Mysk Extractor

Migrate and crossbreed skills between repositories while documenting provenance for reproducibility research.

Processes

This skill defines six processes, named after mycological and cellular biology:

Pedigree Schema

Every extracted skill gets a genealogy file documenting its lineage.

Schema Structure

See genealogy/templates/pedigree.json for the schema template.

Field Reference

Crossbreeding Support

The pedigree array supports multiple entries for hybrid skills:

• Single entry = conidiation (one parent)
• Multiple entries = plasmogamy (multiple parents fused)

Experimental Design: Isolated Agents

Why Exclude Human Input

The pedigree schema captures all variables needed to theoretically reproduce a skill extraction:

• Source content (via sourceCommitID)
• Transformation logic (via destOperator at destCommitID)
• Model version (via destModel)

However, if human prompts influence the extraction, they become an uncontrolled variable that breaks reproducibility.

Isolating Claude's Non-Determinism

By designing extraction processes to:

1. Spawn subagents rather than work in the main conversation
2. Feed structured data rather than conversation history
3. Exclude human prompt context from the transformation

We control for all external variables, leaving Claude's inherent non-determinism as the only remaining variable.

This enables experiments like:

• Run identical extractions N times → measure output variance
• Compare variance across different operator designs
• Correlate variance with task complexity

The pedigree file serves dual purposes:

1. Genealogy record - trace skill ancestry
2. Experimental control log - prove all controllable variables were fixed

Implications for Process Design

Both fragmentation.md and plasmogamy.md processes MUST:

• Use the Task tool to spawn isolated subagents
• Provide all context via structured parameters, not conversation history
• Record the exact operator and model in the pedigree

Templates

Workflow (Fragmentation)

See substrate/fragmentation.md for the full single-source extraction workflow.

High-level phases:

1. Configure - Define source/destination and cleanup rules
2. Copy - Transfer skill files
3. Cleanup - Generalize project-specific references
4. Genealogy - Create pedigree record
5. Commit - Finalize with descriptive message

Checklist

• Process spawns isolated subagent (no conversation context)
• Subagent receives structured data only
• Skills copied to destination
• Project names generalized
• Domain terms universalized
• Pedigree file created with all fields populated
• destCommitID recorded after commit