When Meta-Skills Collide

The Original Meta-Skill

Skill-Coach was the experiment that started all of this. Its job: look at any skill and make it better. Better triggering, cleaner structure, more honest about what it can and can't do. Eight reference files, a recursive self-improvement workflow, and then: run it on itself five times.

skill-coach/
├── SKILL.md                           ~400 lines
├── CHANGELOG.md
├── scripts/
│   ├── validate_skill.py
│   ├── check_self_contained.py
│   └── test_activation.py
└── references/
    ├── antipatterns.md                anti-pattern catalog with case studies
    ├── shibboleths.md                 expert vs novice vocabulary patterns
    ├── validation-checklist.md        complete review and testing guide
    ├── self-contained-tools.md        scripts, MCPs, and subagent patterns
    ├── scoring-rubric.md              quantitative 0-10 skill evaluation
    ├── skill-composition.md           cross-skill dependencies
    ├── skill-lifecycle.md             versioning and deprecation
    └── mcp_vs_scripts.md              when to use Skills vs Agents vs MCPs

How It Works

Skill-Coach applies a six-step creation process and a progressive disclosure philosophy:

The description formula: [What] [Use for] [Keywords] NOT for [Exclusions]. Its own description is the example: "Guides creation of high-quality Agent Skills... Activate on: create skill, review skill, skill quality... NOT for general coding advice, slash commands, MCP development."

The recursive self-improvement workflow uses its own scripts:

python scripts/validate_skill.py <path>       # structural check
python scripts/check_self_contained.py <path>  # phantom reference check
python scripts/test_activation.py <path>       # activation rate check

Address ERRORS first, then WARNINGS, then SUGGESTIONS. Update CHANGELOG.md. Re-run until clean.

What Skill-Coach^5 Found About Itself

Each generation found something the previous one was too close to see:

By SK₅: tighter triggering, self-consistent examples, a workflow that matched its own structure. Cleaner, not longer.

Why It Matters Here

Skill-Coach established that a meta-skill can improve itself. SA and SC are meta-skills with different improvement philosophies. The question this experiment asks: what happens when you cross-apply them instead of self-applying them? Do they find the same things?

They don't.

Thank You, Anthropic

This experiment exists because Anthropic open-sourced their skill creation tooling under the Apache 2.0 license. They didn't have to. The license permits reproduction, derivative works, and public display with attribution. We include their complete skill-creator with full provenance — every file, every script, every agent definition. For the latest version: their repository. Everything here is a snapshot from commit b0cbd3df, March 7, 2026.

The Setup

Each skill got its own source folder, the target folder (read-only), and a writable output copy. Tools: Read, Write, Edit, Glob, Grep, Bash.

claude -p "You are Skill Architect. Your source folder: {sa_path}.
           Evaluate and improve skill-creator at: {sc_path}.
           Write all improvements to: {output_path}." \
  --allowed-tools Read,Write,Edit,Glob,Grep,Bash(python:*) \
  --permission-mode bypassPermissions

Each agent read its own skill folder — references, scripts, examples — before scoring the target. Layer 1: SKILL.md only. Layer 2: full folder access. Layer 3: after each skill self-improved, cross-evaluate again.

Three rounds. Six crossings. The braid is below.