Cross-model project folder

cross-model-project-starter/
├── PROJECT_MAP.md
├── AGENTS.md
├── CLAUDE.md
├── GEMINI.md
├── Decisions/
│   └── DECISION_TEMPLATE.md
├── Processes/
│   └── PROCESS_TEMPLATE.md
├── References/
│   └── REFERENCE_INDEX.md
├── Inputs/
│   └── README.txt
└── Outputs/
    └── README.txt

What each part does

PROJECT_MAP.md

The front door for a person or model:

• what the project is
• what matters now
• current owner
• definition of done
• where each kind of truth lives

Keep it short. A map points somewhere; it does not contain the whole company.

Decisions

Record choices you do not want reopened every session:

• what was chosen
• owner and date
• evidence and trade-offs
• consequences
• measurable condition for revisiting it

Do not hide an unresolved proposal inside the decision log as if it were approved.

Processes

Store repeatable instructions:

• outcome
• required inputs
• exact steps
• objective quality checks
• human approval point
• edge cases learned from real runs

Update the process when reality disagrees.

References

Stable rules, approved examples, definitions, source links, brand rules, and facts. Give every reference an owner and last-reviewed date.

Inputs and Outputs

• Inputs: raw, unapproved material waiting for work
• Outputs: approved final artifacts

Do not let an agent treat an Input as approved truth. Keep the state visible in the folder name and file.

Why three adapter files exist

There is no single magic filename every model reads identically.

• Claude Code reads CLAUDE.md. Anthropic explicitly recommends a thin CLAUDE.md that imports @AGENTS.md when a repository already uses AGENTS.md.
• Codex reads AGENTS.md and AGENTS.override.md from the project hierarchy, subject to its project-document budget.
• Gemini CLI reads GEMINI.md by default and supports @file.md imports. Its context.fileName setting can also include AGENTS.md.

The starter uses:

CLAUDE.md → @AGENTS.md
GEMINI.md → @./AGENTS.md
AGENTS.md → PROJECT_MAP.md + shared operating rules