other agents.

Caveats

• The evidence is explicitly “smoke-test grade” (1–4 runs per cell, LLM judges), so the numbers are directional rather than rigorous benchmarks.
• The value concentrates at “traps”—authority conflicts, false completion claims, unattended runs—with no lift on simple attended tasks.

Verdict

Worth studying if you are building agent instruction sets or trying to make smaller models reliable on complex, multi-step work. Skip it if you need a general-purpose prompt that improves every query; this is a surgical protocol for situations where agents usually lie or drift.