Use ChatGPT for planning and Codex for local execution
They set up a handoff where ChatGPT designs the work and checks it, while Codex makes the changes and runs commands on the local machine.
- What they did
- The poster connected ChatGPT to a local Codex installation using MCP-style file access. After seeing ChatGPT work through a local CAD task by asking clarifying questions and coordinating subagents for close to an hour, they put together a slim workflow that sends planning and review to ChatGPT and leaves the hands-on local steps to Codex. They added simple safety limits to the process.
- How it went
- A short CAD trial suggested ChatGPT was better at the higher-level thinking, even though the CAD result came out only middling. The arrangement is still experimental, and the poster asked others how they would structure it. It was made for the Windows version of Codex, and because computer-use inside native apps often failed, they used Codex web Chrome control instead.
- Tips
- Separate the thinking stage from the doing stage, and make sure irreversible actions require a checkpoint. If direct app control is flaky, use the more dependable route for local execution.
Try it yourself with ChatGPT
You are the planning side of a two-agent workflow for [project/repo]. Read my goal, ask any needed clarifying questions, and write a step-by-step plan. Then stop and wait for my approval before any edits happen. After I approve, hand the work to the local executor to make the changes, run [test/build command] after each major step, and return a brief report with the files changed, checks run, and any problems found. Do not delete, submit, send, overwrite, or push anything until I explicitly confirm. Goal: [paste goal] Context: [paste relevant files, constraints, or prior plan]
Fill in the [brackets]. Agents can make mistakes, so review anything before it is sent, paid or deleted.
1 more report of this
Get 3 agent uses like this every morning
With the day's top AI news. Free, and one click to unsubscribe.
Related agent uses
Use an agent to patch a stubborn app bug
They pointed a coding agent at two annoying software behaviors and had it debug one and add a self-improvement feature to the other.
Use a coding agent to fix a Slack-reported bug
A Codex agent handled a Slack issue on its own, and an ops agent later confirmed the fix in the same thread.
Use a long-running agent to chase a stubborn bug
A developer let Codex work overnight against a repeatable test, using debug prints and hypotheses to track down an unexplained bug.
Set up a separate sandbox QA pass
Codex turned an existing Claude login into a second reviewer that checked an AI system without changing the userβs credentials.
Use spare AI quota to draft reviewable PRs
A founder pointed an agent at goals and task lists so it could turn leftover Claude and Codex quota into pull requests for review.
Use ChatGPT to narrow down Swift error causes
A Swift developer pastes an error and nearby code into ChatGPT to get a quick first guess, then checks the output before changing anything.