agent-system-starter

A minimal .agent/ setup to make AI coding workflows more predictable.

Works with:

• Claude Code
• Codex / OpenAI agents
• Gemini CLI
• Google Antigravity
• Cursor
• Any custom agent setup

---

Why this exists

Most AI workflows look like:

prompt → retry → tweak → retry

This repo shows a better way:

define behavior + rules + workflows → get consistent output

This is the missing layer between prompts and production AI systems!

---

Quick Start

1. Copy .agent/ into your repo, plus the loader file(s) for your tool:

   • CLAUDE.md — Claude Code
   • AGENTS.md — Codex / OpenAI agents / Antigravity
   • GEMINI.md — Gemini CLI
   • .cursor/rules/ — Cursor

2. Start with:

   • behavior.md
   • rules/base.md
   • one workflow

3. Use your AI tool as usual

---

What you get

• fewer random changes
• less repetition in prompts
• more predictable outputs

---

Structure

.
├── CLAUDE.md            # Claude Code loader (imports AGENTS.md)
├── GEMINI.md            # Gemini CLI loader (imports AGENTS.md)
├── .cursor/rules/       # Cursor loader (points to AGENTS.md)
├── AGENTS.md            # canonical load order — Codex / Antigravity read this directly
├── .agent/              # the pattern: instructions
│   ├── behavior.md
│   ├── rules/
│   ├── workflows/
│   ├── skills/
│   └── memory/
└── .claude/             # optional: Claude Code permissions (not part of .agent/)
    ├── settings.json                # team defaults
    └── settings.local.json.example  # personal template

---

Tool permissions (optional)

The .agent/ pattern covers instructions, not permissions. Bash and action permissions belong in each tool's native settings — they're not part of .agent/.

Claude Code uses a two-file pattern:

• .claude/settings.json — team defaults, committed. Hard blocks every contributor should share (rm -rf /*, reading .env, etc.).
• .claude/settings.local.json — personal layer, gitignored. Your own allow / ask lists. Copy .claude/settings.local.json.example to activate.

Other tools use their own user-level or IDE-level config.

---

Key Idea

Prompts tell the AI what to do. This defines how it behaves while doing it.

---

Start Simple

Don't use everything.

Start with:

• behavior
• rules
• one workflow

Add more only when needed.

---

Deeper explanation

This repo separates:

• intent (prompts)
• behavior (system structure)

📖 Full blog: https://medium.com/@nerdapplabs/the-agent-pattern-a-simple-anatomy-for-ai-systems-76a4d30a9a9b