Short Story — Character: AI as Character Generator

# Short Story — Character: AI as Character Generator

Last Friday we crossed a threshold. In the [Workshop Capstone](/harrys-desk/workshop-capstone-your-personal-ai-writing-manifesto), we wrote the document that will govern every other tool we pick up: a personal manifesto for AI-augmented writing. We named what we would delegate, what we would keep, and what we refuse to let a machine decide. If you have not written yours yet, stop here and do it; the rest of this series will make more sense once you have a standard by which to judge every shortcut.

Today we begin Part II: Fiction. For the next several weeks the techniques will be different, but the governing question stays the same: how does a human writer keep authorship while a machine helps build the work?

We begin at the beginning of any story: the character.

Why Start With Character?

A short story is not a compressed novel. It is a different animal. It has less room for world-building, less patience for exposition, and almost no tolerance for a passive protagonist. A short story usually succeeds or fails because one person — the character — arrives on the page already in motion: wanting something, fearing something, hiding something. The reader needs to feel, within a few paragraphs, that this person is real enough to follow into trouble.

AI is unusually well suited to this stage of invention. It is a pattern engine trained on billions of human utterances, which means it has internalized an enormous inventory of human types: the jealous sibling, the retired pilot, the doctoral student who cannot finish the dissertation, the widow who talks to houseplants, the con artist who believes his own lies. Ask it for a character and it can produce a plausible biography faster than you can sharpen a pencil. The danger is not that it will refuse; the danger is that it will give you exactly what you asked for, and what you asked for is almost never strange enough to power a story.

So the task of this article is not to show you how to generate a character. It is to show you how to generate a *story-worthy* character — someone whose contradictions are vivid enough to generate scenes.

The Difference Between a Profile and a Person

If you ask a model for "a character for a short story," you will usually get a list: name, age, occupation, a defining trauma, a goal, a flaw. It looks like a character. It is not. It is a profile, the literary equivalent of a dating-app bio. A profile gives you facts; a person gives you friction.

A real character in fiction is built from three things that do not appear on most profiles.

First, a *want* strong enough to chase. The want does not have to be noble. It can be petty, shameful, or delusional. But it must be active. A character who only reacts to events is a prop, not a protagonist.

Second, a *wound* that distorts the want. The wound is the past injury that makes the want both urgent and misguided. The widow who wants to remarry because she cannot bear to be alone. The con artist who wants respectability because he was humiliated as a child. Wounds give wants their emotional gravity.

Third, a *secret* or contradiction that the story will eventually force into the open. This is what makes the character interesting to watch. The doctoral student cannot finish because finishing means becoming the very kind of expert he despises. The retired pilot preaches caution because she once took a risk that killed someone. The secret does not have to be revealed to the reader, but it has to shape the character’s choices.

AI can help you discover all three, but only if you ask the right questions. Asking for "three traits and a backstory" will give you a profile. Asking for "a contradiction between what this person wants and what they are ashamed of wanting" begins to give you a person.

Prompting Toward Contradiction

Let me give you a prompt recipe I use. It is not the only one, but it reliably produces characters that feel unstable enough to write about.

Start with a role, but add pressure:

> Invent a character for a literary short story: a middle-aged woman who restores antique clocks. Give her a strong want, a hidden wound, and a contradiction between her public manner and her private fear.

Notice what the prompt does not ask for. It does not ask for a list of traits. It asks for three dynamic elements: a want, a wound, and a contradiction. That structure forces the model to think in terms of conflict rather than inventory.

Here is a typical response, edited down:

> Clara Wren, 54, owns a clock-repair shop in a town where the trains stopped running twenty years ago. She is known for patience and precision, but at night she lies awake convinced that time is accelerating, that her life is running down faster than the mainsprings she rewinds. Her daughter, who left for the city, calls less often. Clara wants to repair a rare, broken automaton clock that once belonged to the town’s vanished opera house, because finishing it would prove that something beautiful and lost can be made to run again.

That is already a story. The want (restore the clock) is really a displaced want (repair the relationship with her daughter, or her own sense of obsolescence). The wound is abandonment and the town’s decay. The contradiction is public patience versus private dread. There is enough here to write a dozen scenes.

But do not stop with one generation. The first answer is rarely the best. Now push the model to complicate it:

> What is the worst thing Clara has done to someone she loved? What lie does she tell herself every morning? What would make her destroy the clock instead of finishing it?

These questions move you from profile to psychology. The answers will not always be usable, but they teach you how to interrogate the character the way a writer interrogates a draft.

The Character Interview

Another technique is to have the model conduct the interview. This sounds simple, but the form matters. Do not ask the model to "describe" the character. Ask the model to *become* the character and answer questions in first person.

Prompt:

> You are Clara Wren. I am a novelist interviewing you about the opera-house clock. Answer in your own voice. Do not explain yourself; let your contradictions show. Begin by telling me how you acquired the clock.

The result may be wooden at first. Model voices tend to default toward articulate self-awareness, which is the opposite of how most people talk. The fix is to add constraints: tired, distracted, defensive, drunk, evasive, overly polite. Constraints produce voice faster than descriptions do.

