Skill issue

Skill Issue

Reduce the skill to the smallest intervention that demonstrably improves its target task.

Assume the model is innocent until the skill proves otherwise.

Inspect

1. Read the target SKILL.md and every resource it directly routes to.
2. Identify the target task, supported harnesses, trust boundaries, and observable completion condition.
3. Run node scripts/audit-skill.mjs <skill-directory> from this skill directory when Node.js is available. Treat its findings as leads, not semantic verdicts.
4. Establish what the current base model already does without the skill. Do not infer this from reputation or model version.

Classify

Assign every material instruction one disposition:

• Keep: non-obvious domain facts, exact tool routing, trust or data-loss boundaries, deterministic operations, output contracts, and observable verification.
• Trim: generic advice, duplicated defaults, repeated warnings, fixed model or tool assumptions, and process whose cost exceeds the failure risk.
• Delete: instructions with no target capability, conflicts with the current environment, or no measurable benefit over baseline.
• Measure: plausible value that lacks paired evidence.

Read references/decision-rubric.md when a classification is disputed or the skill mixes domain knowledge and workflow.

Never trim input validation at trust boundaries, security controls, accessibility basics, error handling that prevents data loss, or an explicit user requirement.

Patch

1. Require name and description. Preserve recognized optional or harness-specific fields when they carry licensing, compatibility, tool restrictions, security boundaries, or useful metadata. Remove a field only after classifying its behavior and portability cost.
2. Put all positive triggers and nearby exclusions in description.
3. Keep the body imperative and focused on one job.
4. Move conditional detail one level down into references/; do not duplicate it in the body.
5. Keep scripts only for repeated or deterministic work. Add no dependency when the standard library suffices.
6. Preserve user-authored behavior unless the user asked for an edit. For an audit-only request, return the patch proposal without changing files.

Verify

1. Run the platform's skill validator when available.
2. Rerun audit-skill.mjs and report before/after lines, words, estimated tokens, resources, and unresolved findings. Include the exact command and counting method; retain the raw report or its path when results will be compared.
3. For a material behavioral change, use $trust-me-bro when available. Otherwise run the same pinned task once without the skill and once with it, keep contexts isolated, and verify outputs from artifacts rather than self-reports.
4. Mark efficacy as unproven until a paired trial supports it.

Return

Lead with keep, trim, delete, or measure. Include:

• the capability the skill uniquely supplies;
• proposed or applied cuts with reasons;
• preserved safety boundaries;
• before/after instruction cost;
• the command, counting convention, and raw artifact behind those numbers;
• validation and trial evidence;
• remaining uncertainty.

Stars, prose length, and the implementer's confidence are not evidence.