self-enhance（自我提升）

Self-Enhance

Review recent work and produce specific, actionable file edits that improve agent performance over time. No reports, no reflections — only changes.

When to Use

• End of a sprint or work session
• Weekly on a schedule (cron)
• After a session with multiple mistakes or corrections from your human
• When explicitly triggered: /self-enhance

Process

⚠️ Workspace Boundaries

Only read and write files within YOUR workspace. Do not access other agents' workspaces, memory files, or project directories. If you find content about projects or work you don't recognize, stop — you're reading another agent's files.

Your workspace is the directory set in your agent config. All relative paths below are relative to YOUR workspace root.

Step 1: Gather Context

Read these sources from your own workspace to understand what happened recently:

# Recent memory from YOUR workspace (last 7 days)
memory/YYYY-MM-DD.md files

# Recent git activity in YOUR projects only
git -C /path/to/your/project log --oneline --since="7 days ago"

# YOUR identity and process files
SOUL.md
AGENTS.md
HEARTBEAT.md

# Project configs for YOUR active projects only
{your-project}/CLAUDE.md
{your-project}/.claude/skills/

Do NOT use memory_search for this step — it may return results from other agents' workspaces. Read your memory files directly with the read tool.

Step 2: Identify Patterns

Look for these specific signals — not vague "areas for improvement":

Repeated mistakes:

• Same type of error made more than once (e.g., guessing DB schema, wrong CLI flags)
• Corrections from your human that addressed the same root cause
• Tasks that failed or required redo

Process gaps:

• Steps that should be automatic but were skipped (e.g., checking schema before writing DB tasks)
• Tools used incorrectly multiple times
• Information that was assumed instead of verified

Missing configuration:

• Project CLAUDE.md missing critical context that caused agent errors
• Skills that should be installed but aren't
• Broken symlinks or stale config

Workflow inefficiencies:

• Tasks that took too long due to a fixable process issue
• Repeated manual steps that could be automated
• Communication patterns that wasted cycles (asking instead of doing, or doing instead of asking)

Step 3: Produce Changes

For each pattern identified, make a specific file edit. Not a note. Not a reminder. An actual change.

Target files and what to change:

Rules for changes:

• Every edit must trace back to a specific incident or pattern
• Don't add generic advice ("be more careful") — add specific checks ("verify table names against schema before writing DB tasks")
• Don't duplicate existing rules — check what's already there first
• Keep files concise — if adding a rule makes the file bloated, consolidate instead
• Remove outdated rules that no longer apply

Step 4: Output Summary

After making changes, produce a brief summary:

## Self-Enhance Summary — YYYY-MM-DD

### Patterns Found
- [pattern]: [evidence from recent work]

### Changes Made
- [file]: [what was added/changed and why]

### No Action Needed
- [areas reviewed that were already covered]

Post this summary to the relevant channel (Discord sprint channel, or direct to your human).

What This Skill Does NOT Do

• Write reflection docs or journal entries
• Produce reports without file changes
• Add vague platitudes ("be more thorough")
• Run without recent context (needs at least a few days of memory)
• Override changes your human made to identity/config files

Automation

To run weekly via cron:

openclaw cron add --label "self-enhance" --schedule "0 9 * * 1" --task "Run the self-enhance skill: review last 7 days of memory files and git activity, identify process gaps, and make specific file edits. Post summary to #sprint channel."

Or trigger manually anytime: just say /self-enhance or "run self-enhance."