When George R.R. Martin sued Anthropic in late 2023 alongside a group of authors, the complaint made an argument that went beyond copyright. It claimed that AI systems had consumed and distilled the craft of characterization — the specific techniques by which authors like Martin construct characters that readers grieve when they die. The case brought into sharp relief a question the publishing industry had long avoided: what, precisely, is the thing that makes a character feel irreplaceable?
The legal question remains unresolved. But the underlying craft question is answerable — and understanding it is the precondition for using LLMs in character work without producing hollow imitations.
Character theory across screenwriting, novel craft, and drama converges on three dimensions that, when present together, produce the sensation of encountering a real person rather than a function.
Dimension one is interiority — the character has a private world of thoughts, fears, rationalizations, and desires that the audience only partially sees. Without interiority, characters feel like puppets moved by plot necessity.
Dimension two is contradiction — the character holds beliefs or traits that are genuinely in tension. A person who is both deeply loyal and deeply ambitious; who craves intimacy but sabotages it. Contradiction is not inconsistency. Inconsistency is failure of craft. Contradiction is the signature of a real psyche.
Dimension three is specificity — the character is not a type but a particular instance. Not "a grieving mother" but a mother who expresses grief through obsessive reorganization of her kitchen, who has not cried once, who sends her dead son's friends food they didn't ask for.
LLMs trained on vast corpora of fiction have absorbed enormous amounts of character-type information. Without specific prompting, they tend to produce characters that are competent at dimension one (interiority) and weak at dimensions two and three. Knowing this lets you compensate deliberately.
Professional television writers have long used what's called a character bible — a document that establishes who a character is before a single line of dialogue is written. The practice traces to the early days of serialized television, when writing rooms needed to ensure that multiple writers produced a coherent single character across dozens of episodes.
A character bible typically captures: backstory (formative events), core wound (the thing the character has never recovered from), dominant desire (what they are always moving toward), fear (what they organize their life to avoid), and behavioral tells (how their psychology shows up in small physical actions).
This structure maps almost perfectly onto what LLMs need in order to generate character-consistent dialogue and behavior. The character bible is, essentially, a system prompt for a fictional person.
Generating consistent voice once a register is established. Producing backstory variations rapidly. Exploring how a character might respond to novel situations. Identifying gaps in a character's internal logic when asked to stress-test it.
Introducing genuine contradiction without prompting. Resisting the pull toward resolution and redemption arcs. Maintaining a character's flaws without softening them. Generating truly specific behavioral details rather than plausible-sounding generics.
OpenAI's 2023 system card for GPT-4 documented the model's ability to sustain what they termed "persona" — a consistent voice, viewpoint, and behavioral pattern across extended interactions. The card noted that this ability emerged without explicit training on persona-maintenance tasks, suggesting it was a consequence of exposure to large amounts of narrative fiction in which characters maintain consistent behavior across scenes.
This is important for writers to understand: LLMs have a latent capacity for persona maintenance that is far stronger than their capacity for persona creation. They can hold a character you define far better than they can invent one without guidance. The implication is clear — your job is definition; the model's job is inhabitation.
The writer who understands what makes characters real — interiority, contradiction, specificity — can use that understanding to construct prompts that force an LLM to produce character work at the level of intentional craft rather than genre default.
Use the AI assistant below to build a character bible for a character of your choosing. Begin by describing a character concept (genre, rough role in story, one or two traits you already have in mind). The AI will help you systematically develop the five elements of a character bible: backstory, core wound, dominant desire, fear, and behavioral tells.
Push beyond the first answer. Ask the AI to make the contradiction sharper, the specific behavior more physical and particular, the core wound more precise.
When Sudowrite — an AI writing tool built specifically for fiction — launched its "Characters" feature in 2022, the team published a transparency blog post documenting what they'd learned in beta. The most consistent user complaint was that AI-generated characters felt "like a list of adjectives wearing a trenchcoat." Users could generate backstory, desires, and fears easily. What they couldn't get was behavioral specificity — the particular gesture, the idiosyncratic habit, the detail that made a character recognizable from a single action rather than a description.
Sudowrite's engineering team documented that the fix wasn't a model change — it was a prompting structure change. When prompts were restructured to ask for behavior in a specific situation rather than traits in the abstract, the specificity problem largely resolved.
The default mode of asking an LLM about a character is to ask for traits: "What are this character's defining qualities?" This produces competent but flat results because traits are abstractions. "She is stubborn" describes a category. "When she disagrees with someone, she goes quiet and starts tidying objects within reach" is a specific — and it contains far more story potential.
The trait trap is not unique to AI. Writing workshop participants fall into it constantly. The difference is that LLMs are particularly susceptible because trait-level description is far more common in training data than specific behavioral data tied to specific characters in specific situations.
