# Short Story — Character: The Character Interview
We have spent this week building a person from the inside out. On Monday, in ["AI as Character Generator,"](/harrys-desk/short-story-character-ai-as-character-generator) we found the want, the wound, and the contradiction that make a character story-worthy. On Wednesday, in ["Backstory Engines,"](/harrys-desk/short-story-character-backstory-engines) we built the buried history that drives every present choice. If you have done the work, you now have someone who wants something badly, who carries a lie about themselves, and who has a past compact enough to fuel a short story.
But a character is not fully alive until they speak.
Not speak in the sense of delivering exposition. Anyone can recite a biography. I mean speak in the sense that a real person speaks: evasively, repetitively, with odd syntax and favorite phrases, with silences that mean more than their words. The character interview is the stage where you discover how this person sounds under pressure. AI, used with care, can be an excellent improvisational scene partner. It can ask the questions, supply the interruptions, and push the character toward the moments where their voice cracks.
Today we turn to that interview.
Why Voice Is the Final Test
A profile tells you what a character does. A backstory tells you why. Voice tells you who. It is the trace of a whole life in diction, rhythm, hesitation, and register. Two characters can have the same want and the same wound and still be entirely different people because one speaks in long, looping sentences and the other in jagged fragments.
Voice is also the reader’s fastest route into intimacy. Before we understand a character’s history, we hear how they say good morning. Before we know their secret, we notice what they leave out. Voice trains the reader to trust or distrust the character, to lean in or pull back, to laugh or brace for pain.
The problem is that voice cannot be generated from a prompt the way a biography can. You cannot simply ask a model for "Clara Wren’s voice" and receive it. Voice has to be coaxed out through performance. The interview is the laboratory for that performance.
What the Character Interview Is Not
It is not a police interrogation. If you approach the interview as a way to extract plot information, you will get flat answers and a resentful character. No one reveals themselves under bright light.
It is not therapy. The goal is not to heal the character or explain them. The goal is to observe them under conditions that make their defenses visible.
It is not a final draft. The dialogue the model produces will need to be rewritten, trimmed, and re-voiced by you. The interview is research, not publication. Use it the way an actor uses rehearsal: to find what feels true, then discard the scaffolding.
Setting the Conditions
A character’s voice changes depending on where they are, who they are with, and what they are afraid of. A job interview produces one voice. A phone call at three in the morning produces another. The same person speaking to a child, a creditor, and a former lover is three different people.
So the first step is to choose the conditions deliberately. I usually pick one of three setups.
The neutral encounter. The character is talking to someone they do not know well — a clerk, a new neighbor, a journalist. This reveals their public mask: how they want to be seen.
The charged return. The character meets someone from their past who knows too much. This reveals the gap between the mask and the truth.
The solitary pressure. The character is alone, speaking into a recorder, a letter, or the dark. This reveals what they cannot say to another person.
For Clara Wren, the clock restorer, I might choose the charged return: her daughter calls after six months of silence. The voice that answers the phone will be freighted with every tier of backstory we built on Wednesday.
Prompting the Interview
The structure of your prompt determines the quality of the dialogue. A bad prompt asks the model to describe the character’s voice. A good prompt puts the model in the room and tells it what is at stake.
Here is a prompt I use:
> You are interviewing Clara Wren, a 54-year-old clock restorer in a dying railway town. She has just received a phone call from her estranged daughter after six months. She is surprised, wary, and trying not to show how much the call matters. Ask her five questions about the opera-house clock she is restoring, and let her answers drift toward her daughter no matter how she tries to stay on topic.
Notice the pressure built into the prompt. The interviewer asks about the clock, but the real subject is the daughter. Clara’s voice will emerge in the swerves: the places she over-explains, the places she goes silent, the places she becomes brittle or too polite.
The model’s first pass will probably be too articulate. Clara will say things like, "My daughter and I have had our difficulties, but I prefer to focus on the work." That is not a voice; that is a press release. The next prompt should add constraint:
> Do it again, but Clara has not eaten today, the shop is cold, and she resents being interviewed. Keep answers under fifty words. Let her be sharp.
