Fresh-context review, or why the author is the worst judge

The person who wrote a thing is structurally the worst person to check it, and this is doubly true when "the person" is a model. Both the model that generated the code and you, who steered it there, arrive at the review carrying the exact same assumptions that produced the code in the first place. You can't catch a blind spot with the eyes that made it. The fix is to get eyes that don't share the blind spot.

This is why the frontier teams lean so hard on review from clean context. Anthropic's coding guidance is built around handing work to a fresh reviewer that didn't write it, precisely because the author, human or model, reads what they meant to write instead of what's actually on the page. A reviewer coming in cold has to read what's there, because they have nothing else. They don't know what you intended, so they can't accidentally supply it, which means the gaps between intent and implementation, invisible to you, are the first things they see.

Why does clean context catch what the author misses?

Because so much of what makes code look correct to its author is context the author is silently supplying. You read the function and it makes sense, because you're filling every ambiguity with the intent that's still in your head. The model does the same in reverse: it generated tokens that cohere with the tokens before them, so of course they look consistent to it. A fresh reviewer has none of that supplied context. They see the function as it actually reads to someone who wasn't there, which is exactly the audience that will eventually run it in production.

How to actually get it

Two moves. Split writer from reviewer even when both are AI: have one model generate and a different instance, with clean context, review, because the reviewer that didn't write it has no assumptions to protect. And for the things that matter, get a human read too, because a human catches the class of problem that only matters because it matters to humans, the ones no assertion encodes. Neither replaces the other. The fresh AI reviewer catches the mechanical gaps; the human catches the ones that are only visible from outside the machine entirely.

The takeaway: the author shares the blind spot that made the bug, so verification needs eyes with clean context, human and AI both. Never let the thing that wrote it be the only thing that judges it.