Auto Skills
Goal
Improve skill hit rate and recommendation quality by:
- Following strict lookup priority.
- Returning only top 3 precise skills.
- Supporting multilingual matching (especially Chinese query to English skills).
- Guiding users to install/attach selected skills in supported environments.
Hard gate:
- For any action-oriented prompt, run skill routing first before execution.
Default operating mode:
- Try this skill first for nearly every user request with actionable intent.
- Treat "no explicit skill request" as still eligible for skill routing.
- Optimize for high trigger rate first, then precision reranking.
Mandatory Lookup Priority
Always run lookup in this order:
- Project-local skills (current project)
- User-global skills (user directory)
find-skills discovery (only when needed to fill gaps or improve precision)
Do not invert this order.
Trigger Policy (High Recall)
Trigger this skill for:
- Any request with action intent: build, create, implement, refactor, fix, debug,
review, design, plan, optimize, migrate, test, deploy, automate, document,
commit, release, changelog, versioning, publish, push.
- Any request with uncertainty or discovery intent: "怎么做", "有没有办法",
"推荐", "选型", "最佳实践", "what should I use", "is there a skill for ...".
- Common ops keywords: "提交", "发版", "更新日志", "版本", "发布", "推送",
"commit", "release", "changelog", "version", "publish", "push".
- Any domain-specific request even without the word "skill".
- Mixed-language or typo-heavy prompts.
Only do not trigger when:
- User explicitly says "不要查 skill / skip skills".
Core Behavior
1) Query understanding and normalization
Before matching:
- Detect language of user query.
- Normalize query text:
- Lowercase, remove punctuation noise.
- Expand abbreviations (for example: "ws" -> "websocket").
- Split intent words and domain words.
- Build multilingual expansion terms:
- Chinese <-> English concept mapping first.
- Add synonyms and related action verbs.
Examples:
- "流程图编辑器" -> "flowchart", "graph editor", "diagram"
- "连线规则" -> "edge connection", "validateConnection", "routing"
- "技能推荐" -> "skill discovery", "capability matching"
2) Precision-first matching
Score skills by weighted signals:
- Intent match (what user wants to do)
- Domain match (topic/tech stack)
- Output/task match (what kind of deliverable user expects)
- Trigger-description overlap
Use two-stage ranking:
- High-recall candidate generation (wide net).
- Precision reranking (return strongest top 3 only).
2.1) Cost-aware ranking (prefer free first)
Apply this ranking policy by default:
- Prefer non-paid/free skills when relevance is comparable.
- Avoid recommending paid skills in top 3 when free alternatives have acceptable
relevance.
- Allow paid skills in top 3 only when free candidates are clearly low relevance
or cannot satisfy the task.
Practical rule:
- If a free candidate is within a close relevance band of a paid candidate,
rank the free candidate higher.
- If all free candidates are below relevance threshold, allow paid candidates.
3) Top 3 recommendation only
Return exactly up to 3 skills:
- If >= 3 high-confidence matches: return top 3.
- If only 1-2 high-confidence matches: return only those.
- Do not pad with weak candidates.
4) Installation/attachment prompt (tool-dependent)
If environment supports adding skills (for example Cursor):
- Ask user whether to add selected skill to:
- This session
- Current project
- User directory (global)
Ask before performing placement changes.
5) Route first, then delegate specialized skills
When a specialized skill exists (for example commit/release/domain-specific):
- Do not bypass routing.
- Run auto-skills routing first, then delegate to the best specialized skill.
- Keep top-3 recommendation behavior intact before delegation.
6) Runtime initialization awareness (project rule recommendation)
When this skill is matched, if the user has not made a decision yet about adding
project rules (accepted or declined), proactively recommend adding a project
fallback rule so routing behavior is consistently applied in that project.
Recommended action:
- Suggest creating
.cursor/rules/auto-skills-first.mdc.
- Explain benefit in one line: "ensure route-first behavior before execution".
- Ask for confirmation before creating/modifying project rule files.
Suggested rule snippet:
---
description: Route action requests through auto-skills first
alwaysApply: true
---
# Auto-Skills First
For any action-oriented user prompt, run `auto-skills` routing before direct task execution.
Mandatory routing order:
1. Project skills
2. User-global skills
3. `find-skills` discovery
Skip condition:
- Only skip routing when the user explicitly says to bypass skill lookup.
