Add Skill

add-skill

Builds one production-quality skill in skills/<name>/, eval-first. Read docs/skill-authoring.md and docs/evals.md once per session before starting; they are the rulebook this workflow enforces. If invoked as /add-skill <args>, treat $ARGUMENTS as the skill name or idea to scope in step 1.

When NOT to use

• Creating a new skills repository → that happens in the workspace above this repo.
• Infrastructure work (hooks, CI, validator, docs) → plain editing, PR-level scrutiny.
• A "skill" that merely restates what the model already does well → don't build it; a skill must fix a specific observed failure.

Workflow

1. Scope, and pick the invocation mode. State in one sentence what the skill does and the concrete failure it fixes. If the sentence needs "and", split into multiple skills and do them one at a time. Check skills/ and the README catalog for overlap — near-neighbor descriptions steal each other's triggers; adjust scope or plan disjoint wording. Then decide who invokes it, because it changes the name, the description, and the eval set: model-invoked (default) or user-invoked (disable-model-invocation: true — the user types /name; correct for side-effecting skills and memorable commands). See the mode table and the alias-over-engine split in docs/skill-authoring.md.
2. Gather. Read .local/ recursively — every subfolder and file is source material (research, examples, constraints). Then research beyond it: web-search the topic for current facts, official docs, and prior art; verify anything load-bearing against primary sources, and cross-validate contested or high-stakes facts (external APIs, schemas, security claims) with /cross when available. Verified facts a skill depends on go into skills/<name>/references/ as cleaned, tracked files — never cite .local/ paths from a skill.
3. Evals first. Create skills/<name>/evals/evals.json per docs/evals.md: ≥8 should_trigger (vary formality, typos, terseness), ≥8 should_not_trigger (near-misses sharing keywords), 3–5 quality cases with plain-language expected_behavior assertions. If you can't write these, the scope is unclear — go back to step 1. For a user-invoked skill both trigger counts are waived — no router weighs its description — so write the quality cases only.
4. Draft SKILL.md. Frontmatter: name == folder name; description on a single line, third person. Model-invoked: [what] + [use when …] + [not for …], 150–400 chars, trigger keywords in the first ~120. User-invoked: one verb-first line that reads well in a / menu. Body per the structure in docs/skill-authoring.md: purpose → when/when-not → numbered workflow → output spec → gotchas → pointers. Under 500 lines; deterministic steps as scripts/ (non-zero exit on failure); long material as references/ with a TOC.
5. Validate. make check must exit 0 — it runs the validator, then scripts/run_evals.py, which scores every trigger prompt against every sibling description and fails on no-overlap prompts, a sibling outranking the skill, or a near-duplicate description. The write-time hook already validated the file; fix every error and triage every warning. Fix the description it points at, never the evals.
6. Trigger self-test. For at least 3 should-trigger and 2 should-not-trigger prompts, reason explicitly: would the description alone (not the body) route this prompt here, against every other skill in the catalog? Fix the description, not the evals. Recommend the user run 2–3 fresh-session probes for the riskiest cases. (Skip for user-invoked skills — nothing routes them.)
7. Register. Add/update the skill's row in the README catalog table, add a CHANGELOG.md entry, and put the skill in exactly one skills.sh.json grouping (engaging one-sentence group descriptions — that file is the repo's skills.sh listing copy). The schema shape is "groupings": [{"title": ..., "description": ..., "skills": [...]}] — the schema rejects other keys (groups/name) and skills.sh silently ignores a non-conforming file; make check validates it. Re-check descriptions of sibling skills for new overlap. Do not commit or push — leave all changes for the owner to review.

Output spec (Definition of Done)

• skills/<name>/ with SKILL.md + evals/evals.json (+ scripts/references as needed)
• make check → 0 errors; README catalog + CHANGELOG updated
• A 3-line summary: purpose, trigger phrase examples, known limitations

Gotchas

• Keep description on ONE physical line — multi-line values have failed to load in some runtimes and the validator blocks them; never work around the validator.
• Don't paste failed eval prompts verbatim into the description (overfitting); generalize the missing verb/noun instead.
• Over-triggering fix = add targeted "Not for …" noun phrases from the actual false-positive overlap; don't delete positive triggers.
• Model upgrades shift routing — note in your summary that evals should be re-run after the next model release.