How to Use AI to Write Chapter Openings That Drop Mid-Action
There's a specific kind of disappointment that comes from asking an AI to write a chapter opening and getting back three sentences of weather description followed by a character checking their watch. It's not that the writing is bad, exactly. It's that it's safe. The AI has done what AI tends to do: it started at the beginning.
In medias res is one of those techniques that sounds simple until you actually try to prompt for it. Drop the reader into action. Start mid-scene. Sure. But when you hand that instruction to an AI without a framework, you usually get one of two failure modes: either a generic action beat that feels disconnected from everything (someone is running, but why should we care?), or the AI quietly adds a paragraph of setup before the action anyway, because that's how it learned stories begin.
The fix isn't just adding "start in medias res" to your prompt. It's understanding why AI defaults to setup, how to use what I call the disorientation dial, and how to build prompts that force the AI to do the hard work of orienting readers through action itself rather than through explanation.
Why Most AI-Generated Openings Default to Setup (And What to Do Instead)
AI models are trained on enormous amounts of fiction, and the statistical truth of that fiction is: most scenes begin with grounding. A location. A character. A situation. The model has internalized this rhythm so deeply that even when you ask for in medias res, it often interprets that as "action-flavored setup" rather than a genuine mid-scene drop.
The other culprit is context anxiety. The AI doesn't know what you know about your characters and world, so it hedges. It front-loads information to make sure the reader (and itself) can follow along. The result reads like a news report — who, what, where, when — rather than a scene you stumbled into.
The key insight: AI doesn't know what your reader is allowed to be confused about. You have to tell it explicitly — and you have to give it permission to leave things unresolved.
The practical fix has two parts. First, you need to block the setup impulse directly in your prompt. Don't just ask for in medias res — tell the AI what it's not allowed to include. No establishing shots. No character introductions. No explanation of why the character is here. Second, you need to give the AI enough specific detail about what's already happening that it doesn't feel the need to invent context. Specificity is a security blanket for AI. The more concrete your situation, the more confidently it can drop into the middle of it.
That means your prompt needs to arrive already knowing things. Not "write a tense scene where two characters are arguing" but "Mara and her brother Dex are three sentences into an argument about whether to call the police — Mara has already said something she can't take back and she knows it."
The Disorientation Dial: Calibrating How Much Context to Withhold
Not all in medias res openings work the same way. A thriller chapter opening can afford to be deeply disorienting — the reader barely knows where they are, someone is bleeding, and they'll piece it together over the next two pages. A cozy mystery chapter opening that drops mid-action still needs to feel warm and accessible, with just enough strangeness to create forward momentum.
Think of this as a dial from 1 to 5. At 1, you're giving the reader almost everything: they know who the character is, roughly where they are, and what's at stake — they've just been dropped into the scene slightly before the big moment. At 5, you're giving almost nothing: a sensation, a line of dialogue, a physical action, and the reader is genuinely scrambling to understand what world they're in.
Here's a rough genre guide:
- Literary fiction: 3–4. The disorientation is often emotional or psychological rather than situational. The reader knows the physical facts but not what they mean yet.
- Thriller/crime: 3–5. High disorientation is earned because the genre trains readers to expect it. Withhold identity, location, or stakes freely.
- Fantasy/sci-fi: 2–3. World-building complexity means total disorientation can tip into frustration. Give enough physical anchor that the reader can hold on.
- Romance: 2–3. Emotional orientation matters more than situational. The reader should feel the emotional temperature immediately even if the specifics arrive later.
- Cozy/upmarket: 1–2. The contract with the reader is comfort. Even mid-action openings should feel accessible and warm.
When you prompt the AI, you can actually specify this dial directly. Telling the AI "withhold the character's name and the nature of the threat for the first paragraph — the reader should only have physical sensation and dialogue" is a concrete instruction it can follow. Telling it "this should feel disorienting but not frustrating — a 3 on a scale where 5 is completely lost" also works better than you'd think, because the model has enough context about disorientation as a craft concept to calibrate.
Prompt Architecture for In Medias Res Openings