Comment by victorbjorklund

When you form a mental model and then write code from that, thats a very lossy transformation. You can write comments and documentation to make it less lossy, but there will be information that is lost to an reviewer, who has to spend great effort to recreate it. If it is unknown how code is supposed to behave, then it becomes physically impossible to verify it for correctness.

This is less a matter of "mindset", but more a general problem of information.

You don't become a good editor, inspector, what have you, by having other people/machines write all the code for you. To become and, perhaps more relevant, to stay a good reviewer, you need to regularly write code from scratch to see how it works. On top of that, languages, frameworks and libraries change constantly and you need to write and execute and experiment with new code to see exactly how it behaves so that you can eventually review the code that uses these features. Good reviewers are not born good reviewers!

I've noticed this too.

I remember hearing somewhere that humans have a limited capacity in terms of number of decisions made in a day, and it seems to fit here: If I'm writing the code myself, I have to make several decisions on every line of code, and that's mentally tiring, so I tend to stop and procrastinate frequently.

If an LLM is handling a lot of the details, then I'm just making higher-level decisions, allowing me to make more progress.

Of course this is totally speculation and theories like this tend to be wrong, but it is at least consistent with how I feel.

Completely agree with you. I was working on the front-end of an application and I prompted Claude the following: "The endpoint /foo/bar is returning the json below ##json goes here##, show this as cards inside the component FooBaz following the existing design system".

In less than 5 minutes Claude created code that: - encapsulated the api call - modeled the api response using Typescript - created a re-usable and responsive ui component for the card (including a load state) - included it in the right part of the page

Even if I typed at 200wpm I couldn't produce that much code from such a simple prompt.

I also had similar experiences/gains refactoring back-end code.

This being said, there are cases in which writing the code yourself is faster than writing a detailed enough prompt, BUT those cases are becoming exception with new LLM iteration. I noticed that after the jump from Claude 3.7 to Claude 4 my prompts can be way less technical.

It seems fair to say that it is ~never the overall bottleneck? Maybe once you figure out what you want, typing speed briefly becomes the bottleneck, but does any expert finish a day thinking "If only I could type twice as fast, I'd have gotten twice as much work done?" That said, I don't think "faster typing" is the only benefit that AI assistance provides.

> How could that possibly be true!?

(I'll assume you're not joking, because your post is ridiculous enough to look like sarcasm.)

The answer is because programmers read code 10 times more (and think about code 100 times more) than they write it.

Scaffolding works fine, for things that are common, and you already have 100x examples on the web. Once you need something more specific, it falls apart and leads to hours of prompting and debugging for something that takes 30 minutes to write from scratch.

Some basic things it fails at:

  * Upgrading the React code-base from Material-UI V4 → V5
  * Implementing a simple header navigation dropdown in HTML/CSS that looks decent and is usable (it kept having bugs with hovering, wrong sizes, padding, responsiveness, duplicated code etc.)
  * Changing anything. About half of the time, it keeps saying "I made those changes", but no changes were made (it happens with all of them, Windsurf, Copilot, etc.).

This can’t be stressed enough: it depends on what you’re doing. Developers talking about whether LLMs are useful are just talking past each other unless they say “useful for React” or “useful for Rust.” I mostly write Drupal code, and the JetBrains LLM autocomplete saves me a few keystrokes, maybe. It’s not amazing. My theory is that there just isn’t much boilerplate Drupal code out there to train on: everything possible gets pushed out of code and into configuration + UI. If I were writing React components I’d be having an entirely different experience.

Isn't there some way to speed up with codegen besides using LLMs?

Isn’t this because the LLMs had like a million+ react tutorials/articles/books/repos to train on?

I mean I try to use them for svelte or vue and it still recommends react snippets sometimes.

"On the scale of an hour" means "within an order of magnitude of one hour", or either "10 minutes to 10 hours" or "0.1 hours to 10 hours" depending on your interpretation, either is fine.

Maybe, but also: Cloudflare is one of like fifteen organizations on the planet writing code like this. The vast majority of The Rest Of Us will just consume code like this, which companies like Cloudflare, Auth0, etc write. That tends to be the nature of highly-specialized highly-domain-specific code. Cloudflare employs those mythical Oauth experts you talk about.

It's not just time spent typing. Figuring out what needs to be typed can be both draining and time consuming. It's often (but not always) much easier to review someone else's solution to the problem than it is to solve it from scratch on your own.

Oddly enough security critical flows are likely to be one of the few exceptions because catching subtle reasoning errors that won't trip any unit tests when reviewing code that you didn't write is extremely difficult.

Time spent typing is statistically 0% of overall time spent in developing/implementing/shipping a feature or product or whatever. There's literally no reason to try to optimize that irrelevant detail.

In my experience the problem is never creating the dashboard view (there's a million examples of it out there anyway to copy/paste), but making sense of the data. Especially if you're doing anything even remotely novel.

If you're making a dashboard view your productivity is zero, making it faster just multiplies zero by a bigger number.

Edit: this comment was more a result of me being in a terrible mood than a true claim. Sorry.

Hundreds of lines that you have to carefully read and understand.