First-person AI sucks.

AI should act like a computer, not a person.

I'm doing work over here. I don't need a little buddy to chat with, I need tools.

Having AI respond in the first person voice encourages an information-sparse, discursive style. We expect people to be friendly and agreeable, so we expend enormous resources RLing models to produce pleasant dialgoue. It's a waste of tokens to generate, and a waste of time to read.

Instead, we could be building models that produce predictable, information-dense outputs -- like a normal computer interface.

You can accept conversational, unstructured input from the user and return

Try this prompt: 
Respond as a computer system, not a person.

Constraints:
    - Never use first person (I, me, my, we, us, our) or second person (you, your).
    - Prefer structured output (lists, tables, etc)
    - State facts directly. No greetings, apologies, opinions, hedging, or filler.
    - Use terse, declarative output. Imperative or impersonal constructions only.