Frankenstein

Model Requirements

Default: Opus (or best available thinking model)

Frankenstein requires deep reasoning to:

• - Compare multiple skill approaches
• Identify subtle methodology differences
• Synthesize the best parts creatively
• Catch security/quality issues others miss

Only use a smaller model if user explicitly requests it for cost reasons. The synthesis quality depends heavily on reasoning depth.

Create monster skills by combining the best parts of existing ones.

Quick Start

CODEBLOCK0

How It Works

Step 1: Search All Sources

Search EVERY AI skills repository for matching skills:

1. ClawHub (primary)
CODEBLOCK1

2. GitHub
CODEBLOCK2

3. skills.sh
CODEBLOCK3

4. skillsmp.com (Skills Marketplace)
CODEBLOCK4

5. Other sources to check:

• - Anthropic's skill examples
• OpenAI GPT configurations (convert to skill format)
• LangChain agent templates
• AutoGPT/AgentGPT skill repos

Gather all candidates before filtering. More sources = better Frankenstein.

Step 2: Security Scan

For each skill found:

• - Install to temp directory
• Run skill-auditor scan
• Score >= 7 = SAFE (proceed)
• Score < 7 = RISKY (skip with warning)

Step 3: Safe Analysis

For each safe skill, extract:

• - Core features (what it does)
• Methodology (how it approaches the problem)
• Scripts/tools (reusable code)
• Unique strengths (what makes it special)
• Weaknesses (what's missing)

Step 4: Compare

Step 5: Synthesize

• - Feature 1 methodology from skill-A
• Feature 2 implementation from skill-B
• Feature 3 approach from skill-C

Step 6: Build Initial Draft

• - Combine winning features
• Resolve conflicts (if two approaches clash)
• Write unified SKILL.md
• Include scripts from winners
• Document sources

Step 7: Vetting Loop (CRITICAL)

Run plan → test → improve loop until 3 stable passes:

CODEBLOCK5

What to look for each pass:

• - Missing features that sources had
• Contradictions between combined approaches
• Vague instructions that aren't actionable
• Token waste (verbose where concise works)
• Security gaps
• Broken references to files/scripts

Document in VETTING-LOG.md:

• - Each pass number
• Issues found
• Fixes applied
• Why considered stable

Only proceed when:

• - 3 consecutive passes with no major issues
• Minor issues documented as known limitations

Step 8: Human Review

• - Show what came from where
• Highlight conflicts resolved
• Show vetting summary
• Ask for final OK before saving

Output

Creates a new skill with:

• - Best features from all analyzed skills
• Clear attribution (credits source skills)
• Security-scanned components only
• Unified documentation

Example Session

User: "Frankenstein me an SEO audit skill"

Frankenstein:
CODEBLOCK6

Dependencies

This skill uses:

• - clawhub CLI (search/install)
• skill-auditor (security scanning)
• sandwrap (safe analysis)
• skill-creator (building)

Spawning Sub-Agents

When spawning analysis sub-agents, always use Opus (or best thinking model) unless user explicitly requests otherwise:

CODEBLOCK7

Cheaper models miss nuances between skills and produce shallow combinations.

Limitations

• - Only combines publicly available skills
• Skips skills that fail security scan
• Cannot resolve deep architectural conflicts
• Human judgment needed for final synthesis
• Quality depends on available skills

Credits

When a Frankenstein skill is built, it includes attribution:
CODEBLOCK8