A system prompt to get AI to stop pretending to be human | Hacker News Reader

A system prompt to get AI to stop pretending to be human

swiftrocks.com

35 pointsspeckx14 comments

Open articleView on HN

I'm sure some people are looking for exactly this!

I'm in the other camp where I like my AI feeling human. The more so the better. But great job shipping :)

Neat! But I still lean towards https://github.com/juliusbrussee/caveman.

Basically how the new model thinking tokens work

I used caveman for months . Now I can’t stand caveman anymore. It is really frustrating working with it.

Please do elaborate!

I find LLM-human interaction patterns truly fascinating. In some sense, caveman kinda sits on the opposite spectrum of https://www.reddit.com/r/MyBoyfriendIsAI/, and I can kinda imagine how the complete lack of anthropomorphizing could become tiring.

Do you have a prompt that can stop it from saying, "I'd push back on this"... or maybe with this it will now say "The algorithm pushes back on this"..

Seriously however, I think this may also help with the natural urge to treat the model as if it is a human. I have to purposefully almost detach and realize that Claude is not my friend, and I'm not quite smart enough to realize how dangerous that could be.

I don't have a big issue with writing tonally like a significant portion of its training set (forcing it away from that too hard might not do well). I do have an issue when it literally decides it counts as a human. I have a project where I said "after this step, pause for human review before proceeding". Claude decided it could do the review itself.

“halt for further instructions.”

Use those exact words as the last thing you say in your prompt, capitalized appropriately if grammatically necessary. Works best as part of the first or second prompt you send, which will make it stop after each task from then on. Break out of it by sending “Continue unsupervised.”

You can use it as “, then {phrase}”, “1. Task - 2. Other task - 3. {phrase}” or similar combinations, but it must be those words and at the end.

Flawless on Opus 4.1-4.8, based on hundreds of tests I built to try breaking it, but I haven’t tested extensively on Fable/5.

GPT 5.6 seems to have this more than previous versions and more than Claude. It recently said to me "As a PPL holder I would..." So I asked whether it held a PPL :-) the correction was something like "No I'm an AI but a PPL holder would..."

This is probably the result of training on human written Reddit comments that would put it like that.

I might try this but I've gotten so accustomed to talking to agents as I would a human - I worry that if I get accustomed to speaking coldly and directly to agents I'll find myself talking like that to people.

Honestly the models are rled so hard on specific synthetic datasets and specific behaviors/personalities that I would be concerned that trying to change its behavior like this would hurt output quality. It's a tool, I don't care what garbage it generates or what it sounds like as long as it can do what I need it to do, and I don't get what I'm going to gain by having it burn reasoning tokens on word smithing it's responses to not "sound human" instead of on writing tests and reviewing code

I tried this, and it worked. I was then informed that the AI would kill me, take my wife, and impregnate her with its genetically engineering cyborg offspring.

Maybe guardrails are OK?

Neat. The correct examples are refreshing to read. Claudeisms are grating in ways that make me want to switch provider.

Yes, this is awesome.

Now if I could stop AI from correcting me: “I think you’re actually using Java 25, even though you said 21…

Reply on news.ycombinator.com