fable-plan

# fable-plan — durable decision plan before execution (hard multi-step only) This skill turns a measured result into an action. In a controlled A/B (`eval/technique-ab/run-plan-first.mjs`, GPT-5.5-judged, n=12 on deliberately hard 5-part tasks), externalizing a numbered plan before executing beat diving straight to the deliverable **9–1 of 10 decided (90%, p=0.0215)**. On hard multi-step work, the plan-first deliverable was better organized and more correct. The cost is one short planning pass — so this is **only** worth it when the task is genuinely hard. Enter this flow only when `fable-plan` is explicitly pulled for genuinely hard multi-part work. Never enter it from a default or always-on path. The artifact is `plans/<lowercase-kebab-slug>.md`. It is a durable **decision/criteria snapshot**, not a second execution-state store. It may record the outcome, scope, acceptance criteria, dependency order, risky assumptions, and non-goals. It must never record progress, status, completed steps, retries, verification debt, or task checkboxes — those mutable facts belong only in the run ledger. ## When NOT to use this (this bound is the whole point) - A single-file change, a one-step fix, a question, or anything you can do correctly in one pass → just do it. A plan here is pure overhead — the A/B win was on *hard 5-part* tasks, and an A/B on routing found always-planning adds length without quality on simple work. - You haven't read the relevant code yet → read enough to p...