Why is Claude writing code that I’m supposed to read?
I wasn’t going to read it line-by-line. I’d skim it, run the tests. So why make AI write in a format optimized for human readers who aren’t reading?
If you aren’t reading your LLM’s code line-by-line, you’re a fucking moron and deserve the AI hate.
Hmm it’s an interesting thought. If you’re satisfied with your tests - with their coverage and their rigor - then really there shouldn’t be a need to read code.
You won’t be able to tell if you should be satisfied with your tests unless you review the code.
The code for the tests, yeah. But not the code for the app itself. If you’ve written tests that you are satisfied fully cover all requirements including edge cases, you can satisfy yourself that the app does what you want it to without looking at the app code.
Test-driven development but with an LLM doing the actual development.
That doesn’t work. TTD works because you’re writing the code you’re testing, so you know exactly what’s in there.
If the LLM hallucinates a feature that wasn’t in the requirements and is somewhat hidden, reviewing the tests won’t cut it. The tests may cover all the requirements, but what about all the code doing stuff that’s not required?
Basically, don’t just test for requirements, also test that what isn’t required isn’t there. And that’s basically an infinite amount of possibilities to test for. The only way you detect that stuff is by checking that the code isn’t doing anything extra it shouldn’t be doing.
really, the first? I think Brainfuck was also not built for humans and has been around for a while…
(this apart from the other obvious things to write about this idea…)
That’s not more readable. Sorry, “a + b” is more readable than “a plus b”.
So me think: why waste time, say lot word when few word do trick?
