Evolution Skill for OpenClaw

🧬 Distill reusable patterns from your workspace history into new skills.

中文版 | English

What is this?

Evolution is an OpenClaw skill that scans your workspaces, identifies reusable patterns from your work history, and generates new skill candidates for your review — with full human-in-the-loop control.

Why does it exist?

Over time, you develop workflows, fix bugs, and discover patterns that could be automated. But these insights are lost in chat history. Evolution watches for patterns across all your workspaces and proposes new skills when it finds something worth automating.

Example: You notice that your researcher agent keeps creating documents in the wrong location. Evolution detects this pattern and proposes a smart-archive skill to automatically route documents to the correct workspace.

Requirements

• OpenClaw v2026.4+ (for skill system support)
• Multiple workspaces (main + subagents) — works best with 2+ agents
• No other skills required — fully standalone

Installation

Method 1: GitHub (Recommended)

# Clone directly to your skills directory
git clone https://github.com/EricArcha/EVO-skill.git ~/.openclaw/skills/evolution
openclaw gateway restart

Method 2: Manual

Download or clone this repository to your skills directory:

cp -r evolution ~/.openclaw/skills/
openclaw gateway restart

Note on ClawHub

ClawHub publication pending. Once published, you will be able to install via:

openclaw skills install evolution

Usage

Command	Description
/evolution	Run scan (incremental mode)
/evolution --full	Full rescan (ignores last run)
/evolution --config	Edit configuration
/evolution --setup-cron	Interactive cron setup
/evolution --help	Show help

First Run

On first run, Evolution will ask:

1. "是否扫描过去90天建立基线？" — Scan last 90 days to establish baseline?
2. "是否设置定时运行cron？" — Set up automatic runs?

Both are recommended for best results.

Workflow

Phase 1: Scan
- Scan configured workspaces (memory/, knowledge/, skills/)

Phase 2: Pattern Recognition
- Identify patterns across 4 categories:
  - 工作边界类 (Cross-workspace boundary issues)
  - 经验复用类 (Reusable prompt templates, workflows)
  - 效率提升类 (Repetitive operations that can be automated)
  - 系统健康类 (Memory bloat, stale knowledge)

Phase 3: Quality Filtering
- Apply frequency, reward, stability, boundary, and OpenClaw compliance gates

Phase 4: Generate Report
- Generate distillation report with candidate skills

Phase 5: User Review
- You review and confirm each candidate (Confirm/Defer/Reject/Modify)

Phase 6: Generate Staging
- On confirmation, skill generates staging SKILL.md

Phase 7: Manual Install
- You install manually (strict confirmation each time)

Phase 8: Track Installed Skills
- Track evolution-generated skills for optimization opportunities

Configuration

Edit ~/.openclaw/skills/evolution/config.yaml:

# Workspaces to exclude
exclude_workspaces:
  - workspace-incognito

# Scan time span (days)
scan_span_days: 30

# Baseline retention (days)
baseline_retention_days: 90

# Cron schedule (leave empty to disable)
cron_schedule: ""

# Max candidates per run
max_candidates: 5

# Min frequency threshold
min_frequency: 3

Integration with Weekly Summary

Evolution integrates with weekly summary skills (周汇报) if you have one configured.

If you have weekly summary:

• Evolution status appears automatically in your weekly report

If you don't have weekly summary:

• Works completely standalone
• Or set up a cron: /evolution --setup-cron

Quality Gates

Each skill candidate must pass ALL gates:

Basic Gates (must pass)

• Frequency: Pattern appears ≥3 times
• Reward: Measurable time savings
• Stability: Not a one-time pattern
• Boundary: Doesn't duplicate existing skills

Execution Constraint Gate (must pass)

• Verifiable: Must include a concrete verification method
• Trigger Clarity: Must specify exact trigger conditions
• Actionable: Must define concrete actions
• Feedback Loop: Must define how the skill confirms it executed correctly

OpenClaw Compliance Gate (must pass)

• Architecture Fit: Follows OpenClaw best practices
• Tool Alignment: Can be executed with OpenClaw's tool set
• Context Appropriate: Scope appropriate for a skill

User Response Options

For each candidate, reply:

• 确认 (confirm) → Generate formal SKILL.md in staging/
• 待定 (defer) → Include in next evolution run, don't forget
• 拒绝 (reject) → Record rejection reason (for future filtering)
• 修改 (modify) → Tell me what needs to change

Strictly No Auto-Execution

Evolution NEVER:

• Auto-creates skills without confirmation
• Auto-installs skills
• Modifies system configuration
• Deletes or moves files

Every action requires your explicit confirmation.

File Structure

evolution/
├── SKILL.md              # Main skill instructions
├── evolution-guide.md    # Detailed workflow doc
├── README.md             # This file
├── README_zh.md          # Chinese version
├── LICENSE               # MIT License
├── config.yaml           # User configuration
├── config.yaml.example  # Config template
├── BASELINE.md           # Incremental scan baseline
├── LAST_RUN.md           # Last run timestamp
├── evolution-installed.md # Registry of evolution-generated installed skills
├── staging/              # Candidate skills (before install)
├── reports/              # Historical reports
└── memory/               # Evolution's own memory (runtime generated)

Known Limitations

• Pattern recognition is heuristic-based (not AI-perfect)
• Requires ≥3 occurrences to propose a candidate
• Incremental scan requires first run to establish baseline
• Scan scope excludes logs/, credentials/, sessions/

License

MIT License — see LICENSE

Contributing

Issues and pull requests welcome. Please read the evolution-guide.md for workflow details.