Summary

This project provides a "skill-based architecture" to manage scattered AI agent guidance, distilling project rules, workflows, and lessons into a centralized, routable skill asset under skills/<name>/. It establishes a single source of truth for AI agents like Cursor, Claude Code, and Codex, enhancing efficiency, consistency, and maintainability of AI project configurations.

How It Works

The architecture transforms disparate prompt documents into a standardized skill structure. It features a concise SKILL.md router, organized directories (rules/, workflows/), and "thin shells" in existing locations that route to the central skill, preventing duplication. Key elements include routing.yaml as the source of truth, "description-as-trigger" for activation, and self-maintenance via line-count signals.

Quick Start & Requirements

Pull the repository locally (e.g., git clone) to a location accessible by your AI agent (e.g., skills/skill-based-architecture/ or ~/.cursor/skills/skill-based-architecture/). Trigger the meta-skill via agent instruction (e.g., "Use skill-based-architecture to refactor the project rules"). The agent copies a scaffold into skills/<name>/. Note: Codex requires explicit user authorization for sub-agent delegation. Detailed migration steps are in WORKFLOW.md.

Highlighted Details

• Two-layer routing: SKILL.md (essentials) + routing.yaml (conditional).
• Thin shells: Lightweight entry points across agent formats (Cursor, Claude, Codex) bootstrap routing, preventing duplication.
• Description-as-trigger: Natural language activation phrases.
• Session Discipline & Task Closure: Re-reads SKILL.md per task; AAR for task completion.
• Self-maintenance: Line-count signals trigger evaluation; split/merge procedures keep docs lean.
• Cross-harness compatibility: Works with Cursor, Claude, Codex, Windsurf, Gemini, AGENTS.md tools.

Maintenance & Community

Includes upstream maintenance scripts (scripts/). Specific community channels, contributors, or sponsorships are not detailed in the provided README.

Licensing & Compatibility

No specific open-source license is mentioned. Compatibility for commercial use or closed-source linking cannot be determined from this information.

Limitations & Caveats

Not recommended for projects with < 50 lines of rules, short-lived solo projects (< 2 weeks), or scenarios without team sharing. Codex requires explicit user permission for sub-agent delegation due to tool-level rules.