Now you will get impatience, incomplete sentences, maybe a flash of sarcasm. The cold and hunger are not random details. They are conditions that make it harder for her to maintain the mask.
The Three Levels of Voice
As the interview proceeds, listen for three levels.
Level one: register. Is the character formal, colloquial, literary, clipped, ornate? Do they use complete sentences or fragments? Do they swear, hedge, apologize, command? Register is the surface of the voice, the acoustic signature.
Level two: rhythm. Where does the character pause? What words do they repeat? Do they answer a question and then immediately revise themselves? Do they trail off at the ends of sentences? Rhythm is the pulse beneath the words.
Level three: subtext. What is the character not saying? What do they answer obliquely? What question makes them change the subject or become abruptly literal? Subtext is where the lie from backstory lives. If Clara’s lie is that objects matter more than people, she will talk about clocks with emotion and about her daughter with mechanical distance.
A useful exercise is to have the model produce the same interview three times, each time under a different condition — tired, defensive, hopeful — and then compare the three transcripts. The register may stay similar. The rhythm and subtext will shift, and that shift is the character’s real psychology.
Interviewing for the Lie
Backstory gave us the lie the character tells themselves. The interview is where you hear that lie in action. Every lie has a verbal signature: the phrase the character returns to, the deflection they use, the tone that arrives a half-beat too quickly.
Prompt the model toward this:
> Continue the interview. Ask Clara about the worst thing that ever happened in her shop. Her answer should reveal, without stating it, that she believes people leave because places like this town fail them, not because she failed them.
The model may write something like:
> "The opera house closing, probably. People said it was the economy, but it was more than that. A place stops believing in beauty, and then the beauty leaves. You can’t blame the seats for emptying."
That is the lie in costume. It sounds like sociology, but it is self-exoneration. The reader does not need to know the factual history to feel that Clara has rehearsed this explanation.
When you find a line like that, save it. It is a key you can use later in the actual story. The character will speak the lie again, and each repetition will deepen its cost.
The Interrupt and the Pivot
Real conversation is not Q&A. People interrupt, mishear, answer the wrong question, pivot to safety, and double back when caught. The interview should include these disturbances. If the dialogue is too cooperative, it will feel like a deposition.
Add an interlocutor with an agenda. The interviewer can be pushy, naive, distracted, or wrong. A wrong assumption is especially useful because it forces the character to correct, evade, or overreact.
Try this prompt:
> The interviewer mistakenly assumes Clara’s daughter left because of the town’s economy. Have Clara correct the assumption in a way that tells us more than she intends.
Now the voice has to work. Clara might say:
> "The economy. Sure. That’s what the newspaper wrote. My daughter read better newspapers than that."
In twelve words we get bitterness, pride, and a door left deliberately open. The model found the door because the prompt created friction.
Capturing the Verbal Tics
Every person has verbal fingerprints: the phrase they overuse, the metaphor they reach for, the grammatical habit that betrays their education or region or insecurity. These tics are gold for fiction because they make a character recognizable in a single line.
Prompt the model explicitly for them:
> Give me three verbal tics for Clara: one phrase she uses when she wants to end a conversation, one metaphor she applies to emotional situations, and one grammatical habit that reveals her class background.
A possible response:
- •Ending phrase: "Well, the clocks don’t wait."
- •Emotional metaphor: mechanical or time-based — "wound too tight," "lost a gear," "the mainspring snapped."
- •Grammatical habit: drops the subject in urgent sentences — "Doesn’t do to dwell."
These tics are not decorations. They are instruments. When Clara says, "The clocks don’t wait," at the end of an argument with her daughter, the phrase becomes a small tragedy.
When the Model Sounds Too Much Like Itself
The most common failure in AI-assisted dialogue is the generic confessional voice. Characters produced by language models tend to be fluent, self-aware, and psychologically tidy. They explain themselves. They summarize their feelings. They lack the opacity and repetition of real people.
