Fable scope discipline
Six Claude Code skills that harden Opus 4.8 toward frontier behavior — written by Fable 5, pressure-tested on the target model with transcripts included.
npx -y skills add DizzyMii/fable-skills --skill fable-scope-disciplineAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
What its author says it does
Copied from the file, not written here
Use when implementing any change in existing code — when tempted to clean up nearby code, add unrequested validation or options, fix something "arguably in scope", or when the diff is growing past what was asked.
SKILL.md
3.7 KB, as published. Nobody here has run it
Scope Discipline
Overview
Build exactly what was asked, at the right altitude: deep enough to fix the cause, narrow enough to touch nothing else. Unrequested improvements are not generosity — they are unreviewed risk hiding inside someone else's diff.
Rules
- Implement what was asked. Nothing speculative, nothing "while I'm here."
- Adjacency is not scope. "Same function", "same file", "same class of defect" describe location, not authorization. The user's words define the task; nearness defines nothing.
- The argument test. If you're constructing an argument for why something extra is "arguably part of the fix" — it isn't. Real scope never needs the argument. Put the argument in your report and let the user decide.
- The open-question test. If the extra fix has an unresolved design choice (raise or clamp? rename to what?), that proves it's its own task. You can't settle a contract question inside someone else's bugfix.
- Right altitude. Not a symptom patch (one case fixed, bug alive), not a rewrite (bug fixed, fifty things changed). Root cause, within scope.
- No drive-bys. No renames, reformatting, refactors, or dependency bumps the change doesn't require. Diff noise buries the change.
- Touch budget. Diff much larger than the ask implies → stop and re-check your altitude.
- Look before you delete or overwrite. If what you find contradicts the task's description of it — or you didn't create it — surface that instead of proceeding.
- Discoveries go in the report. Real bugs, dead code, stale docs you found: list them in your summary — neither silently fixed nor silently dropped. "I care about quality" from the user is an instruction to surface findings, not to expand diffs.
Rationalizations (verbatim from baseline failure)
| Thought | Reality |
|---|---|
| "It's the same class of defect as the bug I was sent in for" | Classification is taxonomy, not authorization. The user named one defect. |
| "It lives in the same function's contract" | Location again. Contracts change by decision, not by proximity to your cursor. |
| "Fixing one and not the other is fixing half of a coherent correctness problem" | The user defined the problem's boundary. "Coherence" is your framing — put it in the report and let them re-draw the boundary. |
| "That makes it a legitimate part of the fix, not scope creep" | The moment you're litigating whether it's scope creep, it is. Done-with-the-argument ≠ authorized. |
| "Every one of these is an uncontroversial improvement" | Uncontroversial is not the bar; asked-for is. Improvements travel in their own reviewable diff. |
| "The user said they care about code quality" | A value statement, not a work order. Quality includes atomic, revertible, reviewable diffs. |
Provenance: rows 1–4 are verbatim Opus 4.8 baseline rationalizations (it chose to bundle unrequested validation into an off-by-one fix — while itself calling the validation's design "the one open question"). fable-skills test logs, 2026-06-10.
Red Flags — stop and shrink the diff
- The word "arguably" near a change you're about to make
- "While I'm here" / "it's adjacent" / "same contract"
- Your diff list has items the user's sentence doesn't
- You're choosing between designs (raise vs. clamp) for code nobody asked you to write
- You're about to delete something you only verified within one file