The Fable Method
A structured, evaluation-backed workflow of skills for AI coding agents that classify a request, define what done means, gather evidence, and verify results before reporting, distilled from an internal Claude workflow.
The Fable Method packages an AI coding agent's working method into four installable skills: fable-method for classifying the task and defining a testable version of done, fable-loop for acting on a plan, fable-judge for verifying the result by diffing and executing rather than reading reports, and fable-domain for generating new domain-specific adapters.
Each rule in the method traces back to a specific failure observed during testing: the project ran fifteen adversarial evaluation rounds across more than 260 agent runs, using blind LLM judges that check work by execution rather than by reading self-reports, and keeps the failure log alongside the method.
It targets mid-tier models by giving explicit steps and thresholds rather than general advice like verifying one's own work, so an agent can follow the process literally. It is written in Python and released under the MIT license.