Cet article n’est pas encore traduit, il est donc affiché en anglais.
Claude Code Backlash Explained: 12-Hour Days Just to Press Enter?
A viral post said engineers work 12-hour days just to press enter. What Claude Code does, what two Anthropic studies found, and how to keep your skills.
An anonymous software engineer posted one sentence about his job, and millions of people read it: "People are working 12 to 13 hours a day just to press enter."
He was not saying the AI writes bad code. He said almost the opposite. Claude Code was producing the specs, code, tests, tickets and reports, while the humans were pushed to ship faster. "In reality, nobody is thinking anymore."
I went through the reporting on the post, the interview he gave afterwards, and two studies Anthropic published about its own tool.
▶ WATCH ON YOUTUBEWatch the video
▶ WATCH ON YOUTUBEFull video
What went viral
- The account is v0xium, an engineer who says he recently joined a large company. He asked reporters not to use his real name.
- Business Insider reported nearly 5 million views by the Monday. Futurism later put it above 7 million.
- Elon Musk replied with one word: "Yikes."
- Chamath Palihapitiya warned that companies risk turning technical experts into people pushing a button at a slot machine.
- GitLab CEO Bill Staples said the human has become the orchestration layer for the machines.
Keep the weight in mind: this is one person describing one workplace, anonymously.
What Claude Code does
It is not autocomplete. It runs in your terminal and:
- reads the whole repository, not only the open file
- edits files and runs commands
- installs the test framework and runs the tests
- commits and pushes
- can be called from Slack, and runs inside GitHub Actions
When AI only suggests the next line, you are clearly the author. When an agent reads the repo, writes the patch, runs the tests and updates the ticket, your job quietly becomes supervision.
Anthropic measured this. Across about 400,000 sessions:
| Decisions | Made by |
|---|---|
| Planning (what to build) | Users, about 70% |
| Execution (how to build it) | Claude, about 80% |
A typical session was about four prompts from the human and around ten actions from the model.
The real target is "just ship"
He told Business Insider the core problem is leadership's fixation on shipping more features. Management measures sprint cycles, pull requests and feature volume, even when those features do not improve the product.
Here is the trap. If generating code gets ten times faster and reviewing it stays the same speed, review becomes the bottleneck. The dangerous fix is to remove the review.
Six months later, when production breaks at 2 a.m., a person has to know why that code path exists. Accountability has not been automated.
Study 1: does AI cost you skill?
Anthropic ran a controlled trial with 52 mostly junior engineers learning a Python library none of them knew. One group had AI help during the task. Nobody had AI during the quiz afterwards.
| AI group | No-AI group | |
|---|---|---|
| Speed | About 2 minutes faster, not statistically significant | |
| Quiz score | about 50% | about 67% |
The biggest gap was on debugging questions.
The caveats are real: a small study, mostly juniors, a quiz taken right after the task, and a sidebar assistant instead of an autonomous agent. It is not proof that AI makes you worse. It is evidence that finishing a task and understanding it are two different things.
Study 2: 400,000 sessions
Anthropic analysed about 400,000 Claude Code sessions from about 235,000 users over seven months.
- Expertise still decides the outcome. Novice sessions reached a verified success about 15% of the time. Intermediate to expert sessions: 28 to 33%.
- Debugging fell from 33% of sessions to 19% over those months.
This is observational data. It does not prove AI caused skill loss.
The most useful chart
From the trial: how people used the assistant, and what they kept afterwards.
| Habit | Quiz score |
|---|---|
| Let it generate, then make it explain the code | 86% |
| Write some yourself, have it explain the rest | 68% |
| Ask conceptual questions while you build | 65% |
| Delegate the task and move on | 39% |
| Lean on it a little more each time | 35% |
| Paste the error back until it goes away | 24% |
Same tool, same time, roughly sixty points of difference. The tool is not doing this to you. The habit is. And the habit is usually set by the deadline.
What good looks like
- You define the intent. What problem are we solving, and what does done mean?
- The agent does the mechanical work: searching, proposing the change, drafting tests.
- You review the boundaries: architecture, security, data handling, product behaviour.
- Tests verify the output. They do not take the responsibility.
- You can explain the change without asking the model to explain it for you.
If you lead a team, protect the review, the architecture argument, the incident analysis and the debugging. That is where the senior engineer of five years from now comes from.
Verdict
The post does not prove Claude Code is destroying software engineering. It shows what happens when a company optimises for how much software it produces and stops paying for how much anyone understands. The tool decides how fast you go. The process decides whether you are still an engineer at the end.
Sources
- Full video: https://youtu.be/vlZGBuZ-tuQ
- Anthropic, coding-skills trial: https://www.anthropic.com/research/AI-assistance-coding-skills
- Anthropic, Claude Code usage study: https://www.anthropic.com/research/claude-code-expertise
- Futurism: https://futurism.com/artificial-intelligence/software-engineer-ai-chaos-press-enter-all-day
- My own agent test: Claude Code vs OpenAI Codex, 20 tests
Has AI made you a better engineer, or only a faster one?
Quotes from X are attributed to the reporting above. The video script was researched and drafted with help from Claude.


