If Relearning Software Engineering, First Learn to Spot Which AI Code Is Wrong

Kent C. Dodds posted a tweet, two sentences: everything has changed. If I were starting over from scratch in software engineering now, this is what I would do. Attached below is a video titled If I Started Over From Scratch. In the video frame, a person sits on a purple sofa, drawing in a notebook with a pencil, and the sentence is written above.

Kent has long taught front-end development and is the author of Testing Library; many developers have heard his name. Ordinary readers can think of him as someone who has been in software engineering education for a long time. This tweet places the question on the learning path for beginners; tool updates are just background.

In the comments, some netizens put it more concretely. seridarivus said that starting from scratch now means first learning to read code and review code, and then writing large amounts of code. AI agents will produce a lot of content; what is scarce is knowing which outputs are wrong and why they are wrong. Here, 'agent' refers to an AI programming assistant that can generate code, modify files, and run commands. For ordinary readers, you can think of it as an intern who types very fast but needs human acceptance.

This statement is somewhat counterintuitive. In the past, when learning programming, the default path was to write more. Now AI can quickly generate implementations, so the barrier to writing code has dropped, while the barrier to reviewing code has become more visible. Reading diffs, looking at test failures, tracing dependencies, writing minimal reproductions, and judging boundary conditions—these actions are not cool, but they determine whether you can turn AI output into maintainable engineering.

A netizen named Yash asked: I graduate next year; would Kent still recommend that beginners first go deep into one tech stack, or now that agents can fill in many gaps, should they be broader? There is no unified answer to this question. But in the context of reviewing and debugging, going deep into one stack first is still useful. You need a context that can run before you know where failures come from. If you are only broad and not deep, it's easy to stop at the level of clicking accept for the agent. If you are only deep and not broad, you'll be slow when encountering cross-stack problems. For beginners, deep first and broad later is more stable; the cost is that early on you may not look comprehensive enough.

Carlo's reply is very real: the first sentence 'everything has changed' is too stressful; I need a cup of coffee first. This reaction shows that change is pressure for people just starting out, not a slogan. EMCP Tools added: learn debugging. The agent writes, but you still have to fix it. This sentence is more direct than many learning roadmaps.

If starting over, the beginner tasks could be rearranged: read code, review code, run tests, locate failures, and then make small modifications. Writing quickly is no longer scarce; spotting where it's wrong, why it's wrong, and how to fix it is still scarce. Kent's video title is 'From Scratch,' and the comment section concretizes 'from scratch' into these actions. Original tweet link: https://x.com/kentcdodds/status/2117683366619443388