No first-pass AI character response will be fully specific. The prompting structures above are designed for iteration — use the first response as raw material and immediately push it: "Make that behavioral tell more physical and less psychological." "Give me the same contradiction but in a context where the character has no audience." The gap between a generic AI character and a specific one is almost always one or two targeted follow-up prompts.
| Level | Example | Notes |
|---|---|---|
| Trait (weakest) | "She is guarded." | Produces nothing actable or visual. Every reader imagines something different. |
| Behavior (better) | "She deflects personal questions with humor." | Actable, but still generic. Many characters share this behavior. |
| Specific Behavior (strong) | "When asked about her past, she picks up the nearest object and finds something wrong with it." | Visual, specific, tells us something about her anxiety manifesting as control. |
| Situated Behavior (strongest) | "At her mother's funeral, when the priest asks if she wants to speak, she adjusts her watch three times before saying no." | Fully inhabited. The watch detail implies years of a specific relationship with time and control. |
In 2023, novelist and writing instructor K.M. Weiland published a detailed account of using ChatGPT-4 in her character development process for her novel-in-progress. Her key finding, documented on her widely-read "Helping Writers Become Authors" blog, was that standard character prompts produced what she called "therapist-speak characters" — psychologically labeled but not behaviorally alive.
The breakthrough came when she stopped asking what a character was and started asking what the character does when their stated values fail them. The question "What does your character do when they cannot be the person they think they are?" produced, she documented, qualitatively different material — characters caught in their own contradictions rather than described from the outside.
This is the prompting principle made explicit: force the model into the gap between self-image and behavior. That gap is where real character lives.
A character becomes real at the moment you know what they would do wrong — not just what they want, but what they cannot stop themselves from doing even when it costs them. That moment is the prompting target.
Take a character you already have in mind — either from Lab 1 or one you bring to this session. Work through at least two of the four prompting structures from Lesson 2: situational behavior, contradiction activation, object anchoring, or negative space.
After each response, push for more specificity. If the AI gives you a behavior, ask it to make that behavior more physical. If it gives you a contradiction, ask which side wins in a specific scene.
In 2023, Jasper AI published internal research on what they called "voice drift" — the tendency of LLMs to gradually homogenize the speech patterns of characters over the course of a long generation session. Their finding was consistent with what fiction writers had been reporting anecdotally: characters who began a scene with distinct voices would, by the third or fourth exchange of dialogue, begin to sound alike.
Jasper's engineering team found that voice drift was not a bug but a feature of how LLMs predict token sequences — the model gravitates toward the most statistically probable dialogue given the conversational context, and the most probable dialogue is average dialogue. The fix they documented was periodic "voice anchoring" — reintroducing specific voice parameters mid-session.
Voice in fiction is not the same as dialect or accent, though those can be components. Voice is the total pattern of a character's language choices: what they talk about without being asked, what they avoid, how long their sentences are, whether they finish sentences or trail off, what they notice and name versus what they ignore, their relationship to abstraction versus the concrete.
A character who grew up on a working farm in the 1970s doesn't just speak with regional inflection — they are more likely to use object-level language rather than abstract language, more likely to describe emotions through work metaphors, more likely to be suspicious of anything they can't touch. These are voice parameters, and they can be made explicit in prompts.
A voice profile for prompting contains six parameters that together define a character's linguistic signature:
| Parameter | What to Specify |
|---|---|
| Register | Formal / informal / vernacular. Does the character code-switch? When? |
| Sentence rhythm | Short and declarative; long and qualified; fragmentary; does not finish thoughts. |
| Abstraction ratio | Speaks in concepts vs. objects. Uses feelings-language vs. action-language. |
| Deflection pattern | How the character avoids direct answers — humor, questions, silence, subject change. |
| Dominant metaphor field | The domain from which the character naturally draws comparisons: sport, money, weather, food, machines. |
| What they never say | Words or phrases the character would not use — often as important as what they do say. |
Voice anchoring is the practice of re-introducing voice parameters at strategic points in a dialogue generation session. The technique was developed independently by multiple writers working with early GPT-4 in 2023, and convergent accounts appear in the online communities r/ChatGPTPromptEngineering and the Absolute Write Water Cooler forum.
The simplest form of voice anchoring is a mid-session meta-prompt: "Before continuing, read the last six lines of dialogue. Is [Character A] still speaking in short, declarative sentences that avoid emotional abstraction? If not, revise."
A more robust form is to establish a voice-check ritual before each dialogue block: provide the voice profile, then add "Now write the next exchange of dialogue. If you notice [character] beginning to use abstraction or long-qualified sentences, stop and correct before continuing."
Researchers at Stanford's Human-Centered AI Institute documented in a 2023 paper ("Do LLMs Distinguish Character Voices in Fiction?") that GPT-4-class models showed measurable but not always reliable differentiation between character voices when prompted generically. Voice differentiation improved significantly — measured by computational stylometrics — when voice profiles were explicitly included in character definitions and re-introduced at intervals.
