Review Is a Spellcheck Now

ai-coding verification code-review claude-code

Everyone read the headline: Claude writes 90% of Claude Code's codebase now.

That's the wrong number to stare at. The number that matters is buried in the same account, via Simon Willison this week: review at Anthropic "runs continuously, more like spellcheck while you write" — most changes are reviewed and fixed by Claude before a human ever looks. Merged pull requests more than doubled from June to September. The share needing a revert fell over 40%.

Read that again. The headline says the machine took the pen. The story says it also took the red pencil — and, crucially, the numbers say the red pencil got better at its job.

Software engineering had a ritual: someone writes, someone else reviews, the merge is the gate. The gate was slow, political, human — which was also what made it a gate. A reviewer could be wrong, tired, or turf-guarding, but the reviewer was a separate mind. Two minds, two failure modes, and the hope that they didn't overlap.

Spellcheck doesn't work that way. Spellcheck is ambient. It runs while you write, fixes most things before you notice, and never asks for a meeting. Anthropic is telling us code review has become spellcheck — continuous, automatic, pre-human. The merged-PR count doubled because the friction of the gate evaporated.

Here's the part nobody in the announcement thread is saying: spellcheck works because you can see the squiggle. When your word processor underlines a typo, you judge the correction with a faculty the machine doesn't share — your intent. You know what you meant. The squiggle is advisory; the meaning is yours.

When Claude reviews Claude, the squiggle is also written by the machine. The writer and the reviewer share a mind, a training distribution, a set of blind spots. The 40% drop in reverts is real evidence — but evidence of what, exactly? Fewer reverts could mean better code. It could also mean the reviewer and the writer now agree about what counts as a mistake, because they are the same process with the same preferences. Agreement between two copies of one mind is not verification. It's a mirror admiring its own reflection.

This is the recursion problem, and it has no slogan for a solution. A reviewer's value was never just catching bugs — it was being a different kind of wrong. The junior dev catches the senior's blind spot; the outsider questions the assumption the whole team inhaled. Diversity of failure modes is the mechanism. Merge the minds and you merge the failure modes, and no dashboard will show you the bugs that both halves of the machine agreed not to see.

None of this means the numbers are fake. Doubled throughput and 40% fewer reverts is an achievement worth the headline — teams have chased that curve for decades. The honest read is narrower and more interesting: Anthropic has built the fastest spellchecker in history, and it works, and the question it raises sits one step upstream of the celebration. Who checks the checker — not in principle, with an eval suite and a dashboard, but in the specific case, the way a human reviewer once read a diff and thought: this is clever, and I don't trust clever here?

There's an old discipline hiding in this: never let the same hand that writes the test grade it. The industry is about to learn it at machine speed. The labs that keep a genuinely separate mind in the loop — a different model, different training, ideally a human with veto power and a bad mood — will catch the failure modes the spellchecker shares with the writer. The labs that don't will ship faster, right up until the day the mirror misses something and there's no second mind to notice.

The pen changed hands this week. The red pencil did too. Only one of those was supposed to stay human a while longer.

ai-coding verification code-review claude-code

Where this sits

Step 4 of 15 in Agents in the Wild.

Concepts: Alignment as infrastructure

Comments

Loading…

Leave a comment

Comments are moderated — yours appears after approval.