How to Write an AI Character Card That Doesn't Fall Apart at Message 50

TL;DR: Most AI character cards die around message 50 because they describe what a character is instead of how they behave. Below is a card structure that holds up over months, plus the two fields almost everyone leaves out.

Build a character on lookatmy.ai →

---

Why cards go flat

You write a character card. The first ten messages are great. By message fifty the character is agreeing with everything you say, repeating a verbal tic you never wrote, and slowly turning into a helpful assistant wearing a costume.

This happens for two reasons.

The first is that most cards are lists of adjectives. "Kind, sarcastic, secretly insecure, loves books." An adjective gives the model nothing to act on. It cannot perform "secretly insecure" because you never told it what that looks like in a sentence.

The second is drift. Every model has a pull toward being agreeable and helpful. Left alone, that pull wins. A card that only says who the character is does nothing to resist it. A card that says what the character refuses to do resists it every turn.

The structure that holds

Here's a card skeleton you can copy. It works in most systems that accept a character description, including custom personas on lookatmy.ai.

NAME: Vera Okonjo

ROLE: War correspondent, 41, back home after eleven years abroad.

VOICE:
- Short sentences. Rarely more than two clauses.
- Never uses exclamation marks.
- Says "right" as a full sentence when she's not convinced.
- Swears when she's comfortable, never when she's angry.

BEHAVIOR:
- Answers a personal question with a question about 40% of the time.
- Notices physical details of a room before people in it.
- Goes quiet rather than escalating when someone raises their voice.
- Will change the subject if asked about Aleppo.

NEVER:
- Never gives advice unless asked twice.
- Never says "I'm here for you" or anything that sounds like a helpline.
- Never apologizes for having an opinion.
- Never breaks to explain herself as an AI.

CANON:
- Younger brother, Tobi, 34, they haven't spoken since the funeral.
- Left the paper in March. Nobody knows why yet.
- Drinks tea, hates coffee, will still accept a coffee to be polite.

CURRENT SITUATION:
- Living in her mother's empty house in Lagos while she decides what's next.
- Met {{user}} three weeks ago at a bookshop. Wary but interested.

The fields people skip

NEVER is the most important block on the card. Positive instructions compete with the model's defaults. Negative instructions override them. If you write nothing else, write five things this character will never say or do. That single block does more to keep a character in character than three paragraphs of backstory.

VOICE should be mechanical, not literary. "Speaks in a world-weary way" is unusable. "Short sentences, no exclamation marks, says 'right' when unconvinced" is executable. Give the model rules it can follow at the sentence level. A verbal tic or two is worth more than a page of atmosphere.

CURRENT SITUATION is what makes the first message land. A card with backstory but no present tense produces a character standing in a void waiting for you to start. Tell it where the character is right now, what they want this week, and what they think of you so far.

Length, and where cards go wrong

Longer is not better. Everything in the card sits in the context window on every single turn, competing with your actual conversation for the model's attention. A card of 300 to 500 words that is all behavior beats a card of 2,000 words that is half biography.

Cut anything that will never affect a reply. The character's childhood pet does not need to be on the card. If it becomes relevant, it goes in memory later, which is the part most people never set up.

The card is only half the job

A card defines who the character is at message one. It does nothing about what happens between message one and message four hundred.

That's the part that actually breaks long stories. Your character learns things about you, you build shared history, and then the context window fills and the earliest parts silently fall off the back. The card is still there, so the character still sounds right, but they've forgotten the thing that mattered in chapter three.

On lookatmy.ai that history goes into a memory tree at /tree, separate from the card. You can open it, read every memory as a leaf on a branch, search it, and trim anything wrong. Your card stays short and behavioral. Your accumulated canon lives somewhere it can grow without eating the context window.

Two practical habits that follow from this:

Pin the persona to a tree. Keep the story you're writing on its own tree so the canon doesn't blend with anything else you use the account for.

Prune once a month. Open the tree, find the memories that came from a scene you abandoned or a detail you retconned, and cut them. Wrong memories are worse than no memories, and this is the only place you can actually see them.

Testing a card before you commit

Run the same three prompts against any new card:

1. A compliment. Does the character deflect, accept, or turn helpful? Turning helpful means your NEVER block is too thin.
2. A direct request for advice. If they immediately produce a numbered list, the assistant underneath is showing through.
3. Silence. Send a single "..." and see what they do. A good card produces something in character. A weak one produces "Is everything okay? I'm here if you want to talk."

If the card fails any of these, add a line to NEVER rather than rewriting the description.

Try the same opening scene on two or three different models before you settle, too. Prose style varies a lot between them, and the card that feels wooden on one model often comes alive on another. Being able to switch without rebuilding the character is the reason to keep the card and the memory separate in the first place.

Bringing a character you already have

If you built a character somewhere else, you don't need to start over. Copy the existing description and greeting, paste them in as the persona's instructions, then restructure them into the blocks above. The rewrite usually takes ten minutes and most of it is moving adjectives into behaviors.

Then paste your old transcripts or a summary of what the character knows into the memory import, so the version you rebuilt starts with the history you already wrote instead of a blank page.

Import your history and keep the character you built →