ChatGPT and Codex: plan the app, then verify it locally

Start with behavior, not a screen description alone

A useful app brief describes what the developer should be able to verify. For a reading list, specify adding a book, preserving it after a relaunch and returning the original list after clearing a search. Ask the agent to list unresolved decisions before generating a large implementation.

Plan a reading-list app with local persistence and search.
List screens, state transitions and acceptance criteria.
Identify which checks require a running simulator.
Do not claim any check has passed until it has run.

Choose the correct connection path

For a local Codex environment, the project documents this MCP setup:

codex mcp add ios-agent -- npx -y ios-agent-mcp@latest

ChatGPT web is a different environment. It does not run that command on your Mac. Follow the ChatGPT guide for supported skills or a separately configured HTTPS/private-tunnel connection. Availability depends on account and workspace policy; this project does not provide a public hosted Mac.

Make the handoff explicit

Carry the brief and acceptance criteria into the local project. Ask the connected agent to inspect existing code, retrieve relevant references and propose one implementation slice. Keep the actual app source and test results in the local development workflow. A plan written in a browser is not a simulator result.

Review the completion message

Require the changed files, commands run, outcomes and remaining gaps. Compilation is useful but does not prove persistence or accessible states. Use the simulator for visual evidence and tests for behavior where practical. Do not infer better token efficiency merely because references are retrieved in smaller sections.

Official Codex MCP documentation · Project setup instructions

Decision checkpoints