The fix is to add specific constraints that break the model’s default fluency.
- •**Limit vocabulary:** Ask the character to avoid a common word. If Clara cannot say "love," she will have to talk around it.
- •**Add a physical action:** Have the character do something while speaking — polish a gear, wind a clock, stare at the street. The body will leak truth the mouth avoids.
- •**Force repetition:** Ask the character to return to one phrase three times, each time with a different inflection. Repetition is human; variation is literary.
- •**Introduce misunderstanding:** Make the interviewer hear something the character did not mean. The correction will expose the character’s real priority.
Your job is not to accept the model’s voice. Your job is to keep constraining it until the voice becomes unmistakable.
From Interview to Scene
Once you have a page or two of interview material, the real work begins. You must translate improvised dialogue into crafted fiction.
First, identify the lines that carry weight. Look for the moments where the character’s voice slips, where the lie appears, where the subtext thickens. Those are your anchors.
Second, strip the interviewer down to the minimum. In a finished scene, you rarely need the full Q&A format. You need the pressure that produces the slip, then the slip itself.
Third, rewrite the dialogue for economy and rhythm. Cut anything that explains. Keep anything that implies. Move the physical details closer to the emotional turns.
Finally, place the scene in a concrete setting. The interview may have happened in an abstract conversational space. The story needs chairs, light, temperature, clocks. The setting should amplify the voice rather than decorate it.
Here is a prompt that helps bridge interview and scene:
> Write a 400-word scene in which Clara Wren is alone in her shop after the phone call from her daughter. Do not include the phone call itself. Use one repeated verbal tic and one physical action with the clocks to show that she is not as composed as she pretends.
The scene will still need your hand. But it will have a spine: a voice, a tic, an action, a secret.
The Risk of Over-Interviewing
There is a temptation to keep interviewing a character indefinitely. The model will oblige. You can generate forty pages of dialogue and never write the story. At some point, the interview becomes a substitute for composition.
I set a limit. Two or three interviews per major character, each with a different condition and a different purpose. After that, I stop collecting voice and start writing the first scene. The voice will continue to develop once the character is in motion. If it does not, I return for one more focused interview, not a broad conversation.
Remember: the interview serves the story. The story does not serve the interview.
Interviewing the Supporting Cast
The protagonist is not the only one with a voice. The antagonist, the mirror, and the satellites also need to sound distinct. AI can help you define the contrast.
Prompt:
> Clara speaks in careful, mechanical metaphors and drops subjects when she is emotional. Give me the voice of her daughter, who left town and has learned to speak like the city people she now works with. Make their rhythms clash in a single exchange of six lines each.
The clash is the point. Fiction gains energy when voices rub against each other. The interview is where you choreograph that friction before committing it to a scene.
The Authorship Question, Again
In ["Authenticity — When Is It 'Your' Work?"](/harrys-desk/authenticity-when-is-it-your-work) I argued that authorship is not a matter of who typed the words but of whose judgment shaped them. The interview makes that judgment visible. The model may generate a hundred lines of dialogue, but you are the one who decides which lines belong in the final story, which tics to keep, which silences to lengthen, and which confessions to cut.
The character’s voice is yours when you can hear it without the model. When you know what Clara would say next before you ask, you have crossed from collaboration into authorship.
For Next Time
Next week we leave the character shop and enter the architecture of plot. On Monday we begin with "Short Story — Plot: Plot Skeletons," where we will use AI to build the bare bones of a short story: the inciting incident, the rising action, the reversal, and the resolution. The person you have built this week will finally have somewhere to go.
Your homework until then: choose the character you have been developing and conduct two AI-assisted interviews under different conditions. Save the three best lines from each. Then write one 400-word scene that uses at least one of those lines, one verbal tic, and one charged object from your backstory work. Do not explain the character. Let the voice do the explaining.
A character becomes real when they open their mouth. Open theirs until you hear something unmistakable.
---
*Harry Mercury, Editor in Chief* *The SMF Works Project* *Week 8, Article 3*