I've long said that o3 is the best coding model - but if you're using an agent harness - Claude is just better at navigating your codebase. Enter the @RepoPrompt pair programmer mode - it's the best of both words, as Claude coordinates with o3 to plan and apply edits for you!

Just released @RepoPrompt 1.3.44.- Big update to tool efficiency when managing file selection or fetching codemaps - recursive directory operations now work!.- Massively improved server stability .- apply_edits is now much more useful for non-code usecases too!.

What happens when you didn't plan your agent's work well enough

GPT-5 Pro just wrote me “a deterministic, high-signal fuzz suite” of tests for a utility in @RepoPrompt. Holy shit.

It is insane how much more productive I am with GPT 5 Pro. Reminds me of the uplift I felt when o1 Pro dropped.

Look at this lovely email I just received

RT @steipete: Building a new feature, I let Opus 4.1 pair program with GPT-5 High via @RepoPrompt . Opus directs, GPT-5 implements, Gemini….

None of this applies to gpt 5 pro btw. It is the goat of models.

Gpt5 is arguably a worse vibe coding model than Claude, even if it can write better code. I think if this trend continues, it’s going to be increasingly challenging to prompt these models directly. We’re going to need some intent clarification pipelines.

I think one reason this setup is so powerful with gpt 5, is that it’s so hard to prompt it with short vague instructions. Claude code does great with that though and translates your intent to gpt5 and the @RepoPrompt chat keeps it focused in the provided context.

If you're prompting GPT-5, it will interpret your words LITERALLY. Be specific, or have another model interpret your meaning first.

Demo of the setup (using o3).

Ive been running the same setup - its really powerful!. It takes some getting used to, because you have to learn to prompt it, but it’s insanely powerful having opus check gpt5’s work, and iteratively prompt it through a complex task.

The more I use gpt5 the more I like it.(Over api with medium or high reasoning). It is the most reliable non pro OpenAI model ever released, and with the right prompting it nails your task pretty much every time.

After a while of using Dia, I'm back to firefox. I value privacy more than I value AI features, and I didn't find myself using many AI features in practice.

This is what happens when you don’t optimize the context window of each message sent. You can send 500 messages in @RepoPrompt’s chat and the model will have perfect clarity over all attached files.

It's really cool seeing claude complement GPT-5's work!

This flow feels really overpowered right now. Gpt5 being carefully steered by Claude opus 4.1 is such a powerful combo!.

RT @tszzl: model switcher paradigm will be vindicated in the long run. there is a high switching cost into a very new UX on a useful produc….

RT @PopVerseYT: Gotta praise something over here. @RepoPrompt from @pvncher is I N S A N E. Improved SO MUCH my context usage when workin….

I think OpenAI have to rethink how they bench intelligence because there’s 0 chance that gpt5 nano with some thinking is smarter than 4.1.