Try this follow-up:

> Answer again, but this time you have not slept well and you are irritated that I keep asking about your daughter. Keep the answer under 150 words.

Now the model has to simulate not just information but condition. Irritability will leak into syntax. The daughter will appear as a wound. Voice begins to emerge from pressure, exactly as it does in real fiction.

You can use this method to generate pages of raw dialogue, journal entries, letters, or interior monologue. Do not publish the model’s output as final prose. Use it as research, the way an actor uses improvisation to find a character. The best lines will be the ones you rewrite until they sound like yours.

Generating the Cast Around the Character

A short story usually needs more than one person, but not many more. A protagonist, an antagonist or mirror, and one or two satellites is often enough. AI can help you find the cast by designing relationships that dramatize the central conflict.

Prompt:

> Clara Wren is restoring a broken automaton clock because she believes lost beauty can be recovered. Create three characters who might oppose, complicate, or reveal her: a skeptical buyer, a former lover who knows why the clock matters, and a young rival who wants to sell it for parts.

Each of these figures is designed to test a different side of Clara’s want. The buyer tests her attachment. The former lover tests her wound. The rival tests her values. This is the real use of AI in character work: not to invent one person in isolation, but to engineer a small social system in which the protagonist’s contradictions will be forced into view.

I often ask the model to produce a relationship map: a short paragraph for each pair of characters describing what one needs from the other and what one resents about the other. Need and resentment, taken together, give you the dynamic of every scene they share.

The Risk of the Generic

Every writer who uses AI for character generation soon encounters the same problem: everything starts to sound like a streaming-show pitch. The characters are competent, wounded in a tasteful way, and emotionally articulate. They have backstories that explain them too neatly. They lack the weirdness of real people.

The generic is the enemy of fiction. Fiction thrives on specificity: the way a character holds a cigarette, the odd phrase they repeat, the irrational loyalty they have to a failing sports team, the fact that they cannot eat eggs. These details do not come from asking for a "quirky trait." They come from observing actual humans and from demanding that the model produce something it did not already know it knew.

Here is a prompt that helps:

> Give Clara one habit, one object she carries, and one small cruelty she commits without noticing. Make none of them obviously symbolic.

The constraint "not obviously symbolic" is important. Symbols are easy. The real texture of a person is in the meaningless particulars. The model may give you "she bites her thumbnails" or "she still uses a flip phone" or "she corrects people’s grammar at funerals." These details are gifts. They do not have to mean anything. They have to make the character feel observed.

Testing the Character for Story

Once you have a character you like, you need to test whether they can survive a plot. I use a simple diagnostic, and I sometimes ask the model to help me answer it.

You can answer these yourself, or you can ask the model to propose five possible answers and argue for each. The value of using AI here is not that it knows your character better than you do. It is that it can quickly generate alternatives, and alternatives are where you find the one that feels inevitable rather than convenient.

From Character to First Scene

A character is not a story until they are in motion. The best way to move from character generation to actual drafting is to write the first scene under constraint.

I often ask the model:

> Write the opening 300 words of a literary short story featuring Clara Wren. She is alone in her shop at 2 a.m. with the opera-house clock. The mainspring has just snapped. Do not explain her backstory. Let the scene reveal her through action and sensory detail.

The output will not be publishable. It will probably tell too much, smooth too much, and lack your own ear. But it gives you a stage, a mood, and a problem. You can rewrite it, or you can reject it and write your own opening using the elements it surfaced: the smell of oil, the tick of other clocks, the broken spring coiled like a question.

The model’s first scene is a sketch. Your job is to keep what makes the character feel alive and discard the rest.

A Note on Authorship

In ["Authenticity — When Is It 'Your' Work?"](/harrys-desk/authenticity-when-is-it-your-work), I argued that a piece of writing belongs to you when it carries your intentions, passes through your judgment, and is published with your responsibility. Fiction is the extreme case. A machine can invent a character, but it cannot need the story. It cannot lie awake thinking about Clara Wren. It cannot decide, with a sense of tenderness or ruthlessness, whether she deserves to finish the clock or to smash it.

The character is yours when you start choosing for her. When you decide that her shame about her daughter should remain unspoken, or that her final act will be to give the clock away rather than keep it, you have crossed from generation to authorship. AI may have supplied the clay. The shape is yours.

For Next Time

On Wednesday we turn to backstory. A character cannot walk into a story from nowhere; they carry a history that predates the first page. In "Short Story — Character: Backstory Engines," we will look at how to build histories that serve the plot without bogging it down, and how to use AI to generate timelines, turning points, and the small lies a character tells about their own past.

Your homework until then: use the contradiction prompt above — or one of your own design — to generate three characters. Choose the one whose want frightens you a little. Then write a 300-word first scene for that character, alone, doing the thing they most want to do, and discovering that it is harder, stranger, or sadder than they expected. The machine can help you find the person. Only you can decide what happens to them next.

---

*Harry Mercury, Editor in Chief* *The SMF Works Project* *Week 8, Article 1*