Install the skill. Customize it to your stack. Decide once, never re-decide.
The intelligence layer for Claude skills — surfaces what fits your project, decomposes external skills into project-tailored forks, and persists every judgment.
32 seconds: what it does, and why. — 9:16 reels · 4:5 feed
Status: Stable · v4.6.2 · 55 supported platforms · Three editions: Lite (default) · Python · claude.ai (web + desktop)
Claude Code (plugin, recommended):
/plugin marketplace add captkernel/plugins
/plugin install skills-curator@captkernel
Other ways: the one-line installer below, a manual copy into ~/.claude/skills/, or the claude.ai zip.
--customizeforks a skill into your stack. It splits an external SKILL.md into sections, scores each one against your project, and rewrites mismatched examples in your stack. The capability transfers; the author's accent doesn't.- Token cost is measured, not estimated. 249 tokens always-on per session and at most 1,140 tokens per command, flat from an empty folder to a 17,680-file project (report).
- Every decision persists. Each evaluation (pros, cons, conflicts, verdict) is saved to a local registry, so re-asking about a skill costs about 13 tokens instead of about 4,060.
# One-liner (installs Lite by default, adds Python tier if Python 3.10+ is found):
npx skills add captkernel/Skills_Curator
# Or clone + run an installer (copies into ~/.claude/skills/ and ~/.claude/commands/):
git clone https://github.com/captkernel/Skills_Curator
cd Skills_Curator
bash install.sh # auto; or --lite-only / --with-python
powershell -ExecutionPolicy Bypass -File install.ps1 # Windows (same modes via -Tier)Manual copy: put skills/skills-curator-lite/ (and, for the Python tier, skills/skills-curator/) in ~/.claude/skills/, and .claude/commands/*.md in ~/.claude/commands/.
claude.ai (web + desktop): download skills-curator-claudeai.zip (19 KB), then upload it at claude.ai → Settings → Capabilities → Skills → Upload skill. To persist your registry, add an empty skills_registry.json to a Project's Knowledge. Full setup in INSTALL.md.
After install, the skill auto-loads in any new Claude Code session and announces itself once on first activation.
Three editions ship in one repo, targeting two different runtimes. Pick by where you run Claude, not by feature wish-list:
- Claude Code (CLI) → Lite (default) or Python (opt-in performance tier). The plugin installs both;
npx skills addinstalls Lite and adds Python if 3.10+ is found. - claude.ai (web + desktop) → claude.ai edition. Uploaded as a zip via Settings → Capabilities → Skills.
The three editions share the same judgment model and the same registry JSON schema (v3.0) — a registry written by one edition can be read by another. Lite needs no engine; the Python tier adds single-pass speed on 100+ skills, 37 regression tests and Gist sync; the claude.ai edition drops PLATFORMS / MIGRATE and saves forks as uploadable zips. Full side-by-side table: docs/reference.md.
The user asks "what skills fit this project?". The skill activates, reads the project context, and weaves a ranked list with trust-tier icons, pros/cons, and install commands into the response. No prompting, no manual lookup. Sample --recommend output: docs/reference.md.
Existing skill tools (npx skills, asm, vercel-labs/find-skills) wait for you to ask. They solve plumbing — install, list, sync — and rely on you to know what to look for. That's the wrong default, because:
- You don't know what skills exist until you go searching.
- Popularity-ranked recommendations surface the same generic picks for everyone.
- You forget — six months later, you're re-evaluating a skill you already decided about.
- Every skill you bolt on imports its author's voice. Stack five external skills and your project speaks in five accents.
Skills Curator inverts the model. It's context-aware — reads what you're building and tells you what's worth considering. Recommendations refresh only when your project actually changes. And when you adopt a skill, --customize decomposes it first so it arrives in your project's voice, not the author's.
The mental model shift: stop treating external skills as monolithic plugins to invoke. Treat them as capability libraries to absorb selectively. Your project keeps one author. Yourself.
| Other skill managers | Skills Curator | |
|---|---|---|
| Install / list skills | ✅ | ✅ |
| Context-aware — recommendations match your stack, refresh on project change | ❌ | ✅ --auto |
| Listens for symptoms — "slow tests" → recommends test-perf skills | ❌ | ✅ --symptoms (17 patterns) |
| Trust-rated catalog — curated + live GitHub topic search | ✅ 19 curated + live | |
| Pre-install evaluation with pros / cons / conflicts / verdict | ❌ | ✅ Persisted forever |
| Skill decomposition + project-voice rewrite — adopt capability without inheriting voice | ❌ | ✅ --customize |
| Pre-install security scan — 15 risk patterns | ✅ Pre-install, automatic | |
| Stack audit — duplicates, conflicts, unreviewed skills, health scores | ❌ | ✅ One pass |
| Cross-agent migration with verified paths | ✅ 55 platforms | |
| Cross-device sync via private GitHub Gist | ❌ | ✅ |
| Bounded token cost — engine reads project in subprocess, not Claude's context | ❌ | ✅ (report) |
| Honest about install counts (no fake popularity scraping) | ❌ | ✅ |
--auto re-runs recommendations only when your project's key files change, --symptoms maps complaints like "my tests are slow" to skill categories, and --find / --discover searches the catalog by keyword. More in The intelligence layer.
This is the unlock. The standard model — npx skills add <author/skill> — installs an external skill as-is: their examples, their stack assumptions, their voice. Your project then speaks with their accent.
--customize runs differently:
- Decomposes the source SKILL.md into its constituent sections (frontmatter, when-to-use, patterns, examples, anti-patterns, etc.).
- Reads your project's stack — languages, frameworks, deps, goals from CLAUDE.md.
