-
Notifications
You must be signed in to change notification settings - Fork 1
Skill Recommendation
skm ask and skm recommend answer the question that matters during real work: which installed skill should I use right now?
skm ask "convert a web page to Markdown"
skm recommend "create image cards for Xiaohongshu" --top 5
skm recommend "markdown to html" --why
skm recommend "create a knowledge graph" --advisor codex --whyUse ask when you want a short answer. Use recommend when you want to compare options.
By default, ranking is fully local. It uses:
| Signal | Why it matters |
|---|---|
| Skill name and frontmatter | Direct intent match |
| Category and description | Broader semantic match |
| Chinese and English synonyms | Users can ask naturally in either language |
| Task intent | Images, graphs, slides, meetings, writing, translation, conversion |
| Conversion direction |
markdown to html should not pick the reverse tool |
| Usage count and recency | Recently useful tools get a small lift |
| Both-side availability | Skills installed in both Claude Code and Codex are easier to reach |
Advisor mode is opt-in:
skm recommend "summarize meeting notes" --advisor claude
skm recommend "generate knowledge graph" --advisor codex --whyIt sends only a compact candidate list: name, category, tool source, description, usage count, local score, and local reasons.
It does not send real skill paths, config paths, MCP env values, API keys, passwords, private keys, or session log bodies.
If the local CLI is missing, logged out, offline, or times out, skm falls back to local ranking.
A good recommendation should tell you three things:
- The best skill to try first.
- Why it matched your task.
- What to try next if the first choice is not right.
For detailed score evidence, add --why.
Built for local Claude Code / Codex skill governance. Repository: skill-manager.