What is an AGENTS.md file used for?

You use an AGENTS.md file to define project rules, coding standards, and guidelines to keep Codex aligned with your project's requirements.

How can you review code generated by Codex?

Users suggest running two Codex instances, one as a developer and one as a reviewer. You should give the reviewer a QA checklist and always perform your own self-review.

How do you keep Codex from generating too much code when you want to learn?

You need to explicitly instruct Codex not to generate large blocks of code if you want it to explain concepts or provide tutorials instead.

Codex Best Practices for Planning, Coding, and Review

Detailed planning and breaking tasks into small tickets help manage complexity when using Codex.

Users recommend a multi-agent setup where one instance codes and another reviews against a QA checklist.

You should micromanage code generation with detailed skeletons and always perform your own review to catch errors.

Best practices

Create a multiphase plan Detail the high-level implementation and granular specifics for your project.
Set up an AGENTS.md file Define rules, coding standards, and guidelines to align Codex with your project.
Build small tickets Avoid building large systems at once and create small, verifiable tickets instead.
Use a multi-agent review Run one Codex terminal as a developer and another as a reviewer.
Use a QA checklist Provide the reviewer agent a checklist for functionality, clarity, and testing.
Micromanage code generation Provide a detailed skeleton with structure, function names, and variables.

To effectively use Codex, prioritize detailed planning, strict guidelines, and continuous human oversight to prevent errors and ensure code quality.

Leverage Detailed Planning and Documentation

Create comprehensive plans: Develop a multiphase plan for features or fixes, detailing the high-level implementation and granular specifics for Codex. "Have it create a multiphase plan for creating the feature/fix you want."

Utilize a AGENTS.md file: Define project rules, coding standards, and guidelines within an AGENTS.md file to keep Codex aligned with your project's requirements. "Setup AGENTS.md"

Break down tasks into small tickets: Avoid asking Codex to build large systems at once; instead, create small, verifiable tickets to manage complexity and facilitate review. "Build one small ticket at a time"

Implement a Multi-Agent Review Process

Employ a developer and reviewer agent: Use one Codex instance as a developer agent to implement code and another as a reviewer agent to check for issues after each phase. "Have two codex terminals open."

Establish a QA checklist for reviews: Provide the reviewer agent with a checklist covering functionality, completeness, consistency, clarity, and testing to ensure thorough evaluation. "I recommend having a QA checklist for things you want the reviewer to look for."

Perform self-review and question Codex's confidence: Always review the code yourself, and ask Codex to rate its confidence in the completion and bug-free status of the feature. "Make sure to review yourself."

Address Codex's Limitations and Enhance Learning

Explicitly instruct Codex on learning vs. coding: If you want to learn, explicitly tell Codex not to generate large blocks of code but rather to explain concepts or provide tutorials. "Tell it to not do that lmao."

Micromanage code generation: Provide Codex with a detailed skeleton, including structure, function names, and variables, to ensure the generated code aligns with your design. "Tell it exactly what to code, like micromanage the shit out of it."

Use external LLMs for planning and learning: Leverage other LLMs like ChatGPT for initial planning, brainstorming, and learning, then translate those plans into precise prompts for Codex. "I use ChatGPT to plan and then have it give me a codex prompt to make it work."

Are you interested in specific tools or plugins that Users recommend for use with Codex?