auto-skills（自动技能）

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:

1. Project-local skills (current project)
2. User-global skills (user directory)
3. 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:

1. High-recall candidate generation (wide net).
2. 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:

1. Match Confirmation

- auto-skills matched: <short reason>

1. Top Recommendations (max 3)

- Skill name - Why it matches (1 line) - Suggested scope (session/project/user)

1. 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:

1. Run priority lookup.
2. Produce top 3 precise recommendations.
3. Ask for selection.
4. If supported, ask install scope.
5. 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.