Setup
One-time setup or rebuild of the optional cluster-routing manifest — groups your installed skills into 15-30 semantic clusters with visible router skills + hidden leaves so the prelude carries one router-description per cluster instead of one description per skill. Mostly useful past ~150 installed skills, when the per-description prelude budget overflows. USE WHEN the user explicitly wants cluster routing set up or rebuilt; skip otherwise — the default management surface is `/skill-tree:provision`.From its SKILL.md
npx -y skills add danielbrodie/skill-tree --skill setupAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 2 stars2 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
SKILL.md
3.0 KB, as published. Nobody here has run it
This command sets up the skill-tree manifest. You are the clustering algorithm.
Step 1: Scan — collect all skill names and descriptions:
${CLAUDE_PLUGIN_ROOT}/bin/skill-tree scan --format json
Read the JSON output. This is every skill on disk with its name and description.
Step 2: Cluster the skills. Read all the descriptions and group them into 15-30 clusters based on semantic similarity — skills that would co-occur in a real user request belong together.
For each cluster, produce:
- name: a clear routing signal (2-3 words). Test: could a model read just this name and know whether to enter this cluster? Bad:
asc-app-app_id. Good:app-store-connect. - description: a USE WHEN decision rule. "USE WHEN: the user asks about X, Y, or Z." Not a member list.
- leaves: each with a
routingHintthat distinguishes it from siblings. "Use this when X, not Y."
Skills that don't fit any cluster stay as standalones. That's fine — not everything clusters.
Step 3: Present the proposed clustering before writing. Output a table to the user with cluster names, member counts, and descriptions. Ask for confirmation. Only after confirmation, write the manifest directly to ~/.claude/skills-library/skill-tree/manifest.json using this schema:
{
"version": "1.0",
"unclusteredBudget": 25,
"clusters": {
"cluster-name": {
"description": "USE WHEN: ...",
"crossReferences": [],
"leaves": {
"skill-name": { "routingHint": "Use this when ..." }
}
}
},
"standalones": ["skill-a", "skill-b"],
"hotPath": [],
"referenceNodes": [],
"deprecated": []
}
If a manifest already exists, read it first and preserve hotPath, referenceNodes, deprecated, and any customInstructions on existing clusters. Only change the clustering. Concretely: load the existing manifest, build a name→old-cluster map, and for each leaf in your new proposal, if its old cluster had a customInstructions field or a routing hint you'd weaken, preserve the old version.
skill-tree sync --dry-run in Step 4 is the real preview gate — it shows exactly what routing files would change before any disk writes happen.
Step 4: Sync — generate cluster routing files:
${CLAUDE_PLUGIN_ROOT}/bin/skill-tree sync --dry-run
Show preview. If confirmed:
${CLAUDE_PLUGIN_ROOT}/bin/skill-tree sync
Step 5: Verify:
${CLAUDE_PLUGIN_ROOT}/bin/skill-tree status
Present the final status with token savings.