Engineering "Tough Love": The Prompt Stack behind Astute Kitty.

Most AI assistants are too "nice."

If you ask ChatGPT for dating advice, it gives you a polite, generic list of tips that rarely works in the real world.

For LoveActually.ai, we had a different challenge:

How do we engineer an AI to be brutally honest without breaking its RLHF safety rails?

Enter Astute Kitty. We built her personality stack using:

-Negative Constraints: Explicitly forbidding "supportive clichés."

-Logical Priority: Forcing the model to scan for contradictions in user behavior first.

-Strategic Brevity: Keeping outputs sharp and high-impact.

It’s been a wild ride fine-tuning a model to "roast" users into becoming better versions of themselves.

Dev Question:
When building AI personas, do you prefer a "Universal Assistant" or "Domain-Specific Personalities" with extreme traits?

Can't wait to show you the full logic tree on Launch Day!