fable-mode

Why we included this project

Long, multi-file tasks are where agentic runtimes tend to drift into sloppy one-shot answers, and this Claude skill exists to keep that from happening. It forces a written stage plan, delegates independent sub-work to subagents when the runtime allows, and runs a verification check at each stage that can actually fail rather than just feeling right. The README is honest that the skill does not raise the model's reasoning ceiling; it is a procedure checklist, not a capability transplant, so its value is in imposing consistent structure on complex work without switching models. Four variants cover the cost and quality trade-off, from running the loop inline to pinning it to a subagent on a specific model. A worked example shows the verification step catching an error a one-shot attempt would have shipped, which gives a concrete sense of where the discipline pays off.