A project-agnostic four-agent AI development workflow for Claude Code, Gemini CLI, and GitHub Copilot.
| File | Purpose |
|---|---|
AGENTS.md |
Single source of truth — workflow rules, agent definitions, conventions |
CLAUDE.md |
Claude Code entry point → reads AGENTS.md |
GEMINI.md |
Gemini CLI entry point → reads AGENTS.md |
.github/copilot-instructions.md |
GitHub Copilot entry point → reads AGENTS.md |
.ai/agents/ |
Agent role definitions (orchestrator, tester, developer, reviewer) |
.ai/plan-template.md |
Required format for presenting implementation plans |
.ai/plugins.md |
Plugin assignments per agent |
.ai/lessons-learned.md |
Running log of mistakes and how to avoid them |
.claude/settings.json |
Claude Code configuration (permissions, model, hooks) |
- Copy all files from this folder into the root of your new repository.
- Open
AGENTS.mdand replace the<!-- CUSTOMIZE -->section in Purpose with a description of your project. - Populate
.claude/settings.jsonwith your Claude Code permissions and preferences. - Delete this README or replace it with your project README.
- Workflow changes → edit
AGENTS.mdonly. The three tool entry files automatically inherit it. - Agent behavior → edit the relevant file in
.ai/agents/. - Mistakes log → add entries to
.ai/lessons-learned.mdas you work. - Plugin assignments → edit
.ai/plugins.md.
User approves plan
↓
Orchestrator breaks into tasks
↓
Tester writes BDD tests
↓
Developer implements until tests pass
↓
Reviewer checks quality + security
↓
Orchestrator commits → next task
All workflow updates belong in AGENTS.md.