Meta Skills and the Skill Creator | Generated by AI
Question: What is a “meta skill” and what is the “skill-creator” skill in this environment?
Answer:
Meta Skill
A meta skill is a skill about skills — instead of teaching you how to do X, it teaches you how to create/improve/evaluate skills that do X.
In this environment’s skill system, skills are structured .md files that give Claude task-specific instructions (e.g., how to create a .docx, how to read a PDF). A meta skill operates one level up: it governs the skill creation and improvement process itself.
The skill-creator SkillHere’s what you need to know:
Meta Skill — Concept
A meta skill operates one level of abstraction above regular skills. Regular skills teach Claude how to do a task (make a .docx, read a PDF, design a UI). A meta skill teaches Claude how to create, evaluate, and improve the skills themselves.
The analogy from ML: a skill is a model, a meta skill is the training loop + evaluation harness for that model.
skill-creator — What It Actually Does
It’s the one meta skill in this environment. It implements a full skill engineering loop:
define intent → draft SKILL.md → write test cases → run evals → review outputs → iterate → package
Concretely it covers:
1. Skill authoring — Structured SKILL.md files with YAML frontmatter (name, description) + markdown instructions. Three-layer progressive disclosure: metadata (always in context) → SKILL.md body (on trigger) → bundled resources (on demand).
2. Triggering optimization — The description field in frontmatter is the primary mechanism by which Claude decides to use a skill. The skill-creator runs a run_loop.py optimization script that A/B tests descriptions against a 60/40 train/test eval split, iterating up to 5 times to maximize trigger rate without overfitting.
3. Eval harness — Writes evals.json, runs test prompts, grades outputs via assertion checking, and generates an HTML review viewer (generate_review.py) for human feedback before the next iteration.
4. Packaging — package_skill.py bundles the final skill into a .skill file for installation.
Skill File Structure
my-skill/
├── SKILL.md ← required; YAML frontmatter + instructions
└── (optional)
├── scripts/ ← executable code for deterministic steps
├── references/ ← docs loaded into context as needed
└── assets/ ← templates, fonts, icons
Why It Matters for You
Since you build CLI agents and automation tools (ww, iclaw, zz), the skill system is essentially prompt engineering with version control and evals. The skill-creator gives you a reproducible loop for shipping reliable Claude behaviors — same philosophy as iterating on a fine-tuned model but at the prompt layer. You could author skills for your own workflows (e.g., a skill for auto-generating your blog posts in your style, or a skill for structured code review against your bank’s engineering standards) and use skill-creator to tune and test them systematically.