Experienced devs: when did you stop reviewing everything?
retained text
I've been a lead dev for 8 years. I follow the evolution of AI closely: I started with ChatGPT, then Copilot, then Claude Code. Today, I use Claude Code through worktree-oriented IDEs. I don't switch models with every update. My workflow suits me, so I stay. I've also upskilled my team on it. For a long time, I had a rule: never merge code without human review. But the more I master these tools, the more I parallelize. I often run two projects at the same time. Beyond that, I struggle. And I see one thing: the bottleneck is me. AI produces faster than I can review. Concrete example: I'm porting an unmaintained PHP library to a Symfony bundle. A bit of context, in case: Symfony is a major PHP framework, especially popular in Europe, where it's often the go-to full-stack framework — in the US it has less mindshare among PHP devs, though its components power tools like Laravel and Drupal. At first, it was an experiment. The further it goes, the more I see that it works. The bundle is now much better than the original library. I didn't review the port. But it's not "vibe coding" without a safety net: * static analysis and lint; * tests (also AI-generated, I should point out); * architecture and development rules, designed so a human can get up to speed quickly; * cross-reviews by other models; * functional tests of the whole bundle; * a maintained JS library that does the same thing as the PHP lib also served as a reference for me. For a long time, I was skeptical of large projects that migrate 100% to another stack with AI, like Bun moving to Rust. I find myself facing the same problem, on a much smaller scale. It's not critical code. I don't think there's a big security risk. And yet, I want to use it as is: now that I have my Symfony bundle, I can iterate, add features, and going back is complicated. So I'm in conflict with my initial rule. I don't think I'm the only one in this situation. Fabien Potencier, the creator of Symfony, has just released Symfony Language Tools, Symfony's official LSP server. He [explains](https://symfony.com/blog/announcing-symfony-language-tools-the-official-symfony-lsp-server#written-with-ai) that the majority of the code was written and reviewed by AI models. He says he keeps control over the architecture and the final decision on every change. He also insists on the test suite, and presents the project as an experimental beta. This echoes my questioning: where do we place review when AI writes and reviews a large part of the code? I'm looking for concrete experience reports. On a side note: I find Opus 5.5 to be a real leap. People often say we've "solved software", but I think we're not far from it. I think a developer who knows what they're doing, knows their stack well, and frames AI well... it's a really huge productivity boost. My question: have you crossed that threshold? If so, how, and with what safeguards? If not, why? I'd like to hear from experienced developers, pre-2022. And I want to avoid useless debates about vibe coding. I'm curious to hear your feedback.
document fetched 2026-10-06 00:22 · canonical https://reddit.com/r/ClaudeAI/comments/1wyfopv/experienced_devs_when_did_you_stop_reviewing