菜单
Use to organize a large skill library into category routers, route new skills, and propose new categories when existing routers are insufficient.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
| name | skill-router |
| description | Use to organize a large skill library into category routers, route new skills, and propose new categories when existing routers are insufficient. |
Use this router when organizing, reclassifying, or validating a growing skill library. It keeps first-pass discovery small by routing from exposed category skills to hidden leaf skills. For third-party installs, configure the target roots explicitly instead of assuming the plugin is installed inside the same repository as the skills.
Do not read every leaf SKILL.md by default. Start from routing/skill-inventory.json,
routing/routing-review.md, routing/routing-index.json, and the exposed category router
SKILL files. Read a leaf body only when the inventory summary is ambiguous or when moving it
would change a category boundary.
The target skill roots come from skill-router.config.json or environment
variables. For user-global skills, use hidden leaf roots such as
~/.claude/skill-leaves and expose category routers in ~/.claude/skills.
For project skills, use .claude/skill-leaves and .claude/skills.
node plugins/skill-router/scripts/skill-router.mjs scan to refresh compact metadata.plugins/skill-router/routing/routing-review.md and existing category routers under plugins/skill-router/skills/.Use When and Avoid rules stay clear.plugins/skill-router/routing/routing-plan.json.node plugins/skill-router/scripts/skill-router.mjs apply only after the plan is coherent.node plugins/skill-router/scripts/skill-router.mjs scan and node plugins/skill-router/scripts/skill-router.mjs validate.Create a new category only when at least one of these is true:
Avoid rules vague.use_when, avoid_when, and leaf routes.Write plans as JSON with categories and routes. skill-router.mjs apply is deterministic;
the agent owns judgment, and the script only applies the approved plan.
{
"categories": [
{"id": "category-id", "display": "Category Name", "short": "Short focus", "description": "When to use first", "routing_prompt": "How to choose a leaf", "avoid": ["When not to use it"]}
],
"routes": [
{"category": "category-id", "skill": "leaf-skill", "reason": "Why this leaf belongs here"}
]
}
Use when organizing a large local skill library into category router skills, refreshing a compact skill inventory, or deciding which leaf skill bodies need to be read before agent classification.
Use first for ambiguous failures, messy project context, surprising results, repeated failures, postmortems, or contested product and architecture framing.
Use first for API boundaries, module ownership, domain modeling, contracts, architecture seams, and recurring design structures.
Use first for bounded code edits, refactors, implementation, simplification, control-flow cleanup, or maintainer clarity before choosing a more specific theory skill.
Use first when the user asks for tests, correctness, high reliability, static reasoning, proof, or safety evidence.
Use first for work shaped by users, tools, procedures, runtime evidence, operational artifacts, interfaces, or organizational workflows.