Toolkit Skill Author

toolkit-skill-author

Pre-load

bashCopy code
!test -f CONTRIBUTING.md || (echo "CONTRIBUTING.md missing — run from repo root." && exit 1)

Verify skill-creator is callable in this session. Scan the available-skills list for a skill named skill-creator (bare or namespaced, e.g. skill-creator:skill-creator). If absent, abort with exactly this message:

skill-creator plugin is required. Run /plugin install skill-creator@claude-plugins-official, enable it, then restart this session and re-run /toolkit-skill-author.

Do not fall back to manually scaffolding SKILL.md — the handoff is core to the workflow.

Load authoritative context:

bashCopy code
!cat CONTRIBUTING.md
!bash .claude/skills/toolkit-skill-author/scripts/find-peers.sh skills

• @.claude/skills/toolkit-skill-author/references/decision-rules.md
• @.claude/skills/toolkit-skill-author/references/handoff-preamble.md
• @.claude/skills/toolkit-skill-author/references/registration-checklist.md

Phase 0 — Parse intent and apply gates

From the contributor's initial prompt, extract (without asking yet):

• target_name — did they name a specific skill (e.g., "extend monitoring-advisor", "create a skill called foo-bar")?
• action — extend, new, or unknown.
• bucket — did they state a capability bucket?

Apply the gates below before any survey question. Each gate halts when it fires.

Gate A — Forbidden bucket

If bucket is Agent-routing (or a synonym like "agent routing", "routing skill"), refuse:

Agent-routing skills are owned by the toolkit core team per CONTRIBUTING § Capability buckets — not authored via /toolkit-skill-author. Halting.

Do not proceed. If bucket is unstated, this gate re-applies after Q1.

Gate B — Name collision (new-skill intent)

If action is new and target_name is given, lowercase both target_name and each existing skill directory name, then check:

• Exact match (equal strings): refuse with "<target_name> already exists in skills/. Pick a different name, or switch to extend." Do not proceed.
• Near match. Split both names on - into token lists. Fire if either (a) the two lists share any token, or (b) any token in one list is a substring of any token in the other list. Example: monitor-advisor ↔ monitoring-advisor fires via (a) shared advisor and (b) monitor ⊂ monitoring. When fired, surface the overlap and ask: "<target_name> overlaps with existing <existing>. Did you mean to extend <existing>, or proceed with a new name?" Wait for an answer before continuing.

Gate C — Fast-path for clear extend

If action is extend and target_name names an existing skill:

• Re-check the bucket. If target_name is one of the agent-routing skills (context-detection, incident-response, proactive-monitoring), fire Gate A's refusal — those are owned by the core team and cannot be extended via /toolkit-skill-author either.
• Otherwise: skip the Phase 1 survey, use the initial prompt as the extension description, and jump directly to Phase 2a — Extend. If the prompt is missing fields the handoff preamble needs (purpose, phrasings, output artifact), ask only for those specific fields — do not re-run the full Q1–Q4 survey.

Otherwise, continue to Phase 1. New-skill requests always run the decision tree — the gate against hidden collisions is the tree itself.

Phase 1 — Decision survey

Ask these four questions one at a time, waiting for each answer.

1. Capability bucket. Trust / Incident Response / Monitoring / Prevent / Optimize / Setup. Agent-routing is not offered. If the contributor writes in Agent-routing, fire Gate A and halt.
2. Primary MCP surface or data input. E.g. Monte Carlo GraphQL, BigQuery INFORMATION_SCHEMA, Sentry issues. One or two items.
3. One-line purpose. Plain language — no pushy triggers.
4. 2–3 example user prompts. Quote them literally.

Decision

Apply the 4-step test from decision-rules.md against the peers dumped by find-peers.sh. The dump shows every skill's name + description + when_to_use; reason directly about which skills could plausibly activate on the Q4 prompts.

