claude-skill-usage AGENTS.md

AGENTS.md

Context for any AI coding agent working in this repo. (Claude Code reads CLAUDE.md, which imports this file.)

What this is

A Claude Code plugin that reports how often each Claude skill is actually used, what it costs in tokens, and what to do about it (Keep / Modernize / Turn off / Prune). It reads the user's local ~/.claude transcripts — nothing is uploaded.

Architecture (hybrid — deterministic core + Claude judgment)

1. Collector — skills/skill-usage/scripts/skill-usage.py. Pure Python, no LLM. Parses transcripts for Skill tool-calls, sizes each skill, lints it against a best-practices checklist, assigns a cheap heuristic baseline verdict. Writes skill-usage.json + dashboard.html.
2. Recommendation pass — driven by skills/skill-usage/SKILL.md. Claude fetches Anthropic's current skill docs + reads the user's own CLAUDE.md, then writes real verdicts to recommendations.json.
3. Dashboard — generated dashboard.html (single file, vanilla JS). Merges the two; renders an action strip + kanban bucket board + sortable table.

Layout

.claude-plugin/{plugin.json,marketplace.json}   manifests (this repo is plugin + marketplace)
skills/skill-usage/
  SKILL.md            the skill + the recommendation-pass workflow
  DESIGN.md           full design spec
  scripts/
    skill-usage.py    collector + dashboard renderer
    selftest.py       dependency-free assert tests for the pure functions

Run / test

python3 skills/skill-usage/scripts/skill-usage.py --open   # build + open dashboard
python3 skills/skill-usage/scripts/selftest.py             # must print OK

Hard rules (don't break these)

• Outputs go to ~/.claude/skill-usage/, never next to the script. A plugin's own dir is ephemeral (wiped on update). Output location is OUT_DIR (env SKILL_USAGE_OUT overrides). Do not revert to writing beside __file__.
• Dashboard JS uses createElement/textContent, never innerHTML. It embeds JSON the user's skill descriptions flow into; innerHTML would be an injection risk and trips security hooks.
• Always load dashboard.html in a real browser and confirm zero console errors before claiming it works. python -c/node --check will NOT catch browser-runtime bugs (e.g. a const top collision with window.top). This is the required verification gate.
• Pure functions (lint_skill, heuristic_baseline, merge_recommendations, rec_aggregates) are covered by selftest.py. Add a case when you change their behavior.
• stdlib only (optional tiktoken if present; otherwise a chars/4 estimate). Don't add dependencies.
• Token counts are estimates — never present them as exact.
• Recommendations are advisory. The tool never disables or deletes anything on its own; archive (not delete) is the default when a user acts.