Stop guessing which agent skill to use.

skill-compass checks what you have installed, picks the right skill for the task and flags broken installs.

> What skills do I have?real output, shortened

$ node skills/skill-compass/scripts/inventory.mjs --text
Skills (20)
  code-review  [claude-code]
    Review the changes since a fixed point along two axes.
  diagnosing-bugs  [claude-code]
    Diagnosis loop for hard bugs and performance regressions.
  find-skills  [shared]
    Helps users discover and install agent skills.
  grill-me  [claude-code, manual only, shortcut for grilling]
    A relentless interview to sharpen a plan or design.
  prisma-cli  [codex, cursor, opencode]
    Prisma ORM CLI commands reference.
  ...
Works with

Your agent answers from memory.

I had skills installed for Claude Code, Codex and Cursor, spread across folders I had forgotten about. When I asked my agent what I had, it guessed.

It missed skills in other agents' folders, said a project had no broken links when it had nine, and told me to delete a shortcut that was working fine.

So I built a skill that looks before it answers.

Already have skills?

skill-compass lists them, picks the right one for the task and flags broken installs.

Missing one?

skill-compass tells you, suggests what to search for and looks it up on skills.sh with find-skills.

Same model. Same question.

Claude Opus 5, with and without the skill. Shortened and translated from Portuguese, nothing added.

> which skills do I have installed? I forgot what half of them do
Without the skill
You have 10 skills in ~/.claude/skills:
backend-development | General backend reference: APIs,
  databases, auth, security, caching, microservices...
...
Notes:
- grill-me and grilling do the same thing. You can delete
  that folder and lose nothing.
- This project has no skills of its own in .claude/skills/.
  • Missed find-skills, installed in another agent's folder
  • Missed 9 broken skill links, said the project had none
  • Suggested deleting a shortcut that is working as intended
With skill-compass
Engineering
✓ diagnosing-bugs  diagnosis loop for hard bugs
✓ grilling  stress-tests a plan with hard questions
  ↳ grill-me  manual shortcut that runs grilling
Database (this project only)
• prisma-cli  Prisma CLI commands reference
...
✓ = this agent can load it   • = installed for another agent

⚠ 18 issues, mostly Prisma links saved as text files.
  Ask me for a health check to see how to fix them.
  • Scanned every agent folder, found all 20 skills
  • Flagged the broken installs without being asked
  • Explained the shortcut instead of calling it a duplicate

Ask it like you would ask a colleague.

Pick the right one

Which of your skills fit the task, in the order you would use them, and what is missing. Then it steps aside and the chosen skill runs exactly as written.

"Which skills for a REST API with JWT?"

See what you have

Every skill across Claude Code, Codex, Cursor, OpenCode and Windsurf, grouped by area.

"What skills do I have?"

Find broken installs

Broken links, invalid SKILL.md files, name collisions, skills installed for another agent.

ERROR symlink-as-text (18)

Spot overlap

Duplicates, partial overlaps and complementary skills. You decide what to remove.

"Do I need all these?"

Know when to search

Nothing fits? It says so, and points you to find-skills with a search term.

"Do I have anything for OAuth?"

What changed in testing.

Seven prompts, each answered with and without the skill, checked against 5 to 11 written criteria.

100% vs 76%

Criteria met with the skill, compared with the same model on its own.

+2.4k tokens

Average extra tokens per answer, after trimming the skill and its script output.

33s vs 37s

Average time per answer. The read-only script replaces folder-by-folder reads.

List installed skills38% → 100%
No skills installed yet57% → 100%
Too many frontend skills80% → 100%
Missing skill for a task83% → 100%
Explain one skill in depth86% → 100%
Health check, broken fixtures91% → 100%
Health check, real setup100% → 100%

Read these numbers as indicative. One run per prompt per configuration, on Claude Opus 5 in Claude Code, graded by me. The model alone already does well when asked directly, as the health checks show. The value is consistency: it scans every agent folder every time, does not invent skills, and does not push you to delete things. Prompts and fixtures are in the repo. The fixture-based evals can be rerun anywhere. The others ran on my own setup.

Install once. Call it by name.

Add the skill

npx skills add SWAGGATH4K1NG/skill-compass --skill skill-compass -g

Make it trigger on its own

Add this to AGENTS.md or CLAUDE.md.

## Skills
For questions about my agent skills, use the skill-compass skill.

Start it explicitly

The most reliable way, in whichever agent you use.

Claude Code/skill-compass
Codex$skill-compass or pick it from /skills
Cursor/skill-compass
Windsurf@skill-compass
OpenCodeMention it: "use the skill-compass skill"

A guide, not another layer in the way.

Recommends, then steps aside

The skill you pick runs exactly as its author wrote it. Nothing is summarised, skipped or shortened.

Read-only

It never installs, removes or edits anything. You get the command and run it yourself.

Treats skill files as data

Instructions hidden in a third-party SKILL.md are flagged, never followed.

No agent lock-in

Follows the open Agent Skills format. No agent-specific paths or commands assumed.

New skills land every week.
Your setup drifts.

Watch the repo to hear about new checks and fixes as they ship.

Follow on GitHub