Collide

Collisions and drift (collide)

Two skills whose descriptions say nearly the same thing will fight over the same request, and the loser looks broken. This measures that overlap, and separately checks whether skills copied into several agent homes are still the same file. Local only, no quota spent.

Language rule: write every user-facing output in the language the user is speaking with you. Evidence rule: a collision claim must name the shared terms. "These two feel similar" is not a finding.

Step 1 — Measure

node "${CLAUDE_PLUGIN_ROOT}/scripts/skillbench.mjs" --collide --md

Add --path <repo> to check only the skills in one repo — the useful mode when adding a skill next to existing ones. Requires Node.js 18+.

Scoring is an idf-weighted cosine over the descriptions: rare shared words count for more than common ones, so "evidence" counts and "the user" does not. Copies of the same skill in different agent homes are excluded from collisions by name and handled as drift instead.

Step 2 — Judge each pair, do not report the score

A score is a lead. For each pair above the threshold, open both descriptions and decide:

• Real collision — both would plausibly answer the same user sentence, and neither says what it is not for. This is the case worth fixing.
• Family, correctly separated — same domain, but the descriptions already name the split ("Not for X — that is "). The bothHaveBoundary column tells you this at a glance. Report as healthy, not as a problem.
• Vocabulary artefact — the overlap comes from shared jargon, not shared purpose. Drop it and say you did.

Concretely: read the two descriptions and try to write one sentence that only one of them should answer. If you cannot, it is a real collision.

Step 3 — Fix by adding the boundary, not by rewording both

The reliable repair is asymmetric and small:

1. Decide which skill owns the contested request. Say why in one line.
2. In the other skill's description, add an explicit exclusion naming the winner: Not for <case> — that is <winner>.
3. In the winner's description, add the trigger phrases that were pulling toward the loser.
4. Add the contested sentence to the loser's should_not_trigger and the winner's should_trigger eval list, so the fix is checkable next time.

Rewriting both descriptions at once usually creates a new collision somewhere else. Change one side and re-run.

Step 4 — Drift

The drift table lists skills that exist in more than one agent home with different content. For each, say which copy is newest (line counts and the hashes are in the report; check file times if you need to) and offer to re-sync from that one. Never overwrite a copy without the user confirming which side is authoritative — the divergent copy may hold the newer work.

If there is no drift, say so in one line: identical copies across agents is the good outcome and worth confirming explicitly.

Step 5 — Reply

Short: how many pairs were above threshold, which are real collisions after your judgement (and how many you dropped as artefacts), the single boundary line you recommend adding and to which skill, and the drift status. If nothing collides, say that plainly and do not manufacture near-misses.