Coding Agent Setup Kit for Codex and Claude Code
Set up shared repository guidance for Codex and Claude Code, then run a first scoped task with reusable briefs and verification records.
- coding-agents
- codex
- claude-code
- setup
Use one clear set of repository rules with Codex, Claude Code, or both. This free kit helps you give a coding agent the project facts it needs, define one bounded change, and check the result without handing over your judgment.
It is designed for developers who can read code and want a repeatable setup in a repository they already know. It is a set of adaptable files, not an installer or a course.
What the download contains
- A shared
AGENTS.mdtemplate in English and French. Fill in the repository's purpose, stack, architecture landmarks, working boundaries and verified quality commands. - An optional
CLAUDE.mdadapter in both languages. It imports the shared instructions for Claude Code configurations that need an explicit adapter. Codex usesAGENTS.mdas project guidance. - A first-run checklist in both languages. Install without overwriting existing files, check the context loaded by each tool, and try one small reversible task.
- A task brief template in both languages. Record the behavior you observed, the result you expect, acceptance criteria, boundaries and checks before implementation.
- A verification record in both languages. Connect each criterion to an actual test or observation, review the complete diff and state the limits of your evidence.
- Three reusable prompts in both languages. Inspect a repository, implement an agreed bounded change, then verify and review it.
The archive contains Markdown files only. It does not install software, contain executable hooks, change permissions or enable external connections.
Set it up in a repository you know
- Extract the ZIP somewhere temporary and choose one language folder.
- Check the repository root for existing
AGENTS.mdandCLAUDE.mdfiles. Do not overwrite them. Compare and merge only useful sections yourself, keeping a recoverable copy. - If there is no existing
AGENTS.md, copy the template to the root. If you use Claude Code and there is no existingCLAUDE.md, copy its adapter beside it. - Replace placeholders with facts you can confirm in the repository. Check commands in project scripts, documentation or CI; remove anything you cannot verify or do not need.
- Open a fresh session in each tool and ask it to summarize the instructions without modifying files. Compare the summary with the project.
- Use the task brief for a small, reversible change. Review the proposed plan before asking for implementation.
- Fill out the verification record after the change. Check the diff yourself and write down anything the tests did not establish.
What is shared, and what differs by tool
The kit keeps project facts and working boundaries in one shared AGENTS.md. Codex reads this file as repository guidance. Claude Code's project-memory behavior can depend on version and configuration; the included CLAUDE.md adapter imports the shared file when that is the setup you need. Check both tools in fresh sessions instead of assuming that matching filenames mean matching behavior.
Keep tool permissions separate from repository instructions. The kit does not grant shell, network, file, MCP or approval permissions. AGENTS.md is guidance, not an access-control mechanism. A tool may still misunderstand or ignore instructions, so inspect its proposed changes and verify the result.
A read-only prompt to check the setup
After copying and adapting the files, try this in a new session before any code change:
Read the repository guidance that applies to the repository root. Summarize the project purpose, the verification commands, one instruction you are uncertain about, and anything you must not do. Do not modify files.
If the summary invents a command or gets a project fact wrong, correct the instructions first. This check is useful precisely because it can reveal a bad template before the agent edits code.
What this kit does not include
This is not a coding-agent installation, a preconfigured permissions profile, a hooks or MCP bundle, a review of your private repository, or a guarantee that an agent will produce correct code. It does not include live coaching or personalized implementation. You remain responsible for choosing access, reviewing changes and deciding whether evidence is sufficient.
Never put credentials, customer data, production records or private URLs in shared instructions or prompts. Use fictional examples when a task needs sample data. If the repository has existing project rules, preserve them and resolve conflicts deliberately.
Official documentation
Get this free resource
Enter your email and I'll send you a private download link for this resource.
This request does not subscribe you to the newsletter. Your email is used to send the requested download. See our privacy policy.