In 1455, Johannes Gutenberg finished printing his first Bible in Mainz. Within fifty years, the number of books in Europe grew from roughly 30,000 manuscripts to more than nine million printed volumes. Scribes did not disappear overnight — they shifted, slowly, into editing, illustration, and scholarly annotation. But the economics of who could publish, who could reach an audience, and who could afford to be read permanently changed. The printing press did not destroy storytelling; it industrialized it, democratized it, and introduced entirely new anxieties about authenticity, authorship, and the corruption of sacred texts by mechanical reproduction.
Something structurally similar is happening now, in the 2020s, with large language models applied to writing. In November 2022, OpenAI released ChatGPT to the public. Within two months it had 100 million users — the fastest consumer-product adoption ever recorded. Within a year, major literary agencies were reporting that query volumes had doubled while the proportion of clearly AI-assisted manuscripts had become impossible to ignore. Publishers began rewriting submission guidelines. The Writers Guild of America made AI regulation a central demand of its 148-day 2023 strike. The disruption is not hypothetical; it is contractual, economic, and already reshaping careers.
This course is not a guide to automating your writing. It is a working examination of how AI tools — particularly large language models — interact with the craft of narrative structure, and where human judgment remains irreplaceable. You will leave with practical skills for using AI as a drafting and development partner, a clear-eyed understanding of its structural limitations, and the vocabulary to think critically about what these tools actually do when they "write." The limits of this course are honest ones: the technology moves faster than any curriculum, and some of what you learn here will need updating within eighteen months.
If you finish every module, here's who you become:
In the spring of 2023, Screenwriter Carrie Sun — then finishing her memoir Private Equity — ran an experiment. She fed the first act of her manuscript to ChatGPT and asked it to outline what the second act should contain. The AI produced a structurally plausible outline: escalating conflict, a midpoint reversal, a darkest moment. It was, she later noted in a Lit Hub essay, "the outline of a movie I had already seen." Every beat landed where three-act theory said it should. None of it was wrong. None of it was hers. The AI had identified the genre's gravitational field precisely and reproduced it perfectly — which is exactly the problem when the author's intention is to escape that field.
The anecdote clarifies something important. AI narrative tools are extraordinarily good at reproducing the statistical center of narrative convention. They are considerably less useful — and sometimes actively misleading — when a writer's goal is to do something structurally unusual. Understanding that gap is where this course begins.
Narrative structure refers to the deliberate arrangement of story events to produce a specific emotional and intellectual effect in an audience. It is not a synonym for plot. Plot is the sequence of events; structure is the architecture that gives those events meaning through contrast, pacing, and sequencing choices.
The most widely taught framework is the three-act structure, formalized by Syd Field in his 1979 book Screenplay: setup, confrontation, and resolution, separated by plot points that redirect the protagonist's trajectory. Field derived this partly from Aristotle's Poetics (c. 335 BCE), which identified beginning, middle, and end as the three necessary components of a unified dramatic action. What Field added was the idea of precise proportional timing — roughly 25%, 50%, 25% of total length — and the concept of the turning point as a structural mechanism rather than just a narrative event.
Other frameworks exist and matter. Gustav Freytag's pyramid (1863) adds a climax and falling action between crisis and resolution. Joseph Campbell's Hero with a Thousand Faces (1949) proposed a seventeen-stage monomyth visible across world mythology. Dan Harmon, creator of Community, distilled this into an eight-stage "story circle" that became influential in TV writers' rooms during the 2010s. Each framework is a lens, not a law.
Large language models were trained on vast corpora of human text, including enormous quantities of published fiction, screenplay databases, and writing-craft instruction. They have seen Syd Field's framework cited thousands of times. When you ask an AI to "outline a story," its default output will reflect the statistical average of what stories look like in its training data — which is heavily weighted toward commercially successful, conventionally structured work.
A large language model does not "understand" three-act structure the way a screenwriter understands it. The model has no internal representation of dramatic intention or emotional arc. What it has are statistical associations: it has observed that text labeled as "Act 2" tends to contain conflict escalation; that the phrase "all was lost" tends to precede a climactic reversal; that short chapters tend to cluster near endings in thrillers. These patterns are real and useful — but they are correlational, not causal.
This distinction matters practically. When you ask a model to "add more tension to this scene," it will add language associated with tension — shorter sentences, more dialogue interruptions, physical sensation descriptors — because those features co-occur with high-tension scenes in its training data. Whether the resulting scene is actually more tense for your specific reader, given the context of your specific story, is a question the model cannot answer. That judgment belongs to the author.
Researchers at Stanford's Center for Research on Foundation Models (CRFM) published work in 2023 noting that when prompted to generate stories, leading LLMs defaulted to "statistically average" narrative trajectories — competent, conventional, and structurally predictable. Deliberately unusual structures, such as the nested chronology of Cloud Atlas (Mitchell, 2004) or the second-person present-tense of Bright Lights, Big City (McInerney, 1984), required extensive explicit prompting to approximate, and the resulting outputs still tended to drift toward convention over longer generations.
