Learn

Learning Extraction

Action

Execute the /learn command flow, which orchestrates the learning-extractor agent in two phases with user selection in between.

Execution

1. Check prerequisite: Verify $HOME/.claude-evolve/active exists. If not, tell user to run /evolve init first and stop.
2. Check explicit intent: If the user's request contains explicit type and scope (e.g., "save this as a universal rule"), pre-set those values and skip selection steps.
3. Phase 1 - Discover: Spawn claude-evolve:evolve-learning-extractor with action="discover" and the user's trigger text as topic hint. Agent analyzes conversation context and session signals, returns at most 4 candidate learnings ranked by value.
4. Present candidates: Show discovered learnings to user as a numbered list with summary, suggested type, and scope.
5. User selects: Use AskUserQuestion to let user choose which learnings to capture. Single candidate gets a shortcut (yes/change/skip). Multiple candidates use multiSelect with a "Capture all" option.
6. Confirm settings: Use AskUserQuestion to ask "Accept suggested types/scopes or customize each?" If customizing, ask type then scope per learning with the recommended option listed first.
7. Phase 2 - Create: Spawn claude-evolve:evolve-learning-extractor with action="create" and the user's approved selections (id, summary, detail, type, scope, name, consolidates_with per learning).
8. Report result: Show user what was created or consolidated. If Phase 2 fails, display error and suggest checking toolkit permissions.

See ${CLAUDE_PLUGIN_ROOT}/commands/learn.md for full implementation details.

Learning Types (Learnings become Components)

| Found | Becomes | Location | |-------|---------|----------| | Problem-solving pattern | Skill | $HOME/.claude-evolve/toolkits/{name}/skills/{skill-name}/SKILL.md | | Investigation method | Agent | $HOME/.claude-evolve/toolkits/{name}/agents/{name}.md | | Code pattern/approach | Rule | $HOME/.claude-evolve/toolkits/{name}/rules/{name}.md | | Improvement to existing | Consolidated | Merged into existing file |

Key: No separate storage. Learnings become agents/skills/rules directly.

Error Handling

| Error | Action | |-------|--------| | No toolkit initialized | Tell user: "Run /evolve init first." | | Phase 1 returns no candidates | Report: "No significant learnings found in this session." | | User selects no candidates | Report: "No learnings captured." | | Phase 2 fails to write | Report error, suggest checking toolkit directory permissions | | Consolidation target found | Agent marks target in Phase 1; user approves in selection; agent consolidates in Phase 2 |

Expected Output

## Learnings Captured

**Created:** `agents/{name}.md` (universal)
**Created:** `skills/{name}/SKILL.md` (project-specific)
**Consolidated:** `rules/{name}.md` (universal)

{Description of each component}

Available next session (agents/skills load at startup).
Run `/evolve release` to sync to other machines.

When to Use

Best results when the session contains:

• A root cause diagnosis after debugging
• A correction from the user about approach or tooling
• A reusable pattern or technique discovered during work
• A non-obvious workaround for a tricky problem

Less useful for routine task completions with no novel insights.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: hknc
• Source: hknc/claude-evolve
• License: MIT

Install and usage instructions live in the source repository linked above.