Write the code yourself, let the AI review it

Peter Bloem, an AI researcher and assistant professor, offers a middle path between refusing AI in coding and handing it the keyboard. He calls it craft coding, and the rule is simple: you write all the code by hand, and the AI only reviews it, the way a senior colleague would review a pull request. No autocomplete, no copy-pasting generated code, no pasting your whole repository into a chat. You share a snippet, get feedback, and decide what to act on.

His worry with vibe coding, where the model produces and you accept, is that it manufactures a feeling of understanding you do not actually have, and it quietly deskills you over time. Bloem's own experience is the anchor: he says the code from his PhD ran fine but was buggy in ways he only sees now, and that Claude reliably finds serious issues when he asks it to look. His ten guidelines all point the same way, such as searching and reading the documentation before asking, attempting the problem yourself, running the code first, and only accepting a suggestion you fully understand.

The honest part is where he admits the cost. This approach is slower, and it will not satisfy a company that treats shipping speed as the only metric. He also concedes you lose some deep debugging practice when a reviewer keeps catching things for you.

Why it matters

If you are deciding how your team should use coding models, this is a concrete policy you can adopt tomorrow: write it yourself, let the model review, and reject anything you cannot explain. You keep the security and quality benefit of AI review without the slow erosion of your engineers' skills.