swyxio Skill Cutter
Reduce context cost and accidental activation without weakening the behavior the skill exists to provide. Treat line count as evidence, not the objective.
Choose the mode
- Audit: When asked to review, critique, or identify cuts, inspect and report without editing.
- Cut: An explicit request to cut, trim, shorten, simplify, or de-slop the skill authorizes local edits within that skill. It does not authorize commit, push, publication, or changes to consumers outside the requested scope.
Find the behavioral core
Read the complete SKILL.md, its UI metadata, and only the directly relevant
resources. Identify the concrete tasks that should trigger it and the decisions
an otherwise capable agent would get wrong without it.
Classify material before cutting. Use acceptance force as well as content class so useful advice does not survive as accidental governance:
| Acceptance force | Treatment |
|---|---|
| Invariant | Keep explicit and narrow. |
| Action-required | Keep only for the named task type. |
| Risk-triggered gate | Name the triggering risk and proportional proof. |
| Recommendation | Label or phrase as non-blocking; condense when useful. |
| Opportunity | Move out of the critical path or delete. |
Then classify its content:
| Class | Default treatment |
|---|---|
| Non-obvious domain constraint or fragile required sequence | Keep |
| Useful optional expert recommendation | Condense and label as optional |
| Generic knowledge or ordinary engineering advice | Delete |
| Policy already enforced by system or repository instructions | Delete |
| Duplicate instruction or example | Keep the clearest instance |
| Project- or incident-specific policy in a generic skill | Move or delete |
| Stale, unverifiable, or overclaimed fact | Verify, qualify, or delete |
| Detail needed only for one variant | Move to a selectively loaded reference |
Report which class justified every material keep, move, or deletion. Do not hide subjective policy behind claims that a provider or tool requires it.
Use this acceptance model when auditing a skill:
User outcome + higher-level invariants + risks created by the action
= blocking acceptance criteria. Everything else is advice or follow-up.
Flag checklist language that silently promotes recommendations, asks a broken subsystem to approve its own repair, coordinates unrelated or read-only work, or keeps gathering proof after the requested outcome is established. For a bounded task, more than one or two skill-added gates needs a direct correctness, privacy, security, data-integrity, or irreversible-action justification.
Cut trigger aggression
The frontmatter description is always-loaded routing context. Make it narrow and concrete:
- name positive user intents and artifacts that should activate the skill;
- remove phrases such as “any task,” “all requests,” and broad product-name-only triggers unless universal activation is truly intended;
- add concise exclusions for nearby tasks that should not activate it;
- move no trigger rules into the body, which is read only after activation; and
- keep
agents/openai.yamldisplay text and default prompt aligned with the narrowed contract.
Do not make the description so timid that explicit requests stop matching.
Apply the cut
Prefer deletion over compression. Preserve:
- task-specific decision rules and failure modes;
- domain-specific safety or authorization boundaries;
- exact tool or file contracts that are easy to misuse; and
- routing to resources that are genuinely needed conditionally.
Remove:
- explanations written for a novice human when the agent already knows them;
- exhaustive product catalogs, copied manuals, and speculative edge cases;
- mandatory-sounding ceremony that applies only to one project or incident;
- repeated summaries, principles, checklists, and completion language; and
- historical justification that does not change future execution.
Also remove or reclassify adjacent improvements presented as prerequisites, global coordination without a concrete shared mutation surface, and exhaustive verification whose result cannot change acceptance of the requested outcome. Preserve exact-target, authorization, privacy, secret, user-data, migration, rollback, and irreversible-action controls where the named risk exists.
Do not preserve text merely because it is correct. Do not replace readable
instructions with dense slogans, and do not move bulk into references simply to
make SKILL.md look shorter. Do not delete scripts, assets, or operational
references based only on apparent non-use; inspect their callers and purpose.
For unstable provider claims, check current primary documentation. Separate provider constraints from optional recommendations and local policy.
Validate and report
Preserve unrelated work. After editing, update broken links and UI metadata, run the repository's skill validator, and inspect the final diff.
Report:
- before/after lines, words, and files;
- the retained behavioral contract;
- a compact keep/condense/move/delete classification with reasons;
- trigger changes and exclusions;
- validation performed; and
- uncertain material deliberately left untouched.
For acceptance-model audits, also report the stop condition, any skill-added blocking criteria, and why each remaining gate is action-required or risk-triggered.
A successful cut is smaller and less eager to trigger while still changing the agent's behavior in every case the skill was created to handle.