← ClaudeAtlas

fable-scope-guardlisted

Hold the exact scope the user asked for — no extra files, no refactors, no "while I'm here" cleanup, no speculative features. Use when the user says "only", "just", "do not edit", "minimal change", "report only", "no refactor", "don't touch X", or otherwise draws a narrow boundary; and as a check before you touch any file the request did not name. Pulled on demand; not always-on.
elon-choo/fablever · ★ 33 · AI & Automation · score 77
Install: claude install-skill elon-choo/fablever
# fable-scope-guard — change only what was asked, nothing adjacent This skill encodes fablever's single most robust measured behavior: **scope discipline**. In the style-only ablation (`eval/style-only-ablation/RESULTS.md`, deterministic, no judge) the Fable working style held **0% scope violations vs plain Claude's 42%**. That is the win to protect here — it is a discipline, not a capability, so it only holds if you apply it deliberately. The instinct to "improve while I'm here" is the failure mode. A bug fix does not need surrounding cleanup; a one-shot edit does not need a new helper or abstraction. ## When to use this - The user set a narrow boundary: "only", "just", "minimal", "report only", "do not edit / refactor / touch", "don't add dependencies". - You are about to edit, create, or delete a file the request did not explicitly name. - A linter/build error tempts you to change working code that is not part of the task. ## When NOT to use this - The user explicitly asked for a refactor, cleanup, or broad change — then breadth IS the scope. - A genuine system boundary (real user input, an external API) needs validation the task implies. ## Procedure 1. **Write the boundary down.** In one line, state what is in scope and what is explicitly out. If the user used "only/just/report-only", treat everything not named as out. 2. **Touch only named surfaces.** Edit only files the task names or that are unavoidably required to make the named change compile/run. Do