skill-cutter

Skill Cutter

Reduce context cost and accidental activation while preserving the decisions that make the skill useful. Length is evidence, not a target.

Establish scope

• Audit requests authorize inspection and findings. Do not edit the target.
• An explicit request to implement findings, cut, trim or simplify authorizes local edits to the named skill scope, not publication or consumer installation.
• Read the complete SKILL.md, available UI metadata and directly relevant resources. Preserve unrelated work and identify the user's requested outcome.

Classify before editing

Keep non-obvious domain constraints, fragile required sequences, exact file/tool contracts and action-specific authorization or data-integrity controls.

Material	Treatment
Useful optional expert advice	Condense and mark non-blocking.
Generic knowledge or repeated policy	Delete.
Duplicate example or instruction	Keep the clearest useful instance.
Incident/project-specific rule in a generic skill	Move to its actual scope or delete.
Stale or unsupported claim	Verify with primary documentation, qualify or delete.
Useful detail for one variant	Link a selectively loaded reference.

Do not move bulk merely to reduce the entrypoint's line count. Inspect callers before deleting scripts, assets or references; apparent non-use is insufficient.

Keep acceptance proportional

Blocking criteria come from the user outcome, higher-level invariants and risks created by the action. Distinguish:

• Invariant: explicit and narrow, such as preserving sample identity.
• Action-required: applies only to the named task, such as a file contract.
• Risk-triggered gate: names the risk and proportional proof.
• Recommendation/opportunity: advice or follow-up, not a prerequisite.

Flag unrelated coordination, optional infrastructure presented as mandatory, repeated approval despite established authorization, self-approval by a broken subsystem, and proof gathering that cannot change acceptance. More than one or two skill-added gates for a bounded task needs a concrete correctness, privacy, security, data-integrity or irreversible-action reason.