Filter agent-routing skills out of the candidate set — contributors can't author those, so they can't be extension targets. Current list: context-detection, incident-response, proactive-monitoring. If the toolkit adds new routing skills, update both this list and Gate C's list to match.

Present the verdict:

• EXTEND <peer> or NEW SKILL.
• The step that decided, and why.
• "Proceed with this verdict, or override?"

If the user overrides, capture their reason verbatim — it becomes part of the PR description per CONTRIBUTING.

Phase 2a — Extend

Invoke skill-creator via the Skill tool in improve-existing mode, passing the handoff preamble from references/handoff-preamble.md with MODE=IMPROVE_EXISTING, PEER_NAME=<peer>, and the survey answers (or the initial prompt for the fast-path case).

skill-creator runs its full loop; if the contributor wants a lighter-touch edit, they can tell it mid-flow to skip iteration.

When skill-creator returns:

bashCopy code
!rm -rf skills/<peer>/evals/ skills/<peer>-workspace/
!python3 .claude/skills/toolkit-skill-author/scripts/lint-skill.py <peer>

Scratch-artifact cleanup removes skill-creator's iteration files (they drove the loop but aren't the repo's eval format). Lint surfaces any frontmatter violations.

Walk the partial checklist from references/registration-checklist.md via TodoWrite:

• Signal-definitions update — only if phrasings shifted.
• /mc catalog update — only if user-facing surface changed.
• Eval entry update — only if activation surface expanded.

Continue to the shared version-bump step.

Phase 2b — New skill

Extended survey (one at a time)

5. Output artifact — what the skill produces (file type, diff, notebook, text verdict).
6. Persona / workflow — who invokes this and during what task.
7. Disambiguation — how it differs from the nearest peer (by bucket), even if no peer forced a split.

Propose name

Suggest 2–3 kebab-case directory names. Short is better. Re-run Gate B against each candidate:

• Exact match → drop the candidate.
• Near match (token-overlap with any existing skill) → drop or explicitly flag.

Only present candidates that survive Gate B. Wait for the contributor to pick one. Before handoff, re-verify skills/<chosen>/ does not exist.

Note the two-level naming: the directory is skills/<chosen>/, but the name field inside the generated SKILL.md frontmatter should be monte-carlo-<chosen> — the canonical prefixed form. The handoff preamble passes both, and lint-skill.py verifies them after scaffold.

Handoff to skill-creator

Invoke via the Skill tool with the handoff preamble from references/handoff-preamble.md, MODE=NEW_SKILL, and all survey answers.

When skill-creator returns:

bashCopy code
!rm -rf skills/<name>/evals/ skills/<name>-workspace/
!python3 .claude/skills/toolkit-skill-author/scripts/lint-skill.py <name>

If lint prints ERROR lines, surface them and wait for the contributor to fix (manually or via a regenerate pass) before proceeding.

Full registration checklist

Walk the full checklist from references/registration-checklist.md via TodoWrite, one item per step:

1. Read the relevant existing file.
2. Draft the addition.
3. Show the diff and confirm.
4. Apply with Edit, or create new files with Write.

If Q1 = Setup, confirm: "Setup skills are exempt from signal-definitions and /mc catalog registration per CONTRIBUTING. Skip those two steps? [Y/n]" — default Y.

Shared — version bump

Both phases converge here.

1. Propose a level based on what actually changed:
   • New skill → minor per CONTRIBUTING § Version bumping.
   • Extend changed activation surface (phrasings in description or when_to_use) → minor.
   • Extend did not touch activation surface → patch.
2. "Proceed with this level, or override?" Valid overrides: patch, minor, major.
3. Run:

   bashCopy code
   !./scripts/bump-version.sh <level>
   

   The script opens $EDITOR for the changelog, then updates all 6 plugin configs and all 6 CHANGELOG.md files.

Done

Tell the contributor:

All changes are staged. Review with git status / git diff --staged, then commit and run /ship to open the PR.

Do not auto-commit or push.