Repo-first coding guardrails for ChatGPT, Codex, Claude, and Gemini.
Short account-global prompts do not carry repo commands, architecture constraints, test expectations, or local invariants well enough. Put the real control plane where the work happens.
- Repo-level files
- Project-level instruction surfaces
- Skills or Gems
- One-shot bootstrap prompts
- Account-global fallback
| Platform | Best file | Default use |
|---|---|---|
| Codex | exports/codex/AGENTS.md |
Repo commands, invariants, and verification rules |
| Claude Code | exports/claude/CLAUDE.md |
Repo-local memory for code work |
| ChatGPT Projects | exports/chatgpt/PROJECT_INSTRUCTIONS.md |
Project scope when repo files are unavailable |
| Claude Projects | exports/claude/PROJECT_INSTRUCTIONS.md |
Project scope when repo files are unavailable |
| Gemini Gems | exports/gemini/GEM_INSTRUCTIONS.md |
Portable behavior layer paired with repo docs |
| Any chat UI | exports/any-llm/BOOTSTRAP_PROMPT.md |
Disposable first-message contract |
| Account-global fallback | exports/chatgpt/GLOBAL_FALLBACK_SHORT.txt |
Emergency-only fallback |
The common failure pattern is simple: a short global prompt is asked to carry test commands, lint rules, protected paths, architecture constraints, and local invariants across every coding task.
That is the wrong control plane.
Repo-specific behavior belongs in a repo file or a project instruction surface.
exports/ Platform-ready instruction files
skill/ Source for the ChatGPT skill bundle
templates/ Project appendix template
docs/ Install, proof, demo, launch, and review material
dist/ Packaged ChatGPT skill
- Copy
exports/codex/AGENTS.mdto the repo root asAGENTS.md. - Append the fields from
templates/PROJECT_APPENDIX.md. - Fill in commands, critical paths, and invariants.
- Copy
exports/claude/CLAUDE.mdto the repo root asCLAUDE.md. - Append the fields from
templates/PROJECT_APPENDIX.md. - Fill in the repo-specific appendix.
- Create a project.
- Paste
exports/chatgpt/PROJECT_INSTRUCTIONS.mdinto Project instructions. - Append the project appendix fields.
Use dist/skill.zip when a reusable ChatGPT skill is the better fit.
The proof layer documents what is demonstrably included and what still requires live usage evidence.
docs/proof/WHAT_IS_PROVEN.mddocs/proof/DEMO_SCENARIOS.mddocs/demo/TRANSCRIPT_REPO_FIRST.mddocs/demo/TRANSCRIPT_FALLBACK_ONLY.md
Publication assets live in docs/launch/.
- release notes
- repo settings guidance
- social copy
- launch checklist
- social preview image
The pack keeps the upstream four-principle spine and adds:
- a stricter operating contract for non-trivial tasks
- a verification ladder and reporting rules
- exports for ChatGPT, Codex, Claude, and Gemini
- fallback demotion for the account-global file
- a project appendix for repo commands and invariants
- publication collateral and proof material
See docs/UPSTREAM_REVIEW.md.
Inspired by an upstream coding-guidelines package. See NOTICE.md and docs/ATTRIBUTION.md.
See NOTICE.md and docs/ATTRIBUTION.md.
MIT
