Personally I've found one of the biggest gains with coding agents is in helping me read code. Actually - that's a lie. I don't read the code. Mostly (unless my spidey-sense goes off) I ask the LLM to read the code and tell me what it does.

And then I make a decision based on that.

I guess I'm wondering if the article is missing half the picture. Yes - AI is wrong some of the time (and that % varies based on a host of variables). But it can read code as well as just write it. And that does matter as it changes the trade-offs this article is weighing up.

It's been pretty great for ramping up into codebases too. "Give me a summary of project in current checkout in markdown form."

There are times when reading the code is necessary, but oh boy are LLMs so much faster at finding the part of the codebase I want to read.

Several projects I work on call into or interact with gigantic codebases. A couple years ago I would have to allocate 30 minutes some times to either trace through the code base or setup and attach a debugger to step through the code until I found the part I needed to know about.

Now I send an LLM to go find it and it comes back with a list of files and line numbers in minutes.

It’s still not perfect. I had a codebase walk this morning where even GPT-5.5 extra high failed to find something I knew was in there on the first pass. It got it with some more directed prompting. If you delegate trust 100% to an LLM you will get bit eventually.

The "It’s harder to read code than to write it" was always silly. The example spolsky gives to support it is basically "devs like to rewrite other devs code, therefore reading is hard" which is obviously bunch of nonsense. That's like saying reading poetry is harder than writing poetry because poets keep writing new poems despite the fact that Shakespeare already wrote it. Now that you can recruit LLM to explain any complicated codebase to you it's even less true.

Exactly, and that's why this maxim about "understanding the code base" being the bottleneck is also somewhat misleading.

Claude is even better at helping you understand the code base then it is at writing code! It can look at a bunch of files and give you an accurate run down in ten minutes.

I like to constrain it as much as possible to ignore variables and function names; the human stuff.

With a new code base my first goal understand how it CRUDs state. What structures and what operations?

Less concerned at the outset if its sorting carrots or processing orders for shirts.

AI seems to work way better for me when I tell it to ignore the use case and focus on surfacing runtime and mathematical operations embedded in the code.

Constraints on new abstraction and telling it to stick to math terms and types and objects also seems to help avoid hallucinations and layers of useless indirection.

It lets you understand the code base at reduced granularity when you want. Or zoom in beyond the written lines to explain _why_ some code is the way it is.