Post by @swelljoe@mas.to

2026-04-17 05:42 UTC

@tokyo_0 @yaelwrites it is sycophantic always, though you can configure it to be less so. I choose configuration options that minimize it (stuff like "professional tone" settings), and I explicitly disallow it in my CLAUDE.md and demand very concise/precise prose. But, it still defers to the human's greater wisdom and dives right in with gusto, no matter how dumb the idea is. They don't know how to say no. They'll change the definition of the problem in order to deliver something that works.

Replies (2)

@swelljoe@mas.to 2026-04-17 05:46

@tokyo_0 @yaelwrites code is different. And, it's the one thing I think LLMs can actually do effectively. LLMs have no taste or desire, so they can't write or make art for shit, but if there is a clear success criteria, as there can be when writing code, they have become astonishingly effective. In the hands of someone who kind of understands what's going on and what is possible and what a good solution would look like, they can deliver working code very quickly.

Open ##1817208
@tschenkel@mathstodon.xyz 2026-04-17 07:45

@swelljoe @tokyo_0 @yaelwrites I'm not certain those extra prompts address the underlying problem. It may only take out the sycophantic language, but still deliver the same confirming content (at least it seems to do so in the limited testing I've done). If that's the case, these prompts make it worse, since they make it more difficult for you to detect and identify. Both consciously, but more important, subconsciously.

Open ##1817214