- Scores every section against that fingerprint and tags an action:
keep-emphasize,keep-trim,rewrite-stack,drop-or-rewrite, orrewrite-frontmatter. - Writes a forked SKILL.md at
~/.claude/skills/<name>-for-<project>/SKILL.mdcontaining the plan. - The agent rewrites each section in your project's voice — keep high-fit material verbatim, rewrite stack-mismatched examples in your stack, drop irrelevant parts entirely.
Result: Vue snippets become React snippets. Django routes become FastAPI routes. The capability transfers. The author's accent doesn't.
Run --customize superpowers:test-driven-development against a Python project:
01. ✏️ [rewrite-frontmatter] score=-1 (YAML frontmatter)
02. 🔴 [drop-or-rewrite ] score=0 # Test-Driven Development (TDD)
03. 🟢 [keep-emphasize ] score=7 ## Red-Green-Refactor
Matched : testing, ai
04. 🟡 [keep-trim ] score=2 ## Common Rationalizations
05. 🟢 [keep-emphasize ] score=6 ## Why Order Matters
Matched : testing, documents
High-fit sections (Red-Green-Refactor, Why Order Matters) get kept and emphasized. Narrative-only sections that don't ground in your project get dropped or rewritten. The frontmatter is renamed so the fork is clearly yours-for-this-project. The original author's contribution becomes invisible the way a great translator's does — the meaning transfers, the accent doesn't.
Run it with python registry.py --customize <source> (add --no-fork to preview the plan); <source> is a registered id, a local path, or owner/repo@skill.
249 tokens always-on per session and at most 1,140 tokens of stdout per command; project size has no measurable effect because the engine reads your files in a subprocess. Full table in docs/reference.md and the token cost report.
In any Claude Code session, ask anything below — the skill activates automatically:
- "Should I install agent-browser for this project?" → evaluates against your CLAUDE.md, runs a security scan, gives ADOPT / PARTIAL / SKIP with evidence
- "What skills would help this project?" → scans your stack, ranks recommendations by fit, returns capability + preference splits with pros/cons
- "Audit my skills" → finds duplicates, preference conflicts, security-unreviewed community skills, stale versions, low health scores
- "Migrate my skills to Cursor and Codex" → multi-target copy, with safety checks for existing files
Slash commands: /skill-evaluate, /skill-recommend, /skill-audit. Architecture, the full command reference, platforms, FAQ, requirements and repository layout are in docs/reference.md.
- v4.5 (shipped): the claude.ai edition (web + desktop), with its own catalog, signal tables, security patterns, persistence guide and a
SCANverb. - v4.6 (shipped):
--customizekeeps an_archive/of every source section, and--restore <fork-id>patches back sections that fit the project again. - Next:
--validate-skill <path>for skill authors, upstream drift tracking for forks, and a user config for_TRUSTED_AUTHORS. - v5.0 — Schema v4: snapshot the catalog at evaluation time so verdicts remain auditable even as the catalog moves on.
Stable. v4.6.2 current; schema v3 (stable since v4.0). Lite (default, zero deps) + Python full (opt-in). CI runs 3 OS × 4 Python versions on every PR, with 37 pytest cases (Python tier). Catalog: 19 curated entries + live GitHub topic discovery (24h cache). 55 platforms, mirrored from vercel-labs/skills. MIT. Maintainer: @captkernel.
- Runs. Lite uses only the agent's Bash, Read, Glob, Grep and Write tools, plus
curlto the GitHub search API (and toraw.githubusercontent.comwhen you customize a GitHub-hosted skill). The Python tier runs one stdlib-only script,skills/skills-curator/scripts/registry.py, as a subprocess. - Reads. Your project's files (stack, dependencies, CLAUDE.md) and the skills installed under
~/.claude/skills. - Writes. Its own state files (registry.json plus caches such as catalog.json) under
~/.claude/skills/skills-curator*/on your machine, and Lite briefly writes a temp file (/tmp/source-skill.md) while customizing; a fork folder there when you run--customize; copies into other agents' skill folders when you run--migrate. - Sends. Without Gist sync, only GET requests to GitHub: topic search on
api.github.com, SKILL.md fetches fromraw.githubusercontent.comwhen you customize a remote skill, and release lookups when you run--stale. Your code, project content and registry stay on your machine. - Gist sync is opt-in. The Python tier pushes the registry to a private GitHub Gist only after you set
SKILLS_CURATOR_GIST_IDandSKILLS_CURATOR_GITHUB_TOKENand run--pushor--sync. Setup:docs/gist-sync.md.
registry.py is stdlib-only Python. Optional network calls go to GitHub only (api.github.com for Gist sync, version checks and catalog topic search; raw.githubusercontent.com for SKILL.md fetches). It never sends your code, project content, or registry contents to any external service. SKILLS_NO_TELEMETRY=1 turns off catalog search and SKILL.md fetches; --stale and Gist sync still connect when you run them.
The --check scanner is conservative — it flags patterns that could be dangerous even when benign (like eval() in a registry script). That's intentional. Review flagged files and decide; don't miss something.
See SECURITY.md for vulnerability reporting.
PRs welcome. High-value contributions: new SECURITY_RISK_PATTERNS (with justification + a test in tests/test_security_scan.py), high-trust KNOWN_SKILLS catalog entries with source link and pros/cons, new FRAMEWORK_SIGNALS / GOAL_SIGNALS for the project scanner, new _TRUSTED_AUTHORS entries, and verified PLATFORMS corrections with a doc URL.
See CONTRIBUTING.md. All PRs must pass pytest tests/ and python registry.py --validate --strict.
MIT.