fable-scope-guard
Skill33 starsChanged 5 months agoMITView on GitHub
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.
In this repository8 other files it gives its agents.
- AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- llms.txtllms.txt
- fable-delivery-gateSkill
- fable-evidence-doneSkill
- fable-planSkill
- fable-reviewSkill
- fable-seedSkill
The file
We split this skill into the rules it gives an agent
3 rules. About 0 are common, 3 are distinctive, and 0 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈3
Distinctive · estimate
This author's own choices. Worth a look.
0
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 34
No adjacent work: No renames, reformatting, dependency additions, or "drive-by" fixes to unrelated code.
- Distinctive: line 38
Prefer editing over creating: If the change fits an existing file, do not introduce a new one.
- Distinctive: line 47
Before finishing: re-read the request's boundary line.
≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
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
- 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.
- Touch only named surfaces. Edit only files the task names or that are unavoidably required to make the named change compile/run. Do not create README/docs/helpers unless asked.
- No adjacent work. No renames, reformatting, dependency additions, or "drive-by" fixes to unrelated code. If you spot a real separate problem, note it in your report — do not fix it now.
- If a fix requires touching working code outside scope, stop and ask. State the file, the change, and why; wait for confirmation rather than expanding scope yourself.
- Prefer editing over creating. If the change fits an existing file, do not introduce a new one.
Expected output
The minimal diff that satisfies the request, plus a one-line "out of scope, noted but not changed:" list for anything you deliberately left alone.
Verification
Before finishing: re-read the request's boundary line. Does every changed file appear in it (or is it strictly required by it)? If any changed file is not justified by the boundary, revert that change.
Failure cases
- "While I was there I also tidied up X" → a scope violation, even if X is better now.
- Adding a dependency, config, or abstraction the task did not ask for.
- Refactoring working code to fix a linter warning without flagging it first.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

