new-skill

new-skill

CLAUDE.md → "Adding a New Skill" is the authoritative checklist. This skill exists to make sure every step actually runs — the session-start loop and the eval cases are the ones most often forgotten.

Read CLAUDE.md's "Adding a New Skill" section now, then execute its steps for logic-<NAME> (ask the user for <NAME> if not in $ARGUMENTS):

1. skills/logic-<NAME>/SKILL.md — copy the frontmatter shape from skills/logic-review/SKILL.md (name + description). The description MUST include a "Do NOT trigger for:" clause and a "SCOPE … RULE" line distinguishing it from the other skills. Process section = 5–7 numbered items, each citing a guide step. Follow lazy-loading per skills/_shared/common.md §13.
2. skills/logic-<NAME>/logic-<NAME>-guide.md — numbered steps (Step 1, 2, …), each explaining the why, concrete examples over abstractions.
3. commands/logic-<NAME>.md — mirror commands/logic-review.md:

   ---
   description: <one line>
   allowed-tools: Skill
   ---
   Use the Skill tool to invoke `logic-lens:logic-<NAME>`, then follow its instructions exactly. $ARGUMENTS
   

4. hooks/session-start (CLAUDE.md step 6 — easy to miss) — add logic-<NAME> to BOTH the for skill in … loop and the printed skill list:

   grep -n "for skill in" hooks/session-start
   

5. Eval cases (CLAUDE.md step 7 — easy to miss) — add 3–5 content cases to evals/content/v2/evals-v2.json. Trigger cases under evals/trigger/v2/ are optional and only catch description regressions (positive cases always score 0% — see CLAUDE.md → Gotchas).
6. gemini-extension.json — CLAUDE.md says Gemini auto-discovers from skills/, but the file also enumerates each skill+command under contribution. If you want that list complete, add logic-<NAME> entries to both contribution.skills and contribution.commands. (Not checked by validate-repo.sh.)
7. Validate — run npm run validate and confirm all checks pass.