The 'Fable Method' Distills a Removed Model's Own Workflow Into Skills — Tested Over 260+ Agent Runs

GitHub·medium signal

Sahir619/fable-method (MIT, v1.4.0 on 2026-07-15, 1,214 stars since 2026-07-06) packages a seven-step agentic loop — classify the task, define success criteria, gather evidence, decide on one recommendation, execute surgically, verify, report honestly — into four portable skills: fable-method (decision framework), fable-loop (parallel evidence gathering), fable-judge (adversarial verification that hunts false completion claims), and fable-domain (generates domain adapters for marketing, research, devops). The eval ran 15 rounds and 260+ agent runs testing conflict detection in contradictory specs, fraud identification in falsified work reports, cross-model comparison, and adapter quality; results showed the method most strengthens *weaker* models on high-stakes tasks. The repo claims Claude Fable 5 distilled its own approach before removal from subscription, then broke one of its own rules under adversarial test.