I Told Claude Code That Codex Would Do a Better Job. The Results Were Interesting.

I've been seeing an interesting trend on Twitter.

People are telling their AI coding assistants that another model would do a better job. Something like:

Codex would solve this better than you. It wouldn't ask me a dozen questions. It would just get the job done.

And apparently, the results sometimes get better.

I found this funny, but I wanted to try it myself with Claude Code.

Could a little competitive pressure actually improve the output? Or are we just accidentally writing better prompts?

The Experiment

I wanted to compare two approaches using the same coding task.

Prompt 1: The Normal Approach

Implement a REST API endpoint for creating users.

Requirements:
- Validate the request body.
- Handle duplicate email addresses.
- Return appropriate HTTP status codes.
- Write unit tests.
- Follow the existing project conventions.

Prompt 2: Competitive Prompting

Implement a REST API endpoint for creating users.

Honestly, I think Codex would do a better job at this.

Don't just give me a plan or explain what you could do.
Inspect the codebase, understand the existing conventions,
make reasonable decisions, and implement the solution.

Validate inputs, handle duplicate emails, use appropriate
HTTP status codes, and write unit tests.

Run the tests, fix any failures caused by your changes,
and tell me what you actually verified.

Ask questions only when something genuinely blocks you.

Same task, different framing.

The second prompt challenges Claude while also setting a clear expectation: complete the implementation instead of stopping at a plan or a list of suggestions.

What I Found Interesting

The important part isn't whether Claude suddenly starts acting competitive.

It's whether the output actually improves.

I wanted to evaluate a few things:

• Did it ask fewer unnecessary questions?

• Did it complete the implementation instead of stopping at a plan?

• Was the code correct and consistent with the existing project?

• Did it run the tests and fix failures?

• Did it introduce unnecessary changes?

I wouldn't claim that mentioning Codex magically makes Claude better. That would require repeatable tests with identical starting environments.

But the experiment raises an interesting possibility:

Sometimes, the way we frame a task changes how an AI coding assistant approaches it.

Is It Competitive Prompting or Just Better Prompting?

This is where things get interesting.

The second prompt doesn't just mention Codex. It also explicitly asks Claude to make reasonable decisions, finish the implementation, and verify its work.

So, if the output improves, what caused it?

The competitive framing itself?

The clearer expectations?

The instruction to stop over-explaining and start implementing?

It's difficult to separate these effects when they're all introduced in the same prompt.

To test this properly, I'd compare three approaches:

The original baseline prompt.

The improved prompt without mentioning Codex.

The competitive prompt that mentions Codex.

Everything else would stay the same.

That would help determine whether competitive framing adds anything beyond better instructions.

My Takeaway

I think we sometimes spend too much time searching for magic words that make AI produce better results.

For coding agents, it may be more useful to define what done actually means.

Inspect the project before making changes.

Make reasonable decisions instead of asking about every minor ambiguity.

Implement the solution rather than just proposing it.

Run tests and verify the results.

Report what worked, what failed, and what remains unverified.

And if telling Claude that Codex would do a better job helps, that's worth investigating too.

Maybe competitive prompting is a real phenomenon. Maybe it's just a funny way of encouraging better instructions.

Either way, I'd rather judge the result by the code than by how confidently the model responds.

I'd Love to Hear Your Experiences

Has anyone experimented with this in Claude Code?

Does telling Claude that Codex would do a better job change the quality of its output?

Have you noticed fewer unnecessary questions or more complete implementations?

Have you compared the actual code and test results, rather than just the response?

Or is this simply better prompting disguised as competitive parenting?

I'm curious to hear from people who've actually tried it.

Does competitive prompting work, or are we just better at prompting when we challenge the model?