logoalt Hacker News

bikeshaving • today at 7:21 PM • 1 reply • view on HN

I’ve found that the “job” of using AI is mainly foible management. What the author describes here is a habit of coding agents to excessively use and create logging infrastructure, and sloppily leave it in production code even when unnecessary. For instance, I created an IPA IME keyboard (https://ipabet.org/), and I found in review that Claude would leave the debugging tooling around unflagged, which essentially made the program a key-logger.

These foibles come and go: early Claude Code refused to let programs crash, mangled all the comments in your codebase, and then one week you find that it insists on ending all functions with `Of` as in `layoutOf()`, `viewOf()`. Finding and mitigating these annoying habits becomes part of the job. Code review becomes articulating with this entity why certain patterns in the code are silly and should be changed.

The author mentions performance and security issues as the domain of “the architect.” For performance, I’ve found that Claude Code has more or less solved this domain, and can profile and optimize code across a variety of stacks and languages using sophisticated, ad-hoc workflows. For security, it’s just an ever-ending nightmare where random habits and sloppy implementation create random vulernabilities, compounded by the fact that Anthropic has added “safeguard filters” whenever you naively say you’re looking for them. Nevertheless, I’m not sure the architecting role as they describe won’t be automated in the near future.


Replies

hinkley • today at 7:39 PM

I can generally tell which PRs are AI on the OSS projects I do reviews for because the release notes are a Steinbeck novel instead of written as a one line description meant to be read among a litany of others. But then they sort of do that for everything don’t you think?