A one-glance overview of your Claude Code skills.
skillscope scans every SKILL.md on your machine — personal, project, and
plugin — and writes a single self-contained HTML report: what each skill does,
what triggers it, how big it is, and what to watch out for.
It answers the questions you actually have once you have more than a handful of skills installed:
- What do I even have? Every skill, grouped by scope, in one page.
- What fires when? The trigger phrases pulled out of each description.
- What's risky? Oversized bodies, over-broad triggers, and skills whose descriptions overlap enough to fire for the same prompt.
- Am I over budget? An estimate of how much of the skill-listing context budget your descriptions consume.
Claude Code loads every skill's description into context on every message so
it can decide what to trigger. Bodies load only when a skill fires. A pile of
vaguely-described skills doesn't just waste tokens — it causes trigger
collisions, where the wrong skill fires or none does. skillscope makes that
pile legible so you can keep it small and sharp.
No runtime dependencies — just Python 3.9+.
Run it without installing:
git clone https://github.com/Troels-en/skillscope.git
cd skillscope
python3 skillscope.py --openOr install it as a command with pipx:
pipx install git+https://github.com/Troels-en/skillscope.git
skillscope --openpython3 skillscope.py [options]| Option | Default | Description |
|---|---|---|
--project DIR |
current dir | Project to scan for .claude/skills |
--out FILE |
skill-report.html |
Where to write the HTML report |
--json FILE |
— | Also write a machine-readable JSON report |
--context-window N |
200000 |
Model context window in tokens, for the budget estimate. Use 1000000 on a 1M-context model. |
--open |
off | Open the report in your browser when done |
Example:
python3 skillscope.py --project ~/my-app --context-window 1000000 --open- Summary — counts by scope, oversized bodies, broad triggers, collision pairs, and estimated description-budget usage.
- Trigger collisions — pairs of skills with high description overlap, scored, with the shared words listed.
- Per-skill cards — description, extracted trigger phrases, non-triggers, body size, invocation mode (you / Claude / both), and a plain-language note.
Skills are flagged when they:
- have a body over 500 lines (the official size guideline),
- have a description +
when_to_useover 1,536 characters (the listing cap), - use broad wording (
any,every,always,must use, …) that makes them fire on unrelated prompts, - have no description at all.
skillscope only reads SKILL.md files on your machine and writes a local HTML
file. Nothing is sent anywhere. No network calls. No API key. No LLM.
- The budget figure is an estimate (description characters ÷ 4 vs 1% of the
context window), not Claude Code's exact accounting. Run
/doctorin Claude Code for the authoritative budget status. - Plugin skills are read from the install manifest, so only installed plugins are scanned — plugins merely available in a registered marketplace are not counted. If the manifest is missing, no plugin skills are reported.
- Analysis is rule-based — it reads structure, not meaning. A skill it marks "healthy" can still be badly written.
MIT — see LICENSE.