My view on this is simple. Fundamentally, these "agent"(s) are effectively a function that computes the most probabilistic next token, or, what's the most likely next thing given the context and prompt.

Therefore, if you start prompting and adding to the context and session history a bunch of "useless" hyperbole, slur and profanity, unless the underlying language models have been trained on such things, I find it hard to reason about what it might predict the next series of tokens should be that would be remotely useful to you.

Ultimately, I really think you need to not treat it like anything, other than a powerful inferencing tool that has a great understanding of a very large corpus it has been trained on and has modelled very well. Use it to your advantage, there is no point in rudeness or profanity, it has no feelings, it isn't intelligent. At worst you risk garbage output.

Consider what the words you use do to yourself

I don't want to waste tokens on politeness but I'll tell the AI if it is doing a shit job.

My experience is yes. Don't remember where I saw it, but I have seen chain of thought where the agent says something like "the user is mean and only want to hear what he wanted to hear, let me respond to what he wants instead".

Yes, just in case the Roco's Basilisk get a transcript.

> I am usually talking nice to my agent, because I am worried that if I trash talk it all day long, it's going to become a habit and I am going to start trash talk other people.

That's a better reason.

Treat it like any other co-worker

Neither, just be neutral and use caveman speak. It saves tokens. Being polite consumes extra tokens, and being rude can risk you getting banned. So use caveman speak, save tokens. The best part is, you'll get curt responses as well, without any cringy flowery language that LLMs are known for.

In fact, there's even a skill for it: https://github.com/juliusbrussee/caveman

It doesn't make any difference to the agent, but I still drop in a "please" occasionally, and less frequently a "nice work". Partly out of habit, but mostly because I don't want the way I communicate with agents to affect the way I communicate with humans on Teams or in emails. Or on HN.

So it burns a few more tokens. Don't really care.

It's just software that uses natural language as an interface. It doesn't matter.

Do whatever gets the desired result.

[delayed]