You should use AI to work slower
I saw this post by Nolan Lawson linked off Hacker News the other day titled “Using AI to write better code more slowly”. https://news.ycombinator.com/item?id=48272984, https://nolanlawson.com/2026/05/25/using-ai-to-write-better-code-more-slowly/
The author suggests using AI to identify and resolve bugs systematically, a workflow that he says can slow his overall velocity, but ultimately leads to higher quality code. Lawson compares this with using AI as a “slop cannon”, to produce large amounts of low-quality code. It’s perhaps a strawman, but quickly producing large amounts of low-quality code is exactly what I’ve encountered when I’ve used AI quickly. Maybe it’s OK for a throwaway script, but even there, from a personal development perspective, wouldn’t it be better for me to understand it?
Vibe coding proponents might argue that we’ve raised the level of abstraction and the average developer should care as little about the code as they do about the assembly their compiler produces. But there is something fundamentally different: compilers are deterministic and verified. They’ll argue that although LLM’s are non-deterministic, this can be compensated with good validation testing. But in this case the validation tests themselves need to be written with real understanding, if those tests are also vibe-coded, the whole approach is circular. What’s more, there’s no adequate validation test for code maintainability itself, and it’s evident that injudicious use of AI can lead to a project that collapses under the weight of its code bloat.
Still, there may be a time and place for vibe coding. In the meantime, I’d like to draw out Lawson’s argument and suggest that developers should be using AI to help them understand, on a line-by-line basis, the codebases they do care about.
Artificial intelligence is a great tool, but a tool is only as good as the judgment of the person using it, and how can anyone claim to have good judgment over something they don’t really understand?
So before using AI to rip everything apart and produce big hairy balls of new code, make sure you understand what you’re working on and producing. Use AI, if you like, to help you understand the code. Find something you don’t understand, highlight it, ask your agent what it does, and apply your common sense and critical thinking to its reply. Cross reference and skim against the documentation, and repeat. This will be slower, it may even feel less productive, but you’ll be developing your understanding, your expertise, and your insight, which will ultimately lead to better software.