skill-consolidator（技能巩固者）

Skill Consolidator

Create our own “best version” skills by consolidating globally installed skills into a smaller, linked set inside this repo, without directly editing the global skills directory.

When to Use This Skill

Use this skill when the user:

• Wants to consolidate overlapping skills into one first-party skill
• Says “merge skills”, “remove duplicates”, or “organize my skills”
• Wants to create a custom skill by strategically combining third-party skills
• Has a cluster of similar skills (e.g. Playwright) and wants one canonical version
• Needs a safe workflow for reviewing and decommissioning old skill copies

Safety & constraints

• Treat globally installed skills as read-only (default target: ~/.agents/skills, or $CODEX_HOME/skills, or ~/.codex/skills).
• Prefer duplication over mutation: copy skills into this repo before editing.
• Never delete skills as part of consolidation; deprecate via documentation and links instead.

Skill locations (repo vs global)

• Repo first-party skills (source of truth): ./skills/ (edit here first)
• Global installed skills (source of truth): ~/.agents/skills/

- ~/.codex/skills/ often contains symlinks pointing into ~/.agents/skills/

Quick start

1. Generate an overlap report from globally installed skills (read-only):

- python3 skills/skill-consolidator/scripts/scan_installed_skills.py --out-dir tmp/skill-consolidator

1. Pick the first cluster in tmp/skill-consolidator/overlap_report.md.
2. Build the canonical result as a first-party skill in skills/<new-skill-name>/.

What “counts” as a good first-party skill (required standard)

Our skills are intentionally docs-driven: SKILL.md + bundled references/ + optional scripts/ are the “product”.

Before considering a consolidation “done”, the first-party skill must meet:

• The required standard: skills/skill-consolidator/references/first-party-skill-standard.md

Workflow

1) Vendor upstream skills (third-party and/or personal global skills)

Goal: treat upstream skills as read-only; copy locally only if you truly need diffs.

Common sources:

• Global: ~/.agents/skills/ (recommended global home)
• Codex links: ~/.codex/skills/ (often symlinks into ~/.agents/skills/)

Recommended default: don’t copy; reference upstream via ~/.agents/skills/.

2) Inventory + overlap report

Run the scanner to produce:

• inventory.json (machine-readable)
• overlap_report.md (human-readable clusters)

Examples:

• Scan globally installed skills directly (read-only):

- python3 skills/skill-consolidator/scripts/scan_installed_skills.py --out-dir tmp/skill-consolidator

• Scan a custom root:

- python3 skills/skill-consolidator/scripts/scan_installed_skills.py --skills-root /path/to/skills --out-dir tmp/skill-consolidator

2) Pick a consolidation strategy per cluster

For each overlapping cluster, choose one primary pattern:

• Merge into one canonical skill: best when content is truly duplicative.
• Hub + leaf skills: best when workflows differ but share concepts (create a hub that routes to specialized skills).
• Shared references: best when multiple skills share long “best practices” sections (extract into references and link).

Use: skills/skill-consolidator/references/combining-best-practices.md for decision rules and anti-patterns. Use: skills/skill-consolidator/references/opinionated-defaults.md to align the “canonical” result with our preferred tech stack and tooling defaults (when relevant).

3) Build the first-party canonical skill in ./skills/

• Create skills/<canonical-skill-name>/SKILL.md that:

- Routes common trigger phrases (in description) - Links to leaf skills or reference docs - Keeps detailed content in references/ and reusable code in scripts/ - Meets the required standard in skills/skill-consolidator/references/first-party-skill-standard.md

If you keep old skills around inside this repo:

• Narrow triggers or remove their SKILL.md so they don’t compete with the canonical skill.

4) Consolidate content with progressive disclosure

• Keep SKILL.md short and procedural.
• Move detailed “how-to” / long reference content into references/.
• When two skills share the same long section, keep it in one canonical location and link to it (do not duplicate).

If you need a shared, cross-skill home for references inside this repo, use a _shared/ folder (see skills/skill-consolidator/references/combining-best-practices.md).

5) Link skills together (without creating deep dependency chains)

• Add a small “See also” section (a few bullets) at the bottom of related skills.
• Avoid multi-hop references (Skill A -> Skill B -> Skill C). Keep links one hop from the skill you’re editing.

6) Install the first-party skill globally (Skills CLI)

After the canonical skill lives in this repo, install it like other skills using the Skills CLI:

• npx skills add <owner/repo@skill-name> -g -y

(This assumes the repo/skill is available from the source you use with npx skills add.)

7) Verify changes

At minimum:

• Re-run the scanner to ensure consolidated skills still parse cleanly.
• Run the script tests:

- python3 -m unittest skills/skill-consolidator/scripts/test_scan_installed_skills.py

If packaging scripts exist in the user’s environment, also validate packaging for the new/edited skills.

8) Second-agent review (required)

Before decommissioning any upstream skills, run a second-agent review of the consolidation:

• Focus: “Did we lose important capability? Are triggers safe? Are links and routing clear?”

Use: skills/skill-consolidator/references/review-and-decommission.md for the review checklist and a copy/paste review prompt.

9) Human check (recommended default)

After the second-agent review passes, do a quick human check to confirm:

• The canonical skill triggers the right scenarios
• The old skills won’t compete with it
• The consolidation is reversible

Use: skills/skill-consolidator/references/review-and-decommission.md.

10) Decommission upstream copies (safe + reversible)

Do not delete upstream skills immediately.

Prefer one of:

• Remove/rename SKILL.md files in any local copies so they can’t trigger
• If you keep old first-party copies around, mark them deprecated and narrow triggers

Use: skills/skill-consolidator/references/review-and-decommission.md for rules and acceptance criteria.

Note: when scanning this repo’s skills/, the overlap scanner won’t include global skills unless you explicitly scan them.