GitHub - ninjaa/openai-codex-exec-plan

2 min read Original article ↗

ExecPlan Template for Codex

AGENTS.md and ExecPlan markdown templates for programming efficiently with AI coding assistants.

"ExecPlan" becomes shared vocabulary between you and your agent—a code word for task-scoped work plans that structure how features get built.

Inspired by Aaron Friel's OpenAI 2025 DevDay presentation: "Shipping with Codex"

Maintained by Aditya Advani

The Workflow

  1. Create .agents/template/PLAN.md — a structured markdown template for feature planning
  2. Configure AGENTS.md — instructs your AI agent to create a new ExecPlan for each task by copying the template into a new feature folder (e.g., .agents/orchestrator/PLAN.md)
  3. Say "do the plan" — the agent executes methodically, usually one-shotting or completing in 2-3 iterations

Why It Works

  • Structured thinking: Forces the agent (and you) to clarify purpose, context, and steps upfront
  • Living document: Progress, decisions, and surprises stay in the plan
  • Clean workspace: Each feature gets its own folder; option to delete when merged
  • Repeatability: The template encodes best practices—TDD, decision logs, verification steps

Quick Start

  1. Copy this repo's .agents/template/PLAN.md and AGENTS.md into your project
  2. (Optional) Add to .gitignore to keep plans local: 
    .agents/*
!.agents/template/
  3. Tell your agent: "Create an ExecPlan for [feature] following the template"
  4. Execute: "Do the plan"

Usage Notes

  • One ExecPlan = One Task: Each plan maps to a task (feature, refactor, bug fix, etc), sometimes with subtasks tracked as well in the Progress section
  • TDD first: AGENTS.md includes test-driven development guidance
  • Concurrency: Running multiple plans in parallel is possible but requires orchestration (still figuring this out myself)
  • Cleanup: Delete feature folders post-merge if working in a shared codebase
  • Slight variation: In the video, Aaron uses .agent but I use .agents to match AGENTS.md naming. I keep forgetting otherwise.

Contributing

If you've adapted this workflow or have insights on running multiple concurrent agents (especially in Codex web or as codex cli background jobs in sandboxes), please share!

License

MIT