How We Built a Zero-Dependency Multi-Agent Skill Compiler & 239-Gate Mutation Harness for Cursor, Claude Code, and Antigravity

Description and use

A zero‑dependency compiler and verification harness were built to eliminate common structural failures in AI agent skill repositories, producing lightweight, synchronized artifacts and rigorous tests.

Problem

Custom AI agent skill repositories frequently suffer from target drift, bloated context, storefront formatting issues, and superficial verification, leading to incompatible, hard‑to‑maintain, and unreliable skills.

Outcome

The factory produces consistent, platform‑agnostic skill artifacts, reduces LLM context waste by 95.1%, fixes storefront formatting bugs, and enforces rigorous verification, preventing the four common structural failures.

Workflow

• Define two canonical inputs: product.manifest.json and pro-bundle/rules/*.md
• Run node factory.mjs compile . to generate six targets
• Generate lightweight root context files (CLAUDE.md, GEMINI.md) capped at 200 lines
• Prepend YAML frontmatter to .mdc files for Cursor
• Emit clean SKILL.md files for Claude Code and Antigravity
• Parse and write Node.js linters to scripts/lint_*.mjs
• Create llms.txt index and bin/cli.mjs for zero‑install testing
• Build POSIX ZIP archives with paywall isolation
• Run audit with factory.mjs audit to enforce 239 hard checks
• Execute adversarial mutation suite challenge_verification_suite.mjs

Evidence

Every week, thousands of developers publish custom AI agent skills, .cursor/rules/*.mdc files, and CLAUDE.md prompt instructions. Yet if you inspect those repositories two weeks after launch, nearly all of them suffer from the same four structural failures:

Source

To solve these four failures permanently, we built the AI Skill & DevTool Commercialization Factory (@longwave-ai-lab/skill-commercialization-factory), an 8-stage, zero-dependency Node.js state machine and dual-layer verification harness.

Source

The compiler generates six synchronized targets deterministically: Target 1: Root Context Index (CLAUDE.md & GEMINI.md <= 200 Lines)

Source