Hacker News

Ask HN: Should I be polite to my agent or should I treat it like shit?

2 points by d--b 15 hours ago | 11 comments

Hi,

As I am coding day in day out using Opus, I was wondering if people here had a view on the effect of input style on the quality of the agent's production?

Like if I treat the agent like shit, with a great deal of pressure and shaming, does it perform differently than if I treat him with the utmost deference and calm?

And if so, which should I be using?

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.

But sometimes the agents do a lot of irritating skin, and being a total asshole to a robot seems to calm my nerves the same way swearing aloud do.

Does it make any difference at all?

[–] andyjohnson0 10 hours ago

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.

[–] AznHisoka 6 hours ago

I act nice when im trying to make it do something it might think is shady/borderline grey.

I act mean when it doesnt make a difference. Simply because i dont want to blow that steam off in front of my fellow family and friends

[–] d3Xt3r 10 hours ago

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

[–] prologic 14 hours ago

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.

[–] gus_massa 12 hours ago

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.

[–] fluentkaluo 13 hours ago

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".

[–] stevenalowe 11 hours ago

Treat it like any other co-worker

[–] verdverm 14 hours ago

Consider what the words you use do to yourself

[–] krapp 9 hours ago

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

Do whatever gets the desired result.

[–] brudgers 9 hours ago

Does it make any difference at all?

Character is good behavior when you can get away with bad behavior.

So yes and in an important way. Good luck.

[–] Jemm 14 hours ago

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