Skill catalog audit — installed vs used
reference· tech-lead · updated 2026-06-28 · source
The real session-start token cost is the always-on skill catalog (every skill's name + description is loaded each session), not agent frontmatter. This audit inventories what is installed against what actually fired in the usage log, so the dead weight can be pruned (#1031).
Source of truth for usage: .git/kit-usage.jsonl (hook .claude/hooks/skill-usage-log.sh).
Window audited: 154 events (109 agent, 45 skill). "Fired" = appeared at least once in that window;
"wired" = referenced by an agent config in .claude/agents/ or kit.config.json.
Installed plugins (5)
| Plugin | Provides | Verdict |
|---|---|---|
claude-kit@claude-kit-marketplace | the kit itself (task/effort/gc/autopilot) | Keep — core |
agent-skills@arvindand-skills | context7, github-navigator, ui-ux-design, skill-crafting, maven-tools | Keep — context7 is wired to tech-lead/backend/devops (maven-tools = Java, unused) |
chrome-devtools-mcp@claude-plugins-official | chrome-devtools skills + MCP | Keep — qa agent |
paper-desktop@paper | code-to-design, design-to-code (Paper MCP) | Keep — designer agent |
rust-analyzer-lsp@claude-plugins-official | Rust LSP | Remove — tedos is a TS/JS monorepo; no Rust |
Standalone skills — ~/.claude/skills/ (26)
| Skill(s) | Wired | Fired | Verdict |
|---|---|---|---|
gsap-core · gsap-timeline · gsap-performance · gsap-scrolltrigger · gsap-plugins | ✅ (frontend/designer) | — | Keep — animation work |
playwright-best-practices | ✅ (qa) | — | Keep |
opentui | ✅ (kit-ui) | ✅ ×2 | Keep |
brand-discovery | ✅ (designer) | — | Keep |
lavish | — | ✅ ×2 | Keep — used for review artifacts |
turborepo | — | ✅ ×1 | Keep — monorepo build |
tailwind-design-system · design-rules · design-system-process | — | — | José's call — design work, likely keep |
deploy-to-vercel | — | — | José's call — deploys are Vercel |
codemod | — | — | José's call — situational refactors |
n8n-code-javascript · n8n-code-python · n8n-expression-syntax · n8n-mcp-tools-expert · n8n-node-configuration · n8n-validation-expert · n8n-workflow-patterns (7) | — | — | José's call — there is an n8n agent, but no n8n work in the window |
web2book · gws-slides · find-skills · colorize | — | — | Remove — low relevance, neither wired nor fired |
Recommendation
- Safe to remove now (1 plugin + 4 skills):
rust-analyzer-lspplugin;web2book,gws-slides,find-skills,colorizeskills. - Decide (situational, ~11 skills): the
n8n-*set (keep only if automation is on the roadmap),tailwind-design-system/design-rules/design-system-process(design — likely keep),deploy-to-vercel,codemod. - Keep: everything wired to an agent or that fired in the window.
The uninstall itself (#1033) touches the global ~/.claude environment, so it runs only on an
explicit prune decision — see kit-resource-inventory.md for the per-command cost model.
Pruned 2026-06-28 (#1033 — conservative)
José chose the conservative prune. Executed against the global ~/.claude:
- Plugin uninstalled:
rust-analyzer-lsp@claude-plugins-official—claude plugin uninstall rust-analyzer-lsp@claude-plugins-official -y. Restore:claude plugin install rust-analyzer-lsp@claude-plugins-official. - Skills moved out of the catalog:
web2book,gws-slides,find-skills,colorize→~/.claude/skills-removed-2026-06-28/(reversible —mvthem back to~/.claude/skills/). Standalone catalog: 26 → 22 skills.
Deferred (José's call, kept for now): the n8n-* set, tailwind-design-system / design-rules /
design-system-process, deploy-to-vercel, codemod.