AgentSkillsCN

recall

在开始工作前,从 agents-notes/ 中搜索并调取相关笔记。适用于开启新任务、遇到熟悉的错误,或在先前会话的情境有助于当前工作时使用。

SKILL.md
--- frontmatter
name: recall
description: Search and retrieve relevant notes from agents-notes/ before starting work. Use when beginning a task, encountering a familiar error, or when context from previous sessions would help.
argument-hint: "[search query or topic]"
allowed-tools: Read, Grep, Glob

Procedure

No arguments — show summary

If $ARGUMENTS is empty:

  1. Use Glob to list all files in agents-notes/**/*.md (excluding README.md)

  2. Group by category folder

  3. Present a summary table:

    code
    | Category  | Count | Recent Notes                          |
    | --------- | ----- | ------------------------------------- |
    | errors    | 3     | drizzle-relation-loading.md, ...      |
    | patterns  | 5     | result-type-error-handling.md, ...    |
    | decisions | 1     | why-trpc-over-rest.md                 |
    | tips      | 2     | database-full-reset.md, ...           |
    | context   | 0     | —                                     |
    
  4. Include total count and a reminder: Use /recall <query> to search for specific topics.

With arguments — search and present

If $ARGUMENTS contains a search query:

  1. Search file contents: Use Grep to search agents-notes/**/*.md for the query terms (case-insensitive)

  2. Search filenames: Use Glob to match agents-notes/**/*<query-term>*.md

  3. Combine results: Deduplicate matches from both searches

  4. Read top matches: Read up to 5 most relevant notes

  5. Present concise summaries: For each note, show:

    • File path
    • The ## Summary section content
    • Tags if present
  6. If no matches found, suggest:

    • Broader search terms
    • Checking /recall (no args) to browse all notes
    • The note might not exist yet — consider using /note to capture it

Important

  • Read-only — never modify or create notes during recall
  • Be concise — show summaries, not full note contents (user can read the full file if needed)
  • Prioritize relevance — if many matches, show the most relevant ones first