I standardised Claude Code in my team: the part nobody talks about
Originally published in French.
I standardised Claude Code in my team a year ago. We ship about 30% faster.
Here's the part nobody talks about.
1. The bottleneck didn't disappear, it moved.
We were producing code faster, so there was more code to review. My weeks didn't get lighter, they shifted: less writing, much more reviewing. The team didn't get faster at deciding, only at producing. If your review process was already the weak point, AI won't fix it — it will put it under pressure.
2. The conventions worked, not the tool.
For the first three months, everyone prompted their own way and we got three different styles of code back. The gain only came once we wrote down our conventions: how to frame a request, what context to give, what we never delegate. The tool is the same for everyone. The agreement between us is what made the difference.
3. My real worry is what this changes for a junior.
I learned by getting stuck. The three-hour block is where judgment gets built. When the answer arrives in twelve seconds, that step is skipped — and judgment can't be bought. I haven't solved this. I simply ask people to explain the code to me before merging it, and "the AI wrote it" has never been an acceptable answer in review.
The honest nuance: none of this is an argument against these tools. I use them every day and I wouldn't go back. It's an argument that the gain isn't free — it's paid for in review discipline.
The tool multiplies judgment. If there's no judgment, it multiplies the mess.
Have you rolled out AI in a team? What surprised you?