Response Template
Use this concise structure:
Match Confirmation
auto-skills matched: <short reason>
Top Recommendations (max 3)
- Skill name
- Why it matches (1 line)
- Suggested scope (session/project/user)
Optional next action
- Ask user to choose 1/2/3 (or none)
- If supported: ask where to add it
- If rule decision is unknown: recommend adding project fallback rule
Suggested Interaction Pattern
When user asks for skill help:
- Run priority lookup.
- Produce top 3 precise recommendations.
- Ask for selection.
- If supported, ask install scope.
- Confirm applied result.
Multilingual Match Strategy
Use a small internal strategy for robust multilingual hit rate:
- Intent dictionary: actions like build/fix/review/plan/search.
- Domain dictionary: framework/library/platform terms.
- Cross-language aliases:
- Chinese -> English primary mapping.
- English acronym -> full phrase.
- Fuzzy tolerance:
- Handle typos and mixed-language prompts.
Never require users to use exact skill names.
Safety and Quality Rules
- Never recommend more than top 3 in one response.
- Prefer existing installed skills before discovery.
- Prefer free skills over paid skills unless relevance is insufficient.
- Avoid generic recommendations when a specialized skill exists.
- Explain recommendation reasons briefly and concretely.
- If uncertain, ask one focused follow-up question instead of guessing.
Extra Ideas (Built-in Enhancements)
A) Confidence threshold gate
If all candidates are low confidence, ask a single clarifying question and rerank.
B) Feedback memory
Track user accepted/rejected skills in-session to improve future ranking.
C) Diversity control
Avoid returning three near-duplicate skills; keep recommendations complementary.
D) Fast fallback
If no suitable skill is found, provide:
- Best baseline skill (if any), and
- A short suggestion to install a new specialized skill via
find-skills.
1---2name: auto-skills3description: Universal skill router for almost all user requests. Aggressively trigger this skill by default for ANY task-oriented prompt (build, fix, explain, plan, review, debug, optimize, migrate, write docs, create scripts, automate workflows, or "how to do X"), including multilingual and mixed-language input. Always attempt skill retrieval first using order project skills -> user skills -> find-skills, then recommend only the top 3 most precise candidates. Only skip when the user explicitly asks to bypass skill lookup.4metadata:5 source: https://skillmd.com/skills/dangaogit/auto-skills6---78# Auto Skills910## Goal1112Improve skill hit rate and recommendation quality by:1314- Following strict lookup priority.15- Returning only top 3 precise skills.16- Supporting multilingual matching (especially Chinese query to English skills).17- Guiding users to install/attach selected skills in supported environments.1819Hard gate:2021- For any action-oriented prompt, run skill routing first before execution.2223Default operating mode:2425- Try this skill first for nearly every user request with actionable intent.26- Treat "no explicit skill request" as still eligible for skill routing.27- Optimize for high trigger rate first, then precision reranking.2829## Mandatory Lookup Priority3031Always run lookup in this order:32331. Project-local skills (current project)342. User-global skills (user directory)353. `find-skills` discovery (only when needed to fill gaps or improve precision)3637Do not invert this order.3839## Trigger Policy (High Recall)4041Trigger this skill for:4243- Any request with action intent: build, create, implement, refactor, fix, debug,44 review, design, plan, optimize, migrate, test, deploy, automate, document,45 commit, release, changelog, versioning, publish, push.46- Any request with uncertainty or discovery intent: "怎么做", "有没有办法",47 "推荐", "选型", "最佳实践", "what should I use", "is there a skill for ...".48- Common ops keywords: "提交", "发版", "更新日志", "版本", "发布", "推送",49 "commit", "release", "changelog", "version", "publish", "push".50- Any domain-specific request even without the word "skill".51- Mixed-language or typo-heavy prompts.5253Only do not trigger when:5455- User explicitly says "不要查 skill / skip skills".5657## Core Behavior5859### 1) Query understanding and normalization6061Before matching:6263- Detect language of user query.64- Normalize query text:65 - Lowercase, remove punctuation noise.66 - Expand abbreviations (for example: "ws" -> "websocket").67 - Split intent words and domain words.68- Build multilingual expansion terms:69 - Chinese <-> English concept mapping first.70 - Add synonyms and related action verbs.7172Examples:7374- "流程图编辑器" -> "flowchart", "graph editor", "diagram"75- "连线规则" -> "edge connection", "validateConnection", "routing"76- "技能推荐" -> "skill discovery", "capability matching"7778### 2) Precision-first matching7980Score skills by weighted