AgentSkillsCN

learning-capture

此技能会将阶段性学习成果加以呈现。可响应：“回顾学习心得”、“看看都学到了什么”、“展示学习收获”、“列出学习要点”，或直接使用“/learn-review”指令。

SKILL.md
--- frontmatter
name: learning-capture
description: >
  This skill captures learnings to staging. Triggers on: "that worked", "it's fixed",
  "working now", "problem solved", "save this", "remember this", /learn command,
  or when user corrects Claude ("use X instead of Y", "no, do X", "always X",
  "never Y", "I prefer X"). Also trigger after solving non-obvious problems together.

Learning Capture

Capture learnings to the staging area in CLAUDE.local.md.

Keep it simple: Just capture the insight. Don't over-structure. /learn-promote will figure out where it belongs.

Two-Phase Approach

  1. •Capture (this skill): Quick, lightweight staging
  2. •Promote (/learn-promote): Smart routing to CLAUDE.md or skills

When to Capture

Capture when you notice:

  • •User corrects you ("use X instead of Y", "no, do X")
  • •A non-obvious solution is discovered together
  • •User states a preference ("I prefer X", "always do Y")
  • •User explicitly asks to save something
  • •Trigger phrases: "that worked", "it's fixed", "problem solved"

Don't capture:

  • •Trivial or one-time instructions
  • •Things already documented
  • •Vague statements without actionable content

Capture Protocol

Step 1: Check for Duplicates

Search CLAUDE.local.md for similar content:

bash
grep -i "keyword" CLAUDE.local.md

If similar exists, don't duplicate. Offer to update if the new learning adds context.

Step 2: Identify the Core Insight

Ask yourself:

  • •What's the actionable takeaway?
  • •Is it reusable for future sessions?
  • •Can it be stated simply?

Step 3: Propose Briefly

Keep proposals minimal - don't interrupt the flow:

code
Got it - want me to save "use uv instead of pip" as a learning?

Or for discoveries:

code
That was a non-obvious fix. Save to learnings?
"Prisma pool errors in serverless → use connection pooling"

Step 4: Save on Approval

Add to CLAUDE.local.md (gitignored):

markdown
# Staged Learnings

- Use uv instead of pip for Python package management
- Always run tests before committing
- ralph loop: check progress.md before starting

Format: Just a simple bullet point. Natural language. No complex tags.

Step 5: Confirm

code
Saved to staged learnings. Run /learn-promote when ready to route.

Quality Filters

Skip if:

  • •Too vague ("do it differently")
  • •One-time ("just this once")
  • •Already exists
  • •Not actionable

Capture if:

  • •Would help future sessions
  • •Required discovery to learn
  • •User explicitly corrected behavior

File Location

All staged learnings go to CLAUDE.local.md (automatically gitignored).

For promotion to final destinations, use /learn-promote.