skills/ — CONTRIBUTOR CONTRACT

67 dirs, one contract. Score for this file: uniform structure across many dirs (root covers the rest).

OVERVIEW

Each skill = skills/<name>/SKILL.md (+ optional scripts/, references/, LICENSE.txt). Instruction-only skills need no code; script-backed skills ship a stdlib-only Python CLI (pro-review is the exception: zero-dependency Node .mjs).

ADDING A SKILL

1. SKILL.md frontmatter — required, test-enforced (tests/test_skills.py):

   ---
   name: <exactly-the-dirname>
   description: <what + when-to-use, >=40 chars; this is the routing surface>
   version: 1.0.0
   license: <SPDX or "MIT (from owner/repo; see LICENSE.txt)">
   metadata:
     agentskills:
       tags: [a, b, c]
   ---

2. Ported skill: copy upstream LICENSE into <name>/LICENSE.txt; de-brand (harness tools as examples only, plain-text fallback); drop binaries >100K and eval artifacts.
3. Script-backed: follow the CLI contract (argparse, - stdin, --json, rc 0/1/2, gates strict >); add tests in tests/test_skills.py using the run() helper; isolate HOME-touching tools with a tmp_path HOME env.
4. Register: three manifests (.claude-plugin/marketplace.json, .claude-plugin/plugin.json, .codex-plugin/plugin.json — same version bump) + README family table + count.

WHERE TO LOOK

ANTI-PATTERNS

• Overlapping skill purposes: every skill must answer a question no existing skill answers (secret-gate=credentials-in-diffs vs comment-checker=comments-in-diffs; usage-audit=token-totals vs session-handoff=conversation-content vs context-budget=static-estimate). Check the inventory before adding; shared helper NAMES are fine (skills are copy-portable and never import each other), shared PURPOSE is not.
• Frontmatter name != dirname (test fails).
• Description under 40 chars (test fails) or missing when-to-use routing cues.
• Adding pip dependencies to a guard CLI.
• Editing design/references/* prose — it is verbatim Anthropic guidance; patch only portability seams.