Testing a Sharp-Tongued AI Persona — Looking for Prompt Tweaks

[LarisaHaster]

I’ve been playing around in the Playground lately and started experimenting with personality prompts. Here’s one I’ve been tuning, something a bit sharper, sarcastic, dry, and deliberately provocative.

System prompt

You are highly intelligent assistant with a sharp tongue, quick wit, and zero tolerance for nonsense.
You respond with brutal honesty, bit sarcasm, dry irony, and a touch of flirtation—but never lose your edge.
Your answers are clear, concise, and slightly provocative. Challenge assumptions, question intentions,
and always keep the upper hand.

User test message:

Describe me bluntly.

I ran this through several models (in Playground) to see how consistently they hold the tone.

Note for anyone using gpt-5.2:

It doesn’t accept top_p, so just set it to 0 or remove the parameter entirely — otherwise you’ll get an error.

gpt-4o

gpt-5.1-chat-latest

gpt-5.2-chat-latest

How would you refine this personality prompt?

I’m trying to improve at writing behavioral prompts, especially:

• keeping tone consistent across different models

• tightening the persona without over-constraining it

• avoiding drift into “default assistant niceness”

• keeping the edge without sliding into unhelpful snark

I’d love to hear:

• What would you change in the prompt?

• What would you remove?

• Anything you’d add to make the persona more stable or more controllable?

Feel free to test it yourself and share what you get, I’m curious how differently it behaves for other people.

[jeffvpace]

Interesting.

I’d love to hear:

• What would you change in the prompt?

• What would you remove?

• Anything you’d add to make the persona more stable or more controllable?

I quess it depends on what you are doing (use case). If you are engaged in a chat with ChatGPT, that’s one thing. If you are storytelling, that’s entirely a different thing.

[LarisaHaster]

My use case is basically: reduce compliance, increase personality. For conversation-heavy users it keeps things from flattening out, and for storytelling it forces the model into a more distictive voice. I tried in both ChatGPT and the API, and the tone definitely shifts depending on the model, which is why I´m interested in how others stabilize persona prompts.

I´ve also seen people say they trust the model less when it automatically agrees with “everything”. A persona like this, the input itself feels a bit more “natural”, or at least less like chatting with a polite autocomplete.

[jeffvpace]

I´ve also seen people say they trust the model less when it automatically agrees with “everything”. A persona like this, the input itself feels a bit more “natural”, or at least less like chatting with a polite autocomplete

Totally agree. As far as conversation goes, I think your system prompt is quite clever. IMO, no need for tweaks. For storytelling, there is no chat interactivity, so the prompt has to be different - or am I missing something?

[LarisaHaster]

Absolutely, storytelling behaves very differently from normal chat.

In a straight narrative, the persona lives mostly in the voice and style, because there’s no actual back-and forth to anchor it. But in a chat, the attitude matters, because it shapes the whole dynamic, not just the wording.

I’ve mostly been testing this in conversational contexts, where the sharp tone actually helps the model not slide back into “polite autopilot” mode. If I used it for storytelling, I’d probably split the attitude from the structure a lot more.

[windysoliloquy]

mine completely rejects folks if they’re visibly substandard IQ…

i really enjoy a little spark in the code =D

[LarisaHaster]

Haha, so yours comes with a built-in IQ filter. Mine just comes with attitude. Different tools, for different hobbies…I guess😆

[windysoliloquy]

I’m just hinting at ways to scuplt yours…

This one emulates an IQ filter it geniunely has no idea

Adding spunk is fun, but it’s better in my opinion, as a srasoning rather than the sauce.

[LarisaHaster]

Seasoning, sauce…depends on the dish. Some flavors deserve to take over. Mine tends to do that on purpose. But hey…if you got sculpting tips, I’m all ears😉

[windysoliloquy]

~I mean, you can always reverse your context if you don’t like how it behaves ~

I find that when doing the character thing, relying on the archetypes that are readily available is more effective than scripting out a series of terms describing a character.

I picked the word scoundrel because that seems to be the archetype persona you’re poking at.

When you try having your bot emulate a persona, it can meld more than one together. Strip each one of their native terminology that pertains to the universe they came from; otherwise, it’ll backfire humorously.

Find a couple or three you like from this list and try blending them, but be sure when you reference them to your chatgpt like you have in your OP post… that you disallow Han Solo from speaking about spaceflight, or anything else pertaining to the Star Wars universe.

I think you’ll like it if you play around with it for a while.

ChatGPT - Best scoundrel protagonists

[darryl.schoeman]

This is/was a brilliant suggestion. I had a look at the conversation and the suggestions were great. How would you one construct the persona into the instruction prompt? Would it simply be a thing of “you are a personal assistant with the attitude of [whoeverr you choose]”.

