[jedbrooke]

> Some small part of me has started to dread reading LLM output because I know what I’m going to find. False assumptions and hallucinations. Emphatic, staccato fragments. Excessive emojis :sparkle:. :rocket: It’s not just me—these are real patterns (:barf:).

It takes like 5 seconds to add a quick CLAUDE.md / AGENTS.md with a quick style guide, or even just “EMOJI ARE FORBIDDEN” and I find it makes llm output significantly more tolerable. That and a quick style guide with some banned words and phrases.

That won’t help with the false assumptions though, gotta use old fashioned careful reading and critical thinking the catch those.

[vaylian]

You are completely right! Do you want me to write these markdown files for you?

</irony>

[stasomatic]

And...

Me: "The gap, stated plainly:" stop using the type of language.

Claude: "You're right. That's one of the constructions your preferences told me to drop, and I used it anyway."

[jedbrooke]

I find sonnet and haiku will do that, but I rarely get that with opus. When it does happen it’s a good cue to start a new session. That’s another benefit of the style guide is it’s a good canary for when the model has gone off the rails. If it’s left the style guide, the rest of the response can likely be discarded as departed from reality.

[zahlman]

FWIW, I've been finding that ChatGPT doesn't use emoji at all when I engage with it like a pair programmer and bounce off design ideas, ask for implementation code, propose refactorings etc.

But when I ask it to do data analysis or modeling, the emoji are all over the place, yes.

(And judging by what I've seen on GitHub over the last year or so, I would never in a million years consider asking an LLM to write a project README or documentation unsupervised.)