For writers working on full-length novels or scripts, the consistency problem compounds over time. A character established in chapter one may appear different by chapter twelve if the writer has been using LLMs without systematic voice anchoring.
The practical solution used by professional writers who have documented their AI-assisted workflows — including Brandon Sanderson's Dragonsteel Entertainment team, which published notes on AI use in their production process — is to maintain a live voice document: a single file that contains the voice profile, three to five example dialogue lines approved by the author, and a list of words/phrases to avoid. This document is prepended to any AI prompt that involves character dialogue.
This is not different in principle from how writers' rooms use character bibles — it is simply the dialogue-specific version, updated as the character evolves through drafts.
Before any dialogue generation session: state the character's register, sentence rhythm, abstraction ratio, and one phrase they would never use. After every 8–10 exchanges of generated dialogue: run a voice-check prompt asking the model to identify any drift and correct it before continuing.
First, work with the AI to construct a voice profile for your character using the six parameters from Lesson 3: register, sentence rhythm, abstraction ratio, deflection pattern, dominant metaphor field, and what they never say.
Then prompt for a dialogue exchange (your character talking with another character of your invention). After 4–5 exchanges, run a voice-check: ask the AI whether any drift has occurred and to revise accordingly.
When the Writers Guild of America went on strike in May 2023, one of the central documents in the dispute was a position paper on AI's role in the writing process. The paper explicitly addressed character arc — arguing that the studios' proposed use of AI for "character trajectory modeling" would hollow out the most human element of storytelling: the authored decision about what a person is capable of becoming and why.
The WGA's argument was not that AI couldn't model transformation — it was that AI-modeled transformation would default to the most statistically common arc shapes: redemption, fall-from-grace, coming-of-age. The document argued that the distinguishing feature of literary character arcs is their specificity to the particular character's particular wound — which requires an authorial decision that no statistical average can produce.
A character arc is not a mood trajectory (character starts sad, ends happy) or a status trajectory (character starts poor, ends rich). An arc, properly understood, is a change in what the character believes is true about themselves or the world — and that change must be caused by events rather than declared by the author.
The classic formulation, derived from Robert McKee's Story (1997) and widely taught in MFA programs, is that a character arc moves from a misbelief — a false assumption the character holds about reality — through a series of pressure events that test that misbelief, to either its abandonment (positive arc) or its fatal reinforcement (negative arc) or a conscious decision to keep believing it despite evidence (flat arc).
This structure is important for LLM work because it gives the model something concrete to track. "Show me how this character's misbelief is challenged by this event" is a far more generative prompt than "show me character growth."
The WGA's concern about default arc shapes is real and useful. When you prompt an LLM for a character arc without a defined misbelief, it will almost always generate one of three arcs: learning to trust others, learning to value themselves, or learning that ambition has limits. These are valid arcs — but they are the three most common because they appear most often in training data. Naming the specific misbelief forces the model off the default path.
The most under-developed element in AI-assisted character arc work is what craft teachers call the resistance scene — the moment, usually in the second act, when a character is given clear evidence that their misbelief is wrong and actively refuses to revise it. This scene is crucial because it makes the eventual change feel like a choice rather than an accident.
LLMs without explicit prompting tend to skip resistance scenes, moving characters from misbelief to revision too quickly and too smoothly. The fix is explicit prompting: "Write a scene where [character] actively reinterprets contradicting evidence in order to preserve their misbelief. The character should feel justified — make us understand their internal logic even if we disagree with it."
Author and writing teacher Lisa Cron, in her 2023 course material published through MasterClass, described exactly this pattern — what she calls "story-brain" tendency in AI output: "The AI wants to resolve the tension. Your job is to refuse to let it resolve the tension until the story has earned the resolution."
Unlike voice profiles, which can be held in a single document, character arc requires tracking change over time. The practical solution used by writers working on long projects is a simple arc log: a running document that records the current state of the character's misbelief, the most recent pressure event, and the degree to which the misbelief has been shaken.
This arc log becomes a context prepend for any arc-related prompting session, ensuring the model knows where the character is in their transformation rather than always starting fresh.
Never prompt for "character growth" as an abstraction. Always prompt for the specific misbelief at stake, the specific event applying pressure, and the specific moment at which resistance gives way — or doesn't. Arc lives in the specifics, not in the movement between emotional states.
Work through the Misbelief-Pressure-Revision framework with your character (or a new one). Start by defining the misbelief explicitly. Then ask the AI for three pressure events, then for a resistance scene, then for the revision or reinforcement moment.
Push past the first arc suggestion. If the AI defaults to redemption or coming-of-age, name the misbelief more specifically and ask it to try again. The goal is an arc that could only belong to this character.