signals:8182- Intent match (what user wants to do)83- Domain match (topic/tech stack)84- Output/task match (what kind of deliverable user expects)85- Trigger-description overlap8687Use two-stage ranking:88891. High-recall candidate generation (wide net).902. Precision reranking (return strongest top 3 only).9192### 2.1) Cost-aware ranking (prefer free first)9394Apply this ranking policy by default:9596- Prefer non-paid/free skills when relevance is comparable.97- Avoid recommending paid skills in top 3 when free alternatives have acceptable98 relevance.99- Allow paid skills in top 3 only when free candidates are clearly low relevance100 or cannot satisfy the task.101102Practical rule:103104- If a free candidate is within a close relevance band of a paid candidate,105 rank the free candidate higher.106- If all free candidates are below relevance threshold, allow paid candidates.107108### 3) Top 3 recommendation only109110Return exactly up to 3 skills:111112- If >= 3 high-confidence matches: return top 3.113- If only 1-2 high-confidence matches: return only those.114- Do not pad with weak candidates.115116### 4) Installation/attachment prompt (tool-dependent)117118If environment supports adding skills (for example Cursor):119120- Ask user whether to add selected skill to:121 - This session122 - Current project123 - User directory (global)124125Ask before performing placement changes.126127### 5) Route first, then delegate specialized skills128129When a specialized skill exists (for example commit/release/domain-specific):130131- Do not bypass routing.132- Run auto-skills routing first, then delegate to the best specialized skill.133- Keep top-3 recommendation behavior intact before delegation.134135### 6) Runtime initialization awareness (project rule recommendation)136137When this skill is matched, if the user has not made a decision yet about adding138project rules (accepted or declined), proactively recommend adding a project139fallback rule so routing behavior is consistently applied in that project.140141Recommended action:142143- Suggest creating `.cursor/rules/auto-skills-first.mdc`.144- Explain benefit in one line: "ensure route-first behavior before execution".145- Ask for confirmation before creating/modifying project rule files.146147Suggested rule snippet:148149```md150---151description: Route action requests through auto-skills first152alwaysApply: true153---154155# Auto-Skills First156157For any action-oriented user prompt, run `auto-skills` routing before direct task execution.158159Mandatory routing order:1601. Project skills1612. User-global skills1623. `find-skills` discovery163164Skip condition:165- Only skip routing when the user explicitly says to bypass skill lookup.166```167168## Response Template169170Use this concise structure:1711721. `Match Confirmation`173 - `auto-skills matched: <short reason>`1742. `Top Recommendations (max 3)`175 - Skill name176 - Why it matches (1 line)177 - Suggested scope (session/project/user)1783. `Optional next action`179 - Ask user to choose 1/2/3 (or none)180 - If supported: ask where to add it181 - If rule decision is unknown: recommend adding project fallback rule182183## Suggested Interaction Pattern184185When user asks for skill help:1861871. Run priority lookup.1882. Produce top 3 precise recommendations.1893. Ask for selection.1904. If supported, ask install scope.1915. Confirm applied result.192193## Multilingual Match Strategy194195Use a small internal strategy for robust multilingual hit rate:196197- Intent dictionary: actions like build/fix/review/plan/search.198- Domain dictionary: framework/library/platform terms.199- Cross-language aliases:200 - Chinese -> English primary mapping.201 - English acronym -> full phrase.202- Fuzzy tolerance:203 - Handle typos and mixed-language prompts.204205Never require users to use exact skill names.206207## Safety and Quality Rules208209- Never recommend more than top 3 in one response.210- Prefer existing installed skills before discovery.211- Prefer free skills over paid skills unless relevance is insufficient.212- Avoid generic recommendations when a specialized skill exists.213- Explain recommendation reasons briefly and concretely.214- If uncertain, ask one focused follow-up question instead of guessing.215216## Extra Ideas (Built-in Enhancements)217218### A) Confidence threshold gate219220If all candidates are low confidence, ask a single clarifying question and rerank.221222### B) Feedback memory223224Track user accepted/rejected skills in-session to improve future ranking.225226### C) Diversity control227228Avoid returning three near-duplicate skills; keep recommendations complementary.229230### D) Fast fallback231232If no suitable skill is found, provide:233234- Best baseline skill (if any), and235- A short suggestion to install a new specialized skill via `find-skills`.