A synthesis of the best AI-assisted development methodologies, designed from the AI agent's perspective for efficient, high-quality delivery.
Five major AI development methodologies exist — gstack, GSD, OpenSpec, Superpowers, and mattpocock/skills. Each has brilliant insights, but each also has blind spots. This project synthesizes their best ideas into a single, cohesive skill system that an AI agent can use to deliver exceptional work.
The AI's perspective matters. Most methodologies are designed for humans to manage AI. This one is designed so the AI itself understands what good engineering looks like and self-directs toward it.
| Source | What We Kept | Why |
|---|---|---|
| mattpocock/skills | Lightweight design, grilling sessions, shared language (CONTEXT.md), caveman compression, /tdd, /diagnose, /improve-codebase-architecture | Minimal process, maximum clarity. Don't own the user's workflow. |
| gstack | Multi-role cognitive perspectives, persistent browser, safety guardrails (/guard, /freeze), "no fixes without investigation" | Different problems need different thinking modes. Safety is non-negotiable. |
| GSD | Fresh context per task, file-based state (.dn/), atomic git commits, goal-backward verification | Context rot is the silent killer of AI quality. State must survive resets. |
| OpenSpec | Spec-driven development, separated reality (specs/ vs changes/), delta-based spec merging, artifact dependency graph | Specs are living documents, not static paperwork. Changes must be traceable. |
| Superpowers | Mandatory skill protocol, git worktree isolation, two-stage review, systematic over ad-hoc | Quality must be structural, not optional. |
- Rigid phase locking — We use spec gates, not phase prisons
- Heavy npm dependencies — Core skills are pure Markdown; tooling is optional
- Named agent systems — Skills are self-contained; agents are created on demand
- Platform-specific code — Skills are platform-agnostic Markdown; install.sh handles distribution
Skills are pure Markdown with YAML frontmatter — the universal format understood by all modern AI coding tools. One set of skills, every platform you use:
| Platform | Skill Directory |
|---|---|
| Claude Code | .claude/skills/ |
| Trae IDE | .trae/skills/ |
| OpenCode | .opencode/skills/ |
| Codex CLI | .codex/skills/ |
| Cursor IDE | .cursor/skills/ |
| Windsurf IDE | .windsurf/skills/ |
| Gemini CLI | .gemini/skills/ |
| GitHub Copilot | .github/skills/ |
| Kilo Code | .kilocode/skills/ |
| Cline (VS Code) | .cline/skills/ |
| Roo Code | .roo/skills/ |
| Qwen Code | .qwen/skills/ |
| Augment Code | .augment/skills/ |
| Auggie CLI | .auggie/skills/ |
| Antigravity | .antigravity/skills/ |
# Install to ALL platforms (recommended — ~80KB, future-proofs your project)
./scripts/install.sh /path/to/your-project
# Or install only the platforms you use today
./scripts/install.sh /path/to/your-project --platform trae,opencode,codex
# Preview what would be installed
./scripts/install.sh /path/to/your-project --dry-runThen in any of your agents, run:
/using-dn-skill
The .dn/ state directory is shared across all platforms — switch between Trae, OpenCode, and Codex seamlessly. A /handoff written in one continues in another.
_meta/ → Boot protocol & efficiency modes
alignment/ → Understand before building
planning/ → Spec → Plan → Issues
execution/ → TDD, prototyping, implementation
review/ → Multi-perspective review & verification
debug/ → Systematic diagnosis & root cause investigation
ops/ → Ship, handoff, browser interaction
safety/ → Guardrails & scope freezing
MIT