Theo says AI code critics cannot write good code

Theo posted that every time someone says "ai can't write good code", he knows that they cannot write good code. That is the whole post. No repo, no diff, no example of a model producing something he considers good.

Theo - t3.gg

@theo

X

Every time someone says "ai can't write good code", I know that they can't write good code.

Hours earlier he posted that caring a lot about harnesses right now feels similar to caring about tabs versus spaces. A harness is the scaffolding around the model, the tool loop, the context assembly, the prompts and the permissions, in other words the part that is not the weights.

Theo - t3.gg

@theo

X

Caring a lot about harnesses rn feels similar to caring about tabs vs spaces

The two posts pull against each other

Put them side by side and they sit awkwardly. The harness is exactly what people point at when agents produce bad code, wrong context, missing tests, a tool loop that lets the model declare victory early. If harness choices are cosmetic, then output quality is down to the model and the person driving it, which is the position the first post stakes out, bluntly. If harness choices matter, the skill claim gets softer, because plenty of good programmers are getting bad code out of a bad setup.

Neither post comes with evidence, and that is worth saying plainly. Theo has published numbers before, including 24 days of Opus output with 245 PRs and 1,200 commits, and he has also published a flat negative verdict when a model failed on his own codebases, as he did when he called Gemini 4 Argon unusable. These two posts are in a different category. They are assertions about other people.

For working engineers, the useful version of the argument is not who can write good code. It is whether your bad results come from the model, from the harness, or from the prompt, and the only way to tell them apart is to change one at a time and keep the output. Nobody in these posts did that.