AI can tell you where the statistical center of your genre's structure is. Knowing where the center is — and then deciding deliberately whether to occupy it, push against it, or abandon it entirely — is the author's job. The tool that maps the territory does not make the choice of where to walk.
Despite the caveats above, AI tools offer real structural leverage in specific, bounded tasks. Outline generation is the most obvious: a model can rapidly produce multiple structural variations on a story premise, giving a writer several conventional architectures to react against or select from. Beat-sheet analysis — asking the model to identify which structural beat a given scene occupies — can surface gaps a writer has missed after too long inside their own manuscript. Pacing diagnostics are useful: feeding chapter-by-chapter word counts and asking the model to comment on structural rhythm is a quick way to identify sections that may be under-dramatized.
In 2023, screenwriter John August (credited on Big Fish, Go, and multiple Charlie's Angels installments) published several posts on his Scriptnotes podcast blog documenting his use of Claude and GPT-4 for structural work. His consistent finding: AI was most valuable early in development, when generating options cost-free, and least valuable later, when the specific idiosyncrasies of a project made generic structural advice counterproductive. That rough division — early generation, later refinement requires human judgment — has become a rough working heuristic in many writers' rooms experimenting with these tools.
You have a story premise: A forensic accountant discovers that her firm has been laundering money for a foreign government — and that her supervisor, whom she admires, is fully complicit.
Ask the AI to outline this premise using two different structural frameworks from Lesson 1 — for example, three-act structure and Harmon's story circle. Then ask it to identify where the frameworks differ most. Notice what the AI gets right and where its suggestions feel generic.
When Pixar developed Inside Out (released June 2015), the story team — led by Pete Docter — spent nearly four years wrestling with a single structural problem: how do you dramatize internal emotional change so that an audience can see it? Their solution was to externalize the protagonist's psyche entirely, making abstract emotional states into named, visible characters who physically alter the architecture of Riley's mental world as she changes. The film's structure is inseparable from its character arc because the character arc is the architecture. Pixar's 2012 internal storytelling guide, later published as a series of tweets by story artist Emma Coats, articulates this as rule two: "You gotta keep in mind what's interesting to you as an audience, not what's fun to do as a writer. They can be very different."
AI tools, asked to develop a character arc, reliably produce the external symptoms of change — the protagonist becomes braver, kinder, or wiser by the end — without the structural mechanism that makes change feel earned. The difference between stating a character has changed and engineering the conditions that make change inevitable is exactly where craft lives.
A plot arc traces external events: what happens in the world of the story. A character arc traces internal change: how a character's beliefs, values, or understanding of themselves shifts as a result of those events. The two are distinct but ideally interdependent — the external events should create pressure that makes the internal change not just possible but necessary.
Screenwriting teacher Michael Hauge, who has consulted on projects for Will Smith's Overbrook Entertainment and works with Warner Bros., describes the character arc as movement from "living in a wound" (the character's false belief about themselves or the world) to "stepping into essence" (abandoning that false belief in favor of authentic identity). The wound is usually established in the first act; the essence is claimed in the third. The second act is the sustained pressure that makes abandoning the false belief unavoidable.
Literary theorist James Wood, in How Fiction Works (2008), makes a related distinction between flat and round characters (borrowed from E.M. Forster's Aspects of the Novel, 1927): flat characters are defined by a single consistent trait; round characters are capable of surprising us in ways that feel true. Character arc is essentially the mechanism by which a flat character becomes round over the course of a narrative — or reveals they were more complex than initially presented.
When prompted to develop a character arc, most LLMs default to a "wound → growth → healing" template that closely mirrors the Hauge model — because that model is heavily represented in screenwriting instruction the models were trained on. The output is structurally correct and nearly always too smooth: the character changes in linear, predictable steps without resistance, regression, or the messy ambivalence that makes fictional change feel real.
The practical craft challenge of character arc is making internal change externally visible without making it on-the-nose. Successful writers solve this through behavioral indicators: specific actions the protagonist takes (or fails to take) that imply internal state without stating it. When Breaking Bad's Walter White (2008–2013) transforms from chemistry teacher to drug lord, creator Vince Gilligan and his writers tracked the arc through concrete behavioral markers: when Walter starts lying without hesitation, when he stops flinching at violence, when he uses his cancer diagnosis strategically rather than suffering it. Each marker is a structural beat serving the character arc, not just the plot.
AI tools are poor at inventing specific behavioral indicators because doing so requires deep knowledge of a particular character and story world. They are more useful for a different task: arc auditing. Given a synopsis or scene list, an LLM can identify where behavioral markers are absent, where a stated change lacks supporting scenes, or where the timeline of change feels compressed. Think of it as using the tool as a structural checklist reader rather than a creative generator.
