continuous-learning
SKILL.md
Learn — Extract & Save Reusable Patterns
Scan current session for reusable knowledge, review with user, save as skill files.
Workflow
- Scan current conversation for extractable patterns
- Present each candidate via AskUserQuestion
- Save confirmed patterns as skill files
- Report summary
What to Extract
| Type | Signal |
|---|---|
| Error resolution | Error + root cause + fix |
| Debugging technique | Non-obvious debug steps |
| Workaround | Library quirks, API limitations |
| Project convention | Codebase-specific patterns |
| User preference | Corrections: 不要/别/必须/always/never/don't |
| Dead end | Approach tried → failed → why it doesn't work |
| Blocker | Limitation discovered: API/tool/library can't do X |
| Anti-pattern | Method that seems right but causes subtle issues |
| Agent introspection | Agent failed → why? Prompt gap / tool limitation / knowledge blind spot |
Skip: Simple typos, one-time fixes, obvious patterns, anything non-reusable.
Agent Introspection
When the agent makes a mistake or fails at a task, go deeper than "what failed":
- What went wrong? — The observable failure
- Why? — Root cause: was it a prompt issue, tool limitation, missing context, or knowledge gap?
- What to update? — CLAUDE.md rule, skill instruction, or tool configuration to prevent recurrence
Quality Gate
Extract only if all four hold: non-trivial, reusable across sessions, specific (clear problem + solution), actionable.
Review UX
For each extracted pattern:
question: "Pattern: {name}\n{summary}\n\nSave?"
options:
- label: "Save to project (Recommended)"
description: "Write to .claude/skills/learned/ (current project only)"
- label: "Save globally"
description: "Write to ~/.claude/skills/learned/ (all projects)"
- label: "Skip"
description: "Discard"
- label: "Edit then save"
description: "Modify before saving"
Save Locations
{project}/.claude/skills/learned/ # Project-level (higher priority)
~/.claude/skills/learned/ # Global
Project skills override global skills with the same name.
Output Format
See references/format.md for the saved skill file structure.
Agent-Agnostic Usage
This skill works with any AI coding agent:
- Claude Code:
/learn - Codex CLI / others: Point the agent to this file and say "follow the workflow above"
Weekly Installs
1
Repository
fancive/claude-skillsFirst Seen
11 days ago
Security Audits
Installed on
codex1