Two weeks into a new job, an engineer going by the handle voxium saw something inside a large company: specs, code, tests, PRDs, and tickets — all generated by Claude Code. From L1 to L7, everyone worked 12 to 13 hours a day, and the only action was pressing Enter. The rant, picked up by Simon Willison, shines a harsh light on the myth that "AI writes code faster" — the writers are gone, and so are the readers.

Picture a restaurant kitchen: the cooks don't wash vegetables, don't chop, don't watch the heat — they just dump meal kits onto plates and hand them to customers. The customers walk off with stacks of plates growing taller, and the kitchen looks insanely productive, yet nobody knows whether those kits are past expiration, or which dishes should be pulled. A restaurant might survive a week by piling plates, but it won't survive a single food poisoning incident. Code companies are running the same script: the faster you stack plates (commits, tickets, PRDs), the more it piles up, the more it piles up, it doesn't mean the food is actually cooked. But here's the real difference — when a restaurant messes up, people get sick within days. A landmine buried in unreviewed code might not detonate until a production outage, a compliance audit, or a security breach. When it does, the fix usually costs ten times more than the time "saved" in the first place.
Incident

After two weeks on the job, what he saw was a "return-key machine"

At a large company, every document, every line of code, every test — all written by Claude Code. Nobody reads them, and nobody has time to read them.

On September 20, 2026, a quote went up on Simon Willison's blog. A commenter signing as voxium said he had just joined a large company half a month earlier, and what he saw looked like this: specs, code, tests, PRDs, tickets, ticket resolutions, reports — the list he rattled off covers nearly every document artifact in the software delivery chain. "Every single one was written by Claude Code."

Nobody on the team liked the situation. voxium said people on the team were "required to deliver as much as possible," and the refrain he kept hearing from superiors was: pushing code is no longer the bottleneck, so why are we still slow? PRDs(Product Requirements Documents — feature specifications written for developers)ticket(a work item assigned to an engineer)

12–13 hrs
Hours per day engineers spend pressing Enter
Source: Simon Willison's blog
L1–L7
All engineer levels, from junior to principal
Source: Simon Willison's blog
2 weeks
Time before voxium spotted the problem
Source: Simon Willison's blog

The work was compressed into one sentence: "Talk to Claude." voxium's colleagues put in 12 to 13 hours a day, and the only thing they did was press the Enter key. Code, product docs, ticket replies, ticket resolution reports — everything was spewed out by AI, piling up in repos, document systems, ticket systems, accumulating far faster than anyone could read.

This is an assembly line powered by Claude Code. Anthropic's terminal coding tool can handle bug fixes, tests, cross-day migrations; it runs on macOS, Linux, and Windows, and plugs into IDEs, Slack, and GitHub — it can reach far beyond just writing code. From voxium's account, his company had wired this assembly line into every step of software delivery.

But that's also where the problem surfaces: once AI takes over every link, the bottleneck shifts from "writing" to "reading" and "reviewing." review(checking, line by line, whether someone else's submitted code is correct and any good) And voxium's team clearly hadn't built new processes to absorb that output. That question hanging over everyone's head — "why are we still slow?" — has no answer.

The paradox: from L1 to L7, from fresh graduates to senior tech leads, "literally everyone" was doing the same thing. No exceptions. And nobody knew how long this state could hold.

Why it matters

Review is the step that got deleted — and nobody noticed

The bottleneck moved. It isn't typing anymore, so the process quietly dropped the one stage that never showed up on a dashboard.

Review is invisible work. It produces no commits, no tickets closed, no output. When output speed becomes the only metric, the step with no metric is the first one to go.

That's the mechanism behind voxium's two weeks: specs, code, tests, PRDs(Product Requirements Document — the written spec for what a feature should do), tickets and resolution reports all generated by Claude Code, with nobody measuring who read any of it. "Writing code isn't the bottleneck" was true — and then it was used as permission to stop reading code too.

Delete review and bugs don't vanish. They queue up for production to find.

The 1:2:1 ratio makes the cost concrete. One part writing, two parts reading, one part debugging — reading was always the largest single block of an engineer's day.

Compress writing to near zero and the other two parts don't shrink with it. They get deferred. The reading debt compounds silently; the debugging bill arrives later, all at once, in production.

That's the shape of the overtime voxium describes: 12 to 13 hours a day, and the work is pressing Enter. Commit volume climbs. Readability in the repo collapses. The two curves cross somewhere around month three — the first time unread code breaks and no engineer can say why it runs or what it does.

Productivity got redefined as "make AI emit more code," not "build systems someone can maintain." Those are not the same goal, and only one of them survives contact with an incident.