Average Code Is Still Yours

LLM-generated code is usually good enough.

That is the useful part. It works, the tests pass, the page loads, the API responds, and you get to keep moving.

It is also the dangerous part.

Good enough code can look like finished code, especially when you are tired, missing context, or trying to unblock five other things. The temptation is to accept the patch, maybe skim the diff, and let momentum do the rest.

But the model does not know your real constraints by default. It does not know the expected scale, the failure modes you care about, the production bar, the tradeoffs you are willing to accept, or the parts of the system that have history.

Harnesses help. Smaller tasks help. Guardrails help. Better context helps. None of that changes the basic responsibility: the generated code is not accountable for the decision to ship it.

You are.

Every time you accept code from an LLM without evaluating it, you are quietly accepting average as your output.

It is your name in the commit.

Are you okay with that?