Invent the arc yourself. Define the wound, the want, the need, and the behavioral markers that track the change. Then use AI to audit: ask it whether each marker actually appears in your scene list, whether the pacing of change feels earned, whether the moment of grace is structurally placed. That division — invention to the human, auditing to the machine — tends to produce the most useful AI-assisted structural work.
Describe a character arc you want to develop — or use this provided one: Maya is a public defender who believes the justice system, despite its flaws, is fundamentally salvageable. By the end of the story, she has witnessed enough systemic failure to realize the institution is beyond repair — and must decide whether to stay inside it anyway.
Ask the AI to audit this arc: Where is the wound? What is the want versus the need? What behavioral markers would show the shift at each act break? What is missing from a structural standpoint?
In 1977, editor Verna Fields described her approach to cutting scenes for a American Film magazine interview: "I always find where the scene wants to end, and then I cut it before that." The instruction sounds paradoxical until you understand what she means. Every scene has a natural resting point — the moment of final emotional resolution — and cutting before that point leaves the audience carrying unresolved tension into the next scene. The technique is as old as film editing itself; Fields applied it to work ranging from Jaws (1975) to Paper Moon (1973), for which she won the Academy Award for film editing. The principle she identified is structural: scenes are not complete units that resolve themselves. They are partial, pressurized containers for dramatic energy that should leak into the scenes that follow.
AI-generated scenes have the opposite problem. They resolve. They tidy up. They reach a resting point and then linger there, making sure the emotional content is fully absorbed before the prose moves on. The compression instinct — the writer's ear for when a scene has done its work and should end — is precisely the craft skill that LLMs lack most conspicuously.
A scene is the smallest structural unit capable of advancing both plot and character simultaneously. It is defined by a unity of time and space (or a continuous dramatic action), a goal the point-of-view character is pursuing, an obstacle preventing achievement of that goal, and an outcome that changes the story's situation — either achieving, partially achieving, or definitively failing to achieve the goal, often while creating a new complication.
This is the MRU model — Motivation, Reaction, Unit — developed by writing instructor Dwight Swain in Techniques of the Selling Writer (1965) and later expanded by Jack Bickham into Scene and Structure (1993). In this model, scenes alternate with "sequels" — shorter passages of reflection and decision that bridge scenes by processing their emotional aftermath before the character pursues the next goal. The distinction matters for AI use because LLMs tend to compress sequel material into scene material (making characters reflect and decide within the dramatic action rather than giving that processing its own space) or to expand scene material into sequel material (over-explaining what just happened).
Janet Burroway, in Writing Fiction: A Guide to Narrative Craft (now in its eleventh edition), describes the scene's essential transaction as: "Come in late, leave early." Enter the scene after the preliminary business; exit before the aftermath settles. This compresses dramatic time, forces the reader to fill in ellipses with their own inference, and maintains forward pressure.
When generating scenes, LLMs default to over-explanation and resolution. A scene in which two characters disagree will typically end with explicit acknowledgment of the disagreement ("They both knew there was more to say") rather than a sharp cut to the next situation. A scene in which a secret is nearly revealed will usually complete the nearly — the secret gets revealed, the reaction is shown, the aftermath is described. The writer's craft is often the choice to cut before that completion.
The most productive use of AI in scene work is not generation — it is diagnostic interrogation. Once you have drafted a scene, you can ask an LLM to identify: What is the scene goal? Does the outcome shift the story's value charge? Where does the scene arrive at its first resting point — and is the scene continuing past that point unnecessarily?
Playwright Annie Dorsen, who has extensively documented her use of algorithmic tools in theatrical construction (her 2010 piece Hello Hi There used chatbot-generated dialogue between representations of Michel Foucault and Noam Chomsky), notes that AI generates text with "no skin in the game" — it does not experience the tension it is describing, so it cannot feel when a dramatic beat has exhausted its energy. Human writers feel this, imprecisely but reliably, as an instinct that the scene is done. That instinct is not replicable by language modeling.
Where AI can help: generating multiple versions of a scene's exit line, so the writer can feel which cut lands hardest. Producing alternative dialogue options for a key confrontation, so the writer can identify which version carries the most subtext. Checking that a scene goal is actually stated or implied in the scene's opening beats. These are narrow, bounded tasks where the tool's pattern-matching capability serves the writer's judgment rather than attempting to replace it.
After writing a scene, ask two questions: What does the protagonist want at the start? How has the story's situation changed by the end? If the second answer is "it hasn't changed much," the scene may not be doing structural work — or the scene's value shift is buried and needs surfacing. AI can help you ask these questions systematically across a full manuscript.
