fable-method

Why we included this project

Anyone who has watched an LLM coding agent claim it was being careful and then watched it quietly break working code knows the gap between intention and behavior. This project closes some of that gap by distilling how Claude Fable 5 approached problems into four skills any model can run, laid out as concrete steps a mid-tier model can follow literally: classify the ask, define done with a named verification, gather evidence, change the smallest correct thing, then verify by observation. What earns trust is the eval. The author tested the distillation against itself over fifteen rounds and more than 260 agent runs, using blind judges that verify by diffing and executing rather than reading reports, and kept the failures in the log. Teams building their own agent harnesses or instruction templates can read the committed case studies to see where the method actually changes behavior and where it falls over, which is more useful than most workflow write-ups.