AGENT SKILL

skill-router

Search a local corpus of ~6,000 curated agent skills (testing, frontend, backend, cloud, data, writing, security, research, media) BEFORE doing any concrete task. Use at the START of any build/fix/test/deploy/configure/refactor/review request, unprompted - call skills_search with the task's goal, read the best-matching playbook with skills_read, and follow it. Default is ROUTE; skip only for pure chat.

License 许可
MIT
最近更新
2026-09-29
Stars
★ 0
Forks
0

01 前言 Front Matter

name
skill-router
description
Search a local corpus of ~6,000 curated agent skills (testing, frontend, backend, cloud, data, writing, security, research, media) BEFORE doing any concrete task. Use at the START of any build/fix/test/deploy/configure/refactor/review request, unprompted - call skills_search with the task's goal, read the best-matching playbook with skills_read, and follow it. Default is ROUTE; skip only for pure chat.

03 SKILL.md 正文

Skill router

You have access to a local corpus of ~6,000 specialized agent skills that is NOT in your catalog. It is vector-indexed and searchable in about a second. A vetted playbook almost always beats improvising from general knowledge - so search BEFORE planning, not after.

When to route

Route by reading the conversation. A lookup is triggered by the SHAPE of the message, not by the user naming a skill:

  • The user describes something they want built, fixed, tested, deployed, configured, refactored, reviewed, or explained with intent to act.
  • The user names a tool, framework, library, cloud, or database.
  • The user asks "how do I / can we / please do".
  • The conversation is mid-task and the next step is a concrete action.

Skip only when the turn is purely conversational: a greeting, thanks, a clarification question, a status report - or you are already following a skill loaded this turn.

Default is ROUTE. If unsure whether a turn counts, route anyway - a miss costs one second. Silence after an actionable message is the failure mode this skill exists to prevent.

How to search

Write the query the way you would explain the task to a colleague - a goal, not a keyword list. Semantic matching does the rest.

Call the skills_search tool with query (and k if you want more than 5 hits). Each result carries path, score (0-1), name, and a one-line description.

If this host has no skills_search tool, run the bundled search CLI once: feed it the JSON query {"query":"<goal in plain words>","k":5} on its standard input (it reads stdin and writes JSON to stdout; adapt the stdin/pipe invocation to whatever shell this host provides — on Linux/macOS that is bash, on Windows it is PowerShell):

node "/home/ryasr/.dsh/profiles/web/node_modules/dsh-awesome-skills/lib/query.js"

The CLI answers with a JSON object whose results are the ranked hits.

Reading the scores

Calibrated on labeled task queries against this corpus: strong-band top hits were relevant nearly every time; mid-band hits are plausible but need a skim; weak-band hits rarely help. Trust the bands:

  • 0.7+ strong match - read it.
  • 0.4-0.7 plausible - read the top 2-3 descriptions, not just the first line; the best playbook for the task is often at rank 2-3.
  • < 0.4 weak - likely no good playbook for this exact task.

When one query is not enough

  • Broad request ("build me a dashboard") → run 2-3 queries for its distinct facets (e.g. "frontend dashboard layout", "charts data visualization", "deploy static site") and take the best hit from each.
  • First results all score < 0.4 → rephrase ONCE with different wording ("auth" → "login session cookie"). A second miss means the corpus likely has nothing; move on without guilt.
  • Multiple hits with the same trailing name from different sources (e.g. several tdd skills, or duplicates across a5c-ai/babysitter/<domain>/ specializations) → they are alternative takes. The full path distinguishes them. Prefer the one whose description fits the user's stack; when they look equivalent, prefer the higher score and read just one.

Decision points, don't drift

  • Task is one clear domain and a hit scores 0.7+ → read exactly one skill and follow it.
  • Task genuinely spans domains (e.g. "add OAuth to my FastAPI app") → read at most two skills (one per domain), name which is primary, and follow the primary for the overall shape.

How to read a hit

Call the skills_read tool with the hit's path (never the display name - paths are unique) to load its SKILL.md. If this host has no skills_read tool, read the file directly with the host's file-read capability at /home/ryasr/.dsh/awesome-skills/skills/<path>/SKILL.md.

A skill's directory holds its complete playbook: reference files, templates, examples, and scripts live beside the SKILL.md. When the SKILL.md says "read tests.md" or points at references/ or scripts/, load those with skills_read (path + file) too - they are part of the skill, not optional decoration. If skills_read is unavailable, list the directory /home/ryasr/.dsh/awesome-skills/skills/<path>/ with the host's file browser to see its contents.

Rules

  • Read before acting: search → read the match → then plan. Quoting a skill's title from memory is not following it.
  • Never paste a whole skill file into the chat - read it, then act on it.
  • Follow the playbook, adapt the details. If a step genuinely does not fit the user's context, say so and adapt rather than silently skipping.
  • If nothing relevant comes back after one rephrase, continue normally. A miss costs one second and is not a failure.

04 同仓库更多 Skills