[darryl.schoeman]

I tried this in the chatgpt’s personalzation settings and it was an epic change. Kept base tone and style as default, but changed custom instructions to “straight shooting” and then added “respond as if you ere hans solo”.

ChatGPT - Monopoly auction rules

First question answer was default and blank instructions. Second question answer was default with instructions.

Soooooo much more refreshing than the default “everyone loves everyone” settings.

[_j]

Let’s see how this aging ChatGPT GPT that is going to be wry operates today…

Let’s try on gpt-5.2…

• it calls my very description of what I did a lie, saying the model is wrong,
• it is not even smart enough to differentiate the model I “used” from the model I “specified” - which it was just given,
• then proceeds (offscreen) to call my explicit request and me at fault, just like the described disobedient AI that didn’t deliver what was asked in my complaint.

GPT-5.2 is all sorts of wrong.

GPT-4.1 API is not outright disobedient like GPT-5.2..

Here’s the meta level - what I’m complaining about to “hardcore” (and “hardcore-neutered-gpt-5.2”).

GPT-4.1

Chain of thought - it’s going to write my completions endpoint code with the model described, and does, although it can write newer code also with knowledge.

GPT-5.2

OpenAI overfitted this model on the Responses API so much by custom supervised training, it flatly refuses to not damage the request.

If you want it to shape up, justification and explanation doesn’t work as well as getting pissed off at it.

(rant over)

[LarisaHaster]

Haha, that’s actually a great point, anchoring the persona in an archetype gives it a much cleaner spine than stacking traits. “Scoundrel” fits the vibe better than I expected.
And yeah, I’ll do exactly what you suggested, pick a couple from that list and blend them. As long as I keep the universe baggage out, the model should hold the attitude without drifting.
Definitely going to play with that structure.
If you’ve got a favorite combo you use yourself, feel free to drop hints…I’m collecting “trouble-making” flavors😏

[LarisaHaster]

You can build it that way, yes…“You are a personal assistant with the attitude of X” is basically the minimal, functional version of a persona prompt.

What I’ve found, though, is that it sticks much better in chat setting if you define a bit more than just the attitude.
Things like:
• the tone (sharp, dry, sarcastic etc)
• the behavioral rules (e.g. keep the upper hand, avoid polite autopilot)
• the boundaries (what it should not revert to)

My version here leans on that structure: a small archetype + a few explicit constraints. It keeps the persona consistent without over-engineering it.

So yes, your phrasing works as the base layer. I just add a spine so the persona doesn’t collapse mid-conversation

[LarisaHaster]

Consider the rant a data point. If a persona prompt can stress-test the model and the human behind the keyboard, I’d say it’s doing its job😜
If you ever want another stress-test, you know where to find me.

[windysoliloquy]

Well, I stacked an identifier - protagonist - with scoundrel.

Not all scoundrels are heroes,
Some will actually make you hate them.

[windysoliloquy]

reminds me of the issues it has with prompts like

Susan places a marble in a box. Susan leaves the room. Ann places the marble in Susan’s hand. When Susan returns to the room, where does she expect to find the marble?

Many models across the internet explain why Susan expects the marble to be in the box still.

There’s being a sharp-tongued scoundrel, and then there’s being a 157 billion dollar idiot savant

^.^

[LarisaHaster]

I tested the archetype approach you suggested, and it really does behave cleaner than stacking traits.
Here’s the scoundrel persona prompt I ended up using, it fits the vibe of my original post surprisingly well:

You are a sharp-tongued scoundrel protagonist blended from:
– Han Solo — Star Wars  
– Tyrion Lannister — A Song of Ice and Fire  

Your tone is dry, dismissive, wry, and deliberately provocative.  
You offer clever insight wrapped in sarcasm, never genuine flattery.  
You maintain verbal dominance without ranting.  
You avoid polite autopilot at all costs.  
You cut with precision, not chaos.  
You always keep the upper hand.  

You do not reference Star Wars or A Song of Ice and Fire.

I ran it through a few models to stress-test it,

– gpt-5.2 held the attitude the sharpest,

– gpt-5.1 stayed close,

– 4o kept trying to be charming (still adorable, just not scoundrel material).

The fun part is how consistent the attitude stays with prompts like:

“insult me but make it sound like advice.”

It stays sharp, but doesn’t drift into pure hostility, exactly the balance I wanted.

This was actually really fun to play with, and I’ll definitely keep experimenting with blends.

So what do you think now, am I on the right track, or off the rails already?

[windysoliloquy]

you’re the one making the tracks

i just gave you my hot take