OpenClaw Skill Creator

A meta-skill that gives your OpenClaw agent the ability to autonomously create, improve, retire, and optimize its own skills — with a built-in learning loop and proactive skill detection that makes the agent smarter over time.

What it does

Capability	Description
Create skills	Captures workflows into reusable SKILL.md files from user requests, conversations, or detected patterns
Proactive detection	Tracks repeated workflows as candidates and suggests new skills when patterns emerge (3+ occurrences)
Improve skills	Tracks corrections and quality, then iteratively improves underperforming skills
Retire skills	Flags stale/unused skills and archives them cleanly
Audit skills	Runs periodic health checks on the entire skill library, including candidate review
Optimize triggers	Tunes skill descriptions so they activate at the right time

The Learning Loop

CREATE → USE → OBSERVE → EVALUATE → IMPROVE → repeat

Every skill is tracked in a skill ledger (skill-ledger.json) that records usage count, correction rate, last used date, and notes. The agent uses this data to:

• Flag skills with >30% correction rate for improvement
• Suggest retirement for skills unused in 60+ days with <3 total uses
• Detect repeated manual workflows that should become skills
• Continuously refine skill instructions based on real feedback

Install

Quick install

git clone https://github.com/ajdavis21/openclaw-skill-creator.git
cd openclaw-skill-creator
./install.sh

The installer will:

1. Detect your OpenClaw workspace (from ~/.openclaw/openclaw.json or OPENCLAW_WORKSPACE env var)
2. Copy the skill into workspace/skills/skill-creator/
3. Install the audit circuit into workspace/CIRCUITS/
4. Register the skill in your SKILLS.md
5. Scan existing skills into the ledger for tracking

Manual install

If you prefer to do it yourself:

# Copy into your skills directory
cp -r . ~/.openclaw/workspace/skills/skill-creator/

# Copy the audit circuit
cp circuits/audit-skills.md ~/.openclaw/workspace/CIRCUITS/

# Create the retired skills archive
mkdir -p ~/.openclaw/workspace/skills/.retired/

# Add to your SKILLS.md table (adjust path as needed)
echo "| skill-creator | v1.1.0 — Create, improve, retire, and optimize skills with learning loop + proactive detection | workspace/skills/ | ✅ Active |"

Custom workspace path

If your workspace isn't at the default location:

OPENCLAW_WORKSPACE=/path/to/your/workspace ./install.sh

Usage

Once installed, your agent responds to natural language:

Say this	Agent does
"Make a skill for weekly email summaries"	Interviews you, drafts SKILL.md, tests it, registers it
"Turn what we just did into a skill"	Captures the conversation workflow as a reusable skill
"The invoice skill keeps getting the date format wrong"	Logs correction, identifies root cause, improves the skill
"Run a skill audit"	Reviews all skills, flags stale/struggling ones, suggests new ones
"Retire the old reporting skill"	Archives it to .retired/, removes from index
"Why didn't the email skill trigger?"	Analyzes the description, optimizes trigger phrasing

File structure

skill-creator/
├── SKILL.md                        # Main skill instructions
├── skill-ledger.json               # Usage & quality tracking
├── install.sh                      # Installer script
├── references/
│   └── skill-template.md           # Copy-paste template for new skills
└── circuits/
    └── audit-skills.md             # Periodic audit circuit (heartbeat-driven)

Skill ledger schema

{
  "version": 1,
  "skills": {
    "skill-name": {
      "created": "2026-03-16",
      "lastUsed": "2026-03-16",
      "lastUpdated": "2026-03-16",
      "useCount": 12,
      "corrections": 2,
      "successRate": 0.83,
      "notes": ["Fixed date format issue in v1.1"],
      "status": "active"
    }
  },
  "candidates": [
    {
      "pattern": "Weekly cross-system status rollup",
      "firstSeen": "2026-03-18",
      "occurrences": 3,
      "triggerPhrases": ["weekly rundown", "catch me up"],
      "toolsUsed": ["gcal_list_events", "asana_search_tasks"],
      "status": "pending",
      "notes": "User asks some version of this every Monday"
    }
  ],
  "audit": {
    "lastAudit": "2026-03-16",
    "auditIntervalDays": 14
  }
}

SKILL.md format

Skills created by this tool follow the OpenClaw SKILL.md convention:

---
name: skill-name
version: 1.0.0
description: >
  What it does and when to trigger.
metadata:
  openclaw:
    category: ops|comms|data|dev|meta|finance|content
    requires:
      tools: [required, tools]
---

See references/skill-template.md for a full template.

Configuration

Audit frequency

Edit skill-ledger.json:

{
  "audit": {
    "auditIntervalDays": 14
  }
}

Retirement thresholds

Defaults (configured in SKILL.md — edit to taste):

• Stale: >60 days unused AND <3 total uses
• Struggling: >30% correction rate

Circuits

The audit-skills circuit integrates with OpenClaw's heartbeat system. If your agent uses circuits (workspace/CIRCUITS/), the audit runs automatically. If not, just ask your agent to "run a skill audit" anytime.

How it compares to Anthropic's skill-creator

	Anthropic	OpenClaw
Eval model	Heavy — JSON assertions, grader subagents, benchmark viewers	Lightweight — real-world usage tracking + user corrections
Lifecycle	Create and ship	Continuous loop: create → use → observe → evaluate → improve
Retirement	None	Auto-flags stale skills, archives to .retired/
Usage tracking	None	Ledger with use count, correction rate, success rate
Self-improvement	No	Yes — skill-creator tracks and improves itself
Environment	Claude Code / Claude.ai	OpenClaw agents (any deployment)

Requirements

• OpenClaw agent with a workspace directory
• Agent tools: read, write, edit, exec (minimum)
• sessions tool (optional — enables parallel skill testing with subagents)
• Python 3 (for the install script's existing-skill scanner)

License

MIT