2025
The Composite Skill Ecosystem
Skills That Build Skills
GitHub RepoThe Challenge
LLM skill development has no quality control. Engineers build prompts in isolation, duplicate effort across teams, and watch their work break silently when models update or requirements drift. There's no standard for what "production-ready" means, no validation before deployment, and no governance for ongoing maintenance. The tooling exists. The tool-making discipline doesn't.
Approach
Built a self-referential architecture where skills create, evaluate, orchestrate, and govern other skills. Separated concerns into distinct lifecycle phases: complexity detection determines right-sized structure before building begins; scientific evaluation validates quality against defined criteria; multi-agent coordination handles orchestration without bottlenecking on a central dispatcher; governance enforces quality at scale through mandatory pause points and audit trails.
Technical Implementation
Skill-Builder: Intelligent complexity detection with three tiers—single files, structured folders, or principles-first mode that extracts domain expertise into transmittable artifacts
Skill-Evaluator: 7-criteria framework grounded in evaluation science (JCSEE standards, Kirkpatrick model), producing scores with evidence and actionable fixes
Agent-Team-Builder: Stigmergic coordination pattern—agents communicate through shared state rather than direct messaging, preventing hierarchical bottlenecks
Project-Steward: 10-agent audit committee, mandatory pause points, git integration with commit strategies
Master-Copywriter: Direct-response methodology built on Schwartz awareness stages and conversion frameworks
Outcome
2,300 lines of production-ready skill definitions across 5 interdependent skills. The ecosystem is self-validating: Skill-Evaluator can assess Skill-Builder's output; Project-Steward can audit the evaluator. Principles-first mode reduced context requirements for complex builds by ~60%. The stigmergic coordination pattern is a novel contribution to multi-agent architecture.