learn

Learn — Skill Auto-Extraction

Distill the current task's approach into a reusable skill, or improve an existing one.

When to Suggest

After completing a task, evaluate TWO conditions:

1. Complexity: 5+ tool calls used
2. Reuse potential: The approach applies to other projects/scenarios, not just this one

If both met, suggest: "This approach might be reusable — want to /learn it?" Never auto-execute. Only suggest.

Extraction Flow

1. Recall

Check if this knowledge already exists:

• Search existing skills for similar functionality
• If a memory system (supermemory, MEMORY.md, etc.) is available, search it too

2. Review

Scan the current conversation for the most recently completed task. Identify:

• Core workflow steps
• Key decisions and why they were made
• Non-obvious tricks or gotchas discovered
• Tools/libraries used and how

If /learn "description" was provided, focus on that specific aspect.

3. Dedup

Search ALL existing skills for similar functionality:

ls ~/.claude/skills/
grep -r "<relevant keywords>" ~/.claude/skills/*/SKILL.md

Three outcomes:

• Similar skill exists: Suggest /skill-improve <skill-name> instead. Do NOT create a duplicate.
• Similar learned skill exists: Suggest updating that skill instead of creating new.
• No match: Proceed to create new skill.

4. Assess Scope

Determine where this knowledge belongs:

• Cross-project reusable workflow → Create skill (continue to step 5)
• Project-specific pattern → Suggest saving to local project memory instead, stop here
• Single factual insight → Suggest saving to memory system instead, stop here

5. Generate

Create skill following skill-creator conventions:

File: ~/.claude/skills/<skill-name>/SKILL.md

Naming: use learned- prefix for auto-extracted skills (e.g., learned-pdf-merge).

Structure:

---
name: learned-<name>
description: <what it does and when to trigger — be specific>
---

# <Title>

<Concise steps. Only include what Claude doesn't already know.>