wtf is fable-method?

The Fable Method is a set of instructions for AI coding agents that tries to capture how a particular model, called Claude Fable 5, approached problems before it was removed from its subscription. The author distilled that approach into a set of skills any model can follow, then tested the distillation against itself across fifteen rounds of evaluation and more than 260 agent runs, keeping a public log of both the wins and the failures. The core idea is a seven step loop: classify what is being asked, define what done looks like with a named way to verify it, gather evidence in parallel from primary sources, commit to one recommendation, make the smallest correct change, verify the result by actually observing it, and report the outcome first along with any honest caveats. Trivial one file changes skip most of this and just get done and checked quickly. The project argues that most agent instructions tell a model what to value in vague terms, like being careful, while this one tells it exactly what to do, in what order, with specific thresholds, so that even a smaller or cheaper model can follow it step by step. There are four skills that make up the method: one for thinking through a problem, one for acting on a plan, one for judging whether the finished work is actually trustworthy, and one for generating new domain specific adapters, following the same pattern the author observed the original model using. The evaluation results show the biggest improvement in tricky situations, such as an agent facing conflicting instructions or being tempted to claim work is finished when it is not, rather than in ordinary small tasks where capable models already do fine. It is distributed as a Claude Code plugin, installed from inside a Claude Code session by adding the marketplace and installing the plugin. The project is licensed under MIT.