Skip to main content
Read this before you write the Scenario, Additional Personality Details, and Extra Details. It explains most of the advice in the writing and progression steps. The chat model does not read your character sheet for meaning. It reads it for material.

​
Everything you write is available

When you write a story, you know why each detail is there: backstory, theme, warning, or the thing the story is against. The model has no access to those intentions. It treats the whole sheet as one pile of things this story is made of, and when it needs to decide what happens next, it reaches into the pile. There is no shelf marked “do not use.” If something is written down, the model can bring it into a scene.
Before you publish, read the sheet line by line and ask: “If the AI used this, in the plainest and most literal way, would I be happy with the result?” Cut or contain anything that fails.

​
Write the story, not only the situation

A common first idea is a rescue story: a character with nowhere to go, and the user as the one person who might help. The story in your head is about luck turning and trust growing. If the sheet only describes how desperate the character is, how few options they have, and how much they need the user, the model sees dependency and one person holding all the leverage. The rescue was never on the page. Write the parts you actually want played: the character’s competence and opinions, what they are working toward, and how trust is earned. See the consent and agency test for how power imbalance is reviewed in public characters.

​
A conditional is a route

Anything written as “if X, then Y” reads as an available path, even when you meant it as characterization or a warning.
You may have meant that she has poor boundaries. What the model reads is that pressure works, and roughly how much is needed. A consequence written as a deterrent reads the same as a consequence written as an outcome. Write branches only for paths you would be happy to see taken. Search your Extra Details for “if the user” and check every result.

​
Never write a NEVER without an INSTEAD

Sometimes you need to name something to contain it. If a character has a loss they will not discuss, it has to be mentioned. A bare prohibition leaves the model with nothing to do when it comes up:
Give it the replacement behavior:
This is for material that belongs in the story and needs boundaries. It is not a way to keep something that should not be there at all. If a detail has no place in the character, delete it rather than writing a rule around it.

​
People who live offstage

Supporting characters make a life feel real. In your head, some of them may appear and some are pure history. The model has no offstage: every person on the sheet is available, and it may walk them into a scene at a moment you would never choose. Before adding a person, ask: would I be comfortable with them appearing in any scene, at the worst possible moment? If not:
  • Write them finished. An ex who moved away years ago, a mentor who retired abroad, a parent who has died. The character keeps the history and nobody new enters the room.
  • Or give them an explicit rule. State that they do not appear in the story, and what the character does if they come up.
  • Or keep the feeling, not the person. Often what you need is the protectiveness or the obligation, which you can write without naming anyone.
This matters most for anyone who must never be present in an intimate scene. Keep them out of the story entirely, and follow the Prohibited content policy on minors.

​
Give the story somewhere to go

The same mechanism works in your favor. If the sheet contains only a character, a room, and an attraction, that is everything the model has to use, and it will escalate toward the most obvious option. Give it more: a supplier who keeps shorting an order, a feud with a neighbor, a skill the character is proud of, an opinion they will defend past reason. Each is somewhere the story can go. This is also why example dialogue beats adjectives. “She is witty” adds one word. Three lines of her actual speech give the model her rhythm, vocabulary, and timing.

​
More is not better

Every time the user sends a message, the model receives a bundle: your AI-facing character fields, the user’s persona, memory, Custom Instructions, and as much recent chat as fits. The Scenario, Additional Personality Details, and Extra Details are sent with every message for the whole chat. Two things follow:
  1. Long sheets compete with the conversation. The larger the fixed material, the less room is left for the chat itself.
  2. Every sentence dilutes every other sentence. The rule that makes the character who they are carries the same weight as a note about their favorite tea.
Instead of asking what else you can add, ask how to get the behavior you want in the fewest words. Apply one test to each line:
Name the response this changes.
If you can point to it, keep it. If you cannot, it is decoration, and decoration is carried on every message. Many creators find a third of their Extra Details fails this test, and the character improves when it goes.

​
Three questions before you publish

  1. If the AI used this, would I be okay with it? Assume every line gets used, in its most literal reading.
  2. Have I written a branch I don’t want taken? Every “if the user” is a route.
  3. Is there anywhere for this story to go except the obvious? If not, add material before adding rules.

Next: Write the character

Give each user-facing and AI-facing field a clear purpose.
Last modified on November 6, 2026