One project. Several instruments. One human.
The 2024 slogan was “Claude builds, ChatGPT debugs.” Sometimes that still works. It is not a law. The law is: PROJECT.md is the brain, agents are interchangeable where the job overlaps, and you keep consequential buttons.
Jobs, not mascots
Plan in chat
Free ChatGPT, Grok, Claude, or Gemini in a browser. Cheap thinking. No folder access required.
Build in an agent
Grok Build, Codex, or Claude Code in the repo. Same PROJECT.md. Pick today’s instrument on the bench.
Unstick with a stranger
When one agent loops, screenshot the error. Paste it into a different AI. Bring the diagnosis back. Do not keep punching the same bruise.
Research elsewhere
Perplexity for citations. Grok for live X/web. Do not make your coding agent pretend it is a search engine.
Visual work
Grok Imagine or ChatGPT images. Then paste a screenshot into the agent (Windows Grok TUI: often Alt+V — not universal).
You, still
Deploy, deletes, auth, billing, secrets — not without a typed yes. See PROJECT.md’s never-list.
Screenshot → other AI → paste the diagnosis back
This is not “ChatGPT is the debugger.” This is “a second mind has not been staring at the same broken file for an hour.”
- Copy the error or screenshot it.
- Open a different product than the one that wrote the code.
- Ask: “Probable cause, in order. What would you try first? Do not rewrite the whole app.”
- Bring that back to the agent in the folder: “Try cause #1. Do not refactor unrelated files.”