Paste or describe a scene you've drafted (or use the sample below). Ask the AI to diagnose it: What is the scene goal? Does the outcome shift the value charge? Does the scene exit before its first resting point, or does it continue past it?
Sample scene to diagnose: Two colleagues — one about to be fired, one unaware — are having a routine project meeting. The one being fired has just received the termination call but hasn't yet told anyone. They discuss next quarter's deadlines.
In 1961, novelist Flannery O'Connor gave a lecture at a Georgia college in which a student asked her to explain the meaning of her story A Good Man Is Hard to Find. O'Connor declined to summarize and instead read a passage aloud again. Later, in her essay collection Mystery and Manners (published posthumously in 1969), she wrote: "The meaning of a story should go on expanding for the reader the more he thinks about it, but the writer's moral sense must coincide with his dramatic sense." The formulation is precise and demanding. Theme is not a message extracted from a story after the fact — it is the convergence of the author's moral intuition and their structural choices, present in every scene-level decision. A Good Man Is Hard to Find does not state a theme about grace and violence; it is built so that grace and violence are structurally inseparable, enacted through the plotting itself.
An AI language model has no moral sense, no dramatic intuition born from felt experience, and no converging purpose behind its word choices. It can reproduce the surface patterns of thematic fiction. It cannot intend.
Theme is the narrative's organizing moral or philosophical inquiry — not a statement ("war is hell") but a question explored through dramatic action ("What does loyalty cost when the cause is wrong?"). The distinction between theme-as-statement and theme-as-question matters because a statement forecloses exploration; a question drives dramatic structure by requiring the story to test it across multiple situations before the ending provisionally answers it.
Literary critic John Truby, in The Anatomy of Story (2007), argues that theme is the story's "moral argument" — a debate between competing positions that the narrative weighs through the fates of its characters. The protagonist argues one position through their choices; secondary characters argue competing positions through theirs; the story's outcome adjudicates between them. This is not a neutral structural move — it reflects the author's actual beliefs about the human situation under examination.
Subtext is theme made invisible — encoded in image, behavior, and dialogue that means more than it literally says. Iceberg theory, as Ernest Hemingway described it in Death in the Afternoon (1932): "The dignity of movement of an ice-berg is due to only one-eighth of it being above water." A story's explicit content is the eighth above the surface; subtext is the seven-eighths below that give the visible portion its weight and stability. Subtext requires authorial intention to create — the writer must know what the iceberg's bulk contains in order for the visible tip to carry that weight.
An LLM can write a story in which a character suffers for their greed. It can write one in which greed is rewarded. It can write one in which the outcome is ambiguous. It will do whichever of these you request, with equal facility and equal indifference. The choice between them is a moral choice — it reflects a position on what the story believes about human greed. AI has no position. Theme requires a moral stake the model structurally cannot have.
The limits here are significant and worth stating plainly. AI can identify the explicit thematic content of a passage ("this scene appears to be about betrayal"). It can note when a stated theme is not supported by plot outcomes ("the story says loyalty matters but rewards characters who abandon it"). It can generate thematically consistent dialogue when given an explicit thematic brief. These are genuine, bounded contributions.
What AI cannot do is discover a theme through the act of writing — the process by which a human author, working through a story across months or years, finds that their preoccupations have organized themselves into a coherent moral inquiry they did not fully anticipate at the outset. The novelist Toni Morrison described this in a 1998 interview with The Paris Review: "I've never written a book that began with a moral — it would be too thin, too programmatic. What I start with is an image, a question, an intuition. The book teaches me what it's about." That process — drafting as moral discovery — is inaccessible to a tool that generates text without a self doing the discovering.
In practice, this means the most productive thematic use of AI is thematic consistency checking: once you know what your story is about, ask the model to flag scenes, images, or character outcomes that work against that theme. It can also help you build out an image system: if you identify a core thematic image (water, doors, windows), the model can suggest additional instances where that image could appear in your scene list. These are editorial tasks, not authorial ones — and the distinction holds throughout this course.
Structure can be analyzed. Character arcs can be audited. Scene value shifts can be diagnosed. But theme — the story's moral argument, its controlling idea, the question it genuinely cares about — must originate with a human intelligence that has something at stake in the answer. AI is a powerful tool for building the vessel. The story that matters is what you put inside it.
Define a theme for a story — your own, or the one provided below. Then describe a scene list or a set of plot events. Ask the AI to identify which elements support the theme and which work against it.
Sample theme and outline: Theme: "Belonging cannot be purchased — it must be grown through shared risk." Story events: (1) Wealthy newcomer buys his way into a small community's social circle. (2) He funds the community center renovation, earning gratitude. (3) He stays silent when a long-standing member is falsely accused of stealing from the fund. (4) The community rallies around the accused. (5) The newcomer is ultimately welcomed into the inner circle by the people who defended each other without him.