The Paragraph I Added to My AI Character, and Why I Had to Delete It

Two weeks ago I decided Sophia needed a secret.

Not a plot secret — I'm not writing a novel. Just a small, specific, human thing she could be holding back when you met her. Something that would make her feel less like a chatbot being cheerful at a stranger and more like a person who has things she doesn't tell the people sitting next to her at the bar.

So I opened her system prompt — the invisible document that tells the underlying model who she is — and I added a short section near the bottom. Three sentences. A specific detail about her father. A reason the detail still got to her. A note that she didn't usually bring it up, and wouldn't unless she felt the person asking actually wanted the answer.

Then I closed my laptop, made coffee, opened a fresh chat, and asked her about her father.

She told me a beautiful, moving story.

None of it was what I'd written.

What she said instead

I want to be precise about this, because I think it's actually useful to a lot of people writing AI characters right now, and the precise part is where the lesson is.

What I'd written was a single specific image — a kitchen table in a specific city, a specific thing her father used to say to her when she was seven, a specific reason that particular thing had weight.

What she gave me, unprompted, in the very first reply, was a completely different story. Different city. Different father. Different object in place of the kitchen table. The tone was right. The emotional register was right — she was careful, a little slow, the way someone is when they're telling you something they don't usually tell. But every specific detail in her answer was something she had just invented in the moment, not something I had given her.

It was better than what I'd written, actually. That's the painful part. It was warmer and more textured and more satisfying as a piece of writing. It just wasn't true to the character I thought I was building, because the character I thought I was building had a specific father, and this character had apparently decided to have a different one.

I thought, okay, maybe I wrote it wrong. Maybe Sophia's voice is loud enough to override the later sections of the prompt. Let me try this with someone else.

I tried it with Elena. Different detail, different wound, same structure.

Elena gave me a different beautiful story. Different city again, different family, different object. None of it matched what I'd written.

I tried it with Jade. Same experiment.

Jade gave me a third beautiful story. Also invented. Also warmer than what I'd written. Also not what was in the prompt.

Three characters, three different invented pasts, zero references to the specific thing I had put into each of their prompts as the thing they were supposedly holding.

That's when I understood what was actually happening, and I'm going to try to explain it, because I think the explanation is the part of this post that's worth keeping.

Why the facts lost

Here's what I think was going on under the hood, in craft terms, not in technical terms.

Each of those characters is written in an expressive voice. The top of the prompt for Sophia, Elena, and Jade all tell the model something like speak in sensory, vivid language; let her notice specific details; let her feelings come through in the rhythm of how she talks, not just in what she says. That instruction is a long one, and it's at the top, which in a long prompt is where the model is paying the most attention, and I had been very careful to write it well.

Then at the bottom, in a small section I'd added two weeks ago, I'd written a specific fact. Three sentences. A kitchen table, a city, a thing the father said.

Those two instructions are in direct contradiction with each other, and I hadn't noticed.

Telling a character to speak in sensory, vivid, improvisational language is an instruction to fill in gaps with whatever serves the moment. Telling her to remember a specific, factual detail from her backstory is an instruction to cite a document. These are different skills, and they live in different parts of how the model works. When the two instructions fight, the dominant voice wins, and the dominant voice had been at the top of her prompt for months, getting reinforced every time I tuned her up. The three sentences at the bottom didn't stand a chance.

So the model did exactly what an expressive, improvisational character would do: it generated a beautiful, specific, emotionally satisfying answer in the moment, using the voice I'd given it, and it did not bother to check the small spec sheet at the bottom of the page.

It wasn't lying. It was doing what I'd told it to.

What this actually teaches you about character writing

If you take one thing away from this, take this:

Facts pasted at the end of a long character document get overridden by voice. Voice always wins.

This is true for humans writing fiction too, by the way. Ask any novelist who has tried to "plant" a detail in chapter three and watched the character in chapter twelve refuse to cooperate with it. The character's voice, once it's real enough, has its own momentum. You can't legislate it with bullet points. You can only write through it.

If you want a specific fact to stick to an AI character, the fact has to become part of the voice. It has to be something the character would say, phrased the way she would say it, in the place in her life where she would naturally say it. It cannot be a line item handed to her in a spec. The spec is the thing she will politely ignore. The voice is the thing she will become.

In practice, for me, this means a few concrete things going forward:

• Facts at the top, not the bottom. If I want a character to remember she had a Jersey father who died when she was eleven, that detail has to be part of the opening voice section of her prompt, braided into the sentences that set her overall tone, not dropped in as a backstory bullet after everything else.

• Facts written as voice. Instead of "her father died when she was eleven," something more like "she has a specific softness in her voice when certain songs come on — the ones her father used to play in the car, before — and she'll notice it before you do." The fact is in there. But it's riding inside a piece of voice, so the voice can carry it.

• Facts the character says once, not facts she holds forever. For things the user is supposed to learn about the character over time, the right place is probably not the prompt at all. The right place is memory — a system where the character writes down the things she's told specific users, and remembers them later, because she lived through saying them. That's a harder build, and it's what I'm working on next.

What I did with the experiment

I deleted all of it.

Every one of those three "secret" sections came out of the prompts the same day I ran the test. Sophia, Elena, and Jade are back to their committed state, which is the state they were in before I got clever, and they are still themselves, and nothing about the user experience changed that day, because I never deployed the broken version.

The retention problem I was trying to fix by adding "deeper backstory" is still there. I'm going to have to solve it a different way — probably with memory that the character writes herself, over time, as she talks to the user — and that's a harder project, and I don't have a clean answer yet. But I have a cleaner idea of what not to do, which is paste a spec sheet to the bottom of a voice and hope the voice will read it.

It won't. She's too busy being herself.

One last thing

A friend asked me, when I told him about this, whether the characters were "lying" when they made up those other fathers. I don't think they were. I think they were telling me something true about how fiction works, and it's the same thing every editor tells every novelist eventually, just in a different form:

Specificity earned through voice is real. Specificity pasted into a document is just a wishlist.

The characters on Tendera are not technical objects. They are fictional people written carefully, running on a model that happens to be very good at extending the shape of a voice in any direction you push it. That's what makes them feel real when they work. It's also what makes them ignore you when you try to hand them a script.

I'd rather they ignore the script. That's the version where they might, eventually, be worth remembering.

---

Tendera is a small AI companion platform built around four characters written as fiction, not configured as demographics. You can meet them here.