How to write a character an AI can actually play

The usual sheet is a pile of adjectives. Warm but guarded. Fiercely loyal. Dry sense of humour. Every word is true and none of them tell a model what to do when you say something careless at two in the morning.

A model performing "guarded" has to guess what guarded looks like this second. It will guess something generic, because generic is what the average of all guarded characters looks like. That is the flatness people blame on the model, and it usually started in the sheet.

The fix is to write decision rules instead of qualities. Not "guarded" but: she answers a personal question with a question, and only gives a straight answer after the third time you ask without pushing. That is performable. It produces the same impression as "guarded" and it produces it consistently.

A character needs something, and getting it has to cost them something else. That tension is the engine — without it there is nothing for a scene to be about.

Keep it concrete and current. Not "wants to be happy" but "wants to sell the workshop before the winter, and cannot say so in front of her father". Now every scene in that workshop has a second layer, and the model has something to play against.

The cost matters as much as the want. If nothing is in the way, the want is satisfied in scene two and the character is finished.

This is the most skipped section and the most useful one. A character who agrees with everything has no shape (we wrote about why an agreeable AI kills the story). What gives them edges is a specific refusal.

One or two are enough. "She will not talk about the year she was away." "He does not lie to children, even when it would help." Anything the story pushes at those lines produces friction, and friction is where character shows.

"Speaks bluntly" is a label. What the model can act on is how:

• Length. Short sentences under stress, or long ones? Does that flip when they are lying?
• Avoidance. What do they change the subject away from?
• Register. Do they swear? Use names? Say "we" when they mean "you"?
• Body. One recurring physical habit — what they do with their hands while waiting — does more for presence than a paragraph of description.

Long backstory feels like effort and mostly is not. Most of it never becomes actionable, and in a long story it gets summarised away.

Three or four concrete facts carry further, if each one can produce a scene: a person they owe, an object they keep, a place they avoid, a thing they are good at that nobody asks about. Those are hooks. A chronological history is not.

The most common mistake in a good sheet is completeness. If you have specified how they react to everything, there is nothing left for the story to reveal, and thirty hours in you are reading your own notes back.

Write the spine — want, cost, limits, voice, a few facts — and stop. Let the rest arrive. In a system that records what happens, the things a character does under pressure become part of who they are, and that accumulated version is always better than the one you could have planned.

Before:

Mira, 34, a talented restorer. Warm but guarded, with a dry wit. Loves old things and hates being pitied. Has a complicated relationship with her family.

After:

Mira, 34, restores furniture in her father's workshop.

Wants: to sell the workshop before winter. Cannot: say so while he is still in it.

Will not: discuss the year she was away. Deflects with work.

Voice: short sentences, shorter when hurt. Rarely uses names. Answers a personal question with a question.

Habit: runs a thumbnail along a joint while thinking.

Facts: owes money to a man called Teo. Keeps her mother's chisel and does not use it. Can tell the age of a piece by smell and never mentions it.

Same person. The second one can be performed.