Post #2308277

2026-05-05 05:12 UTC

Well, I tried Claude Code. It&#39;s less bad than GLM, but alas, I didn&#39;t get the &quot;Hallelujah! I see the light&quot; moment that would let me become an AI grifter. Pity, really. I did notice an interesting thing though: the people who get the best results from coding agents tend to be the people who think they&#39;re awful, which says some interesting things about engineering judgement. Read the article here: https://deadsimpletech.com/blog/engineering-judgement-claude-paradox

Replies (2)

@karl@infosec.exchange 2026-05-05 06:07

@iris_meredith@mastodon.social This post matches my anecdotical observations. The only way to wield LLMs with great results is through a thorough understanding of the system you're building. Essentially it boils down to Kernighan's Law, doesn't it? Don't try to get an LLM to write something you could not grasp yourself, it won't be able to fix its shortcomings and nor will you. That being said, deep understanding of systems (at least for me) takes time. Time saved by LLMs is rarely invested in more expertise, it's rushed into the "next thing" that needs to be built; and this is how the foundations of the knowledge of the system become shallower.

Open ##2450972
@delta_vee@cosocial.ca 2026-05-05 10:48

@iris_meredith@mastodon.social "people who currently have good engineering judgement dislike LLM tools" -- that "current" seems key to me, and is probably a whole post in itself

Open ##2450973