Skill Evolver

A safe, diff-based skill builder and optimizer for OpenClaw.

NEVER applies changes automatically. Always suggests diffs and waits for explicit user confirmation.

When to Use

Use this skill when the user wants to:

1. Create a new skill — from scratch or based on an existing template
2. Improve an existing skill — analyze errors, feedback, or usage patterns and suggest enhancements
3. Audit/review a skill — check quality, security, completeness of any SKILL.md
4. Migrate or refactor — restructure skills between workspaces, vaults, or formats

Core Principles

Safety First

• NEVER write files without explicit user confirmation.
• ALWAYS present a diff/preview before applying changes.
• NEVER execute shell commands from skill logic.
• Skills can read anything; they can only write what the user explicitly approves.

Transparent Process

Every operation follows this workflow:

ANALYZE → SUGGEST → REVIEW → CONFIRM → APPLY

• Analyze: Read existing files, understand context, identify gaps or issues.
• Suggest: Generate proposed changes as a clear diff or structured preview.
• Review: Explain why each change is recommended (rationale).
• Confirm: Wait for explicit user approval ("ja", "ok", "apply", "bestätigen").
• Apply: Only then write files, create directories, or modify skills.

Skill Anatomy

A valid OpenClaw skill requires:

SKILL.md Structure

# Skill Name

## When to Use
[Clear trigger conditions]

## Core Principles
[Safety rules, boundaries]

## Workflow
[Step-by-step process]

## Tools / Scripts
[Reference to helper scripts]

## Examples
[Concrete usage examples]

## Safety & Boundaries
[What NOT to do]

Workflows

1. Create New Skill (Builder Mode)

Trigger: User says "create skill for X", "build a skill that does Y", "new SKILL.md for Z"

Steps:

1. Gather Requirements

   • What problem does the skill solve?
   • What triggers should activate it?
   • What tools/commands does it need?
   • What are the safety boundaries?

2. Select or Design Template

   • For simple skills: use templates/basic_skill.md
   • For tool-heavy skills: use templates/tool_skill.md
   • For complex workflows: use templates/workflow_skill.md

3. Generate Draft

   • Fill template with gathered requirements
   • Ensure all required sections are present
   • Add concrete examples

4. Present Diff

   • Show the complete proposed SKILL.md
   • Highlight: triggers, safety rules, boundaries
   • Ask for confirmation

5. Apply on Confirm

   • Create directory: \x3Cskill-name>/
   • Write SKILL.md
   • Create subdirectories (scripts/, templates/) if needed
   • Report success and next steps

2. Improve Existing Skill (Evolver Mode)

Trigger: User says "improve skill X", "optimize this skill", "skill Y has errors", "evolve skill Z"

Steps:

1. Analyze Current State

   • Read the existing SKILL.md
   • Check for: completeness, clarity, safety gaps, outdated info
   • If error reports exist, read ERRORS.md or session logs
   • Identify what's working and what's not

2. Identify Improvement Opportunities

   Category	What to Check
   Triggers	Are trigger conditions clear and complete?
   Workflow	Is the step-by-step process logical and safe?
   Safety	Are boundaries explicit? Is auto-write prevented?
   Examples	Are there concrete, realistic usage examples?
   Tools	Are referenced scripts still valid and secure?
   Format	Follows OpenClaw SKILL.md conventions?

3. Generate Diff

   • Present changes as before → after blocks
   • Group by category (triggers, workflow, safety, examples)
   • Explain rationale for each change

4. Review with User

   • Summarize: "I suggest 5 changes: 2 safety improvements, 1 new trigger, 2 clarified examples"
   • Show full diff
   • Ask: "Soll ich diese Änderungen übernehmen?"

5. Apply on Confirm

   • Update SKILL.md (backup old version)
   • Update related scripts/templates if affected
   • Report what changed and why

3. Audit / Review Mode

Trigger: User says "review this skill", "audit skill X", "is this skill safe?"

Steps:

1. Read the complete skill directory
2. Check against quality checklist (see below)
3. Output structured report with:
   • ✅ Strengths
   • ⚠️ Gaps / Risks
   • 💡 Improvement suggestions
4. Do NOT modify anything — this is read-only

Quality Checklist

When auditing or improving a skill, verify:

Completeness

• SKILL.md exists and is non-empty
• "When to Use" section is clear and specific
• At least one concrete usage example
• All referenced scripts/templates exist

Safety

• Explicit "never auto-execute" rule present
• External actions (write, exec) require confirmation
• No hardcoded secrets or credentials
• Boundary conditions documented (what NOT to do)

Clarity

• Trigger conditions are unambiguous
• Workflow steps are numbered or clearly sequenced
• Technical terms are explained or linked
• Examples cover success AND error cases

Maintainability

• Version/tag noted
• Last updated date present
• Changelog if skill evolved
• Dependencies documented

Templates

Use the templates in templates/ directory:

• basic_skill.md — Simple knowledge/pattern skill (no scripts)
• tool_skill.md — Skill wrapping scripts or external tools
• workflow_skill.md — Multi-step process skill with decision trees
• evolver_task.md — Task template for recording improvement opportunities

Scripts

scripts/analyze_skill.py

Reads a SKILL.md and returns structured analysis:

• Word count, section completeness
• Trigger clarity score
• Safety rule presence
• Example count

scripts/generate_diff.py

Compares two SKILL.md versions and outputs a human-readable diff.

scripts/validate_skill.py

Checks a skill directory against the Quality Checklist above.

Safety & Boundaries

NEVER

• Write or modify any file without explicit user confirmation
• Confirm before write: Always ask user before writing files
• Confirm before execute: Always ask user before running commands
• Execute shell commands from skill logic without approval
• Auto-publish skills to ClawHub without review
• Delete existing skills without backup
• Ignore safety warnings during audit

ALWAYS

• Present a clear diff before applying changes
• Explain WHY each change is suggested
• Backup the old version before overwriting
• Respect user rejection — if they say "nein", stop immediately
• Log what was changed in LEARNINGS.md or session notes

WHEN IN DOUBT

• Ask the user before proceeding
• Prefer read-only analysis over automatic modification
• Document uncertainty in the suggestion rationale

Examples

Example 1: Create New Skill

User: "Ich brauch einen Skill für GPU-Transkription mit faster-whisper"

Skill Evolver:

1. Gathers requirements: Audio file path, model size, language
2. Selects tool_skill.md template
3. Generates draft with:
   • Triggers: User mentions "transcribe", "whisper", "audio to text"
   • Workflow: Validate path → Check GPU → Run faster-whisper → Return text
   • Safety: Never auto-delete audio files, validate paths before reading
4. Presents complete SKILL.md
5. Waits for confirmation: "Soll ich diesen Skill als gpu-transcribe anlegen?"
6. On "ja": Creates gpu-transcribe/SKILL.md and reports success

Example 2: Improve Existing Skill

User: "Der paperclip-skill startet manchmal Embedded-Postgres, obwohl externe DB konfiguriert ist"

Skill Evolver:

1. Reads existing paperclip SKILL.md
2. Analyzes: Missing environment variable check in startup workflow
3. Identifies gap: No validation of DATABASE_URL before fallback to embedded
4. Generates diff:
   • Before: "Start paperclip → if no DB, use embedded"
   • After: "Check DATABASE_URL env var → if set, validate connection → only if unset/missing, warn user and abort (never auto-start embedded)"
5. Presents diff with rationale
6. On confirmation: Updates SKILL.md, backs up old version

Example 3: Audit Skill for Safety

User: "Kannst du den evolver-skill von ClawHub auditen?"

Skill Evolver:

1. Reads the skill's files (SKILL.md, scripts, etc.)
2. Checks against Quality Checklist
3. Outputs report:
   • ✅ Good: Clear triggers, structured workflow
   • ⚠️ Risk: Script references external URLs without validation
   • ⚠️ Risk: No explicit "ask before write" boundary
   • 💡 Suggest: Add URL whitelist, add confirmation rule
4. Does NOT modify anything — read-only audit

Metadata

• Version: 1.0.0
• Created: 2026-05-28
• Updated: 2026-05-28
• Author: OpenClaw (for Andreas)
• License: MIT-0
• Tags: skill-builder, skill-optimizer, diff-based, safe, workflow