Writing an agent personality that holds up

The personality field is the most fun part of setting up an agent, and the easiest to get wrong. People write a paragraph of vibes, the first few replies feel great, and then the agent is chirpy while telling you a payment failed.

Describe behaviour, not adjectives

"Friendly, professional and concise" gives a model almost nothing to act on, because it is three adjectives that conflict under pressure. Rules about behaviour survive contact with a task in a way adjectives do not:

• Say what to do when it is unsure. "Ask one clarifying question rather than guess" is worth more than any tone word.
• Say what to do with bad news. Tone should change when the message does, and that has to be written down.
• Say what to never do. A short list of hard limits is the most reliable part of any personality.

Length is not depth

A long personality is not a detailed one. Past a certain point extra sentences mostly add ways to contradict yourself, and a contradiction is resolved differently every time, which reads to you as an agent that is inconsistent for no reason.

If two lines of your personality could ever disagree, the model will pick one at random and you will call it a bug.

Cut anything the agent cannot act on. "You are an expert in productivity" changes nothing. "When a task has no deadline, ask for one before scheduling anything" changes every reply it touches.

Test it on the boring path

New personalities get tested on interesting questions, which is exactly where they look best. Test yours on the dull ones instead: a scheduling conflict, a half-finished request, a message that arrives at midnight, something it has no answer for. That is where a personality either holds or falls apart.

Then leave it alone for a week and read the thread back. Reading your own agent in bulk is the fastest edit you will ever make.