Agentic Code Stack

A portable reference package you drop into the root of any project so your AI coding agent (Claude Code, Codex, Cursor, Gemini CLI, etc.) immediately knows what skills, plugins, and MCP servers are available, when to use them, and how to invoke them.

Read PHILOSOPHY.md for the principles behind this stack. The core idea: every coding session runs on two tracks. The pipeline defines sequential stages (shape, align, plan, design, build). The rails are always-on guardrails (docs context, token efficiency, memory tracking, external data) that keep the agent on track at every stage.

What's inside

agentic-code-stack/
├── AGENTS.md               # Primary entry point — agents read this first
├── CLAUDE.md               # Symlink (or copy) of AGENTS.md for Claude Code compatibility
├── README.md               # You are here
├── INSTALL.md              # One-time setup commands per tool
├── PHILOSOPHY.md           # The principles and mental model behind the stack
├── index.html              # Visual map (hover any tool for commands)
├── community-recipes/      # User-contributed setups for different stacks and workflows
├── pipeline/
│   ├── 01-prompt-shape.md
│   ├── 02-align.md
│   ├── 03-spec-plan.md
│   ├── 04-design.md
│   └── 05-build.md
└── rails/
    ├── 01-docs-context.md
    ├── 02-token-efficiency.md
    ├── 03-memory-tracking.md
    ├── 04-external-data.md
    └── 05-agent-interfaces.md

How to use

1. Copy this folder to the root of every new project. Most agents (Claude Code, Codex, Cursor) auto-read AGENTS.md at session start. If your agent uses CLAUDE.md instead, the included copy works as-is. Both files exist so you do not need to choose.

2. First session in a new project: ask the agent to read AGENTS.md (or CLAUDE.md) and acknowledge the stack. Then run any first-time setup from INSTALL.md.

3. Open index.html in a browser when you (the human) want a visual reminder of what's available — hover any pipeline stage or rail tool to see the slash commands.

4. Edit freely. Remove tools you don't use. Add ones you do. The package is meant to be customized per-project.

Links

• GitHub Discussions — Share your setup, ask questions, and see how others configure their stack.
• community-recipes/ — Contribute a recipe for your toolchain. PRs welcome.

The mental model

Pipeline (sequential): Prompt Shape → Align → Spec & Plan → Design → Build

Rails (always-on, parallel to the pipeline):

• Docs & Context — Ref MCP, Context7, Pitlane MCP
• Token Efficiency — Caveman, Pitlane MCP
• Memory & Tracking — Beads, OpenSpec
• External Data — Brightdata MCP

The pipeline tells the agent what stage of work it's in. The rails tell it which tool to reach for at each stage.

Credits & sources

• Grill-me / write-a-skill — mattpocock/skills
• Superpowers — obra/superpowers
• OpenSpec — Fission-AI/OpenSpec
• Get Design (DESIGN.md) — getdesign.md · skill wrapper zephyrwang6/brand-design-md
• Open Design — nexu-io/open-design
• SwiftUI Agent Skill — twostraws/SwiftUI-Agent-Skill
• Caveman — JuliusBrussee/caveman
• Beads — gastownhall/beads
• Pitlane MCP — eresende/pitlane-mcp
• Ref MCP — ref-tools/ref-tools-mcp
• Context7 — upstash/context7
• Brightdata MCP — Bright Data
• Metaprompt — metaprompt.com

All third-party tools belong to their respective authors. This package is a reference layer, not a redistribution.

Community contributions are credited in the individual recipe files and in our GitHub Discussions.