Ankur Sethi
Writer, independent frontend developer, computer gremlin. Long-form writing on https://ankursethi.com.
I run https://blr.indiewebclub.org with https://abhinavsarkar.net and https://tanvibhakta.in.
i'm not thrilled about using LLMs to generate code. i mean, it's alright? if you're very careful and know what you're doing? but IMO tools like Claude Code are dangerous if used irresponsibly.
on the other hand, i'm much more excited about using LLMs to *improve* human-written code. especially if you have the expertise to teach the LLM what bugs, antipatterns, etc. to look for.
https://ankursethi.com/links/mythos-finds-a-curl-vulnerability/
this is such a great use for LLMs! i can fully get behind this.
the interesting part is that you have to have a ton of domain knowledge to pull this off. you can't vibecode your way into this plugin.
https://ankursethi.com/links/using-llms-to-find-python-c-extension-bugs/
did a little experiment over the past four weeks. i have to admit that coding agents are incredibly effective, even when you go full vibeslop mode.
but there's still something deeply uncomfortable about all this.
https://ankursethi.com/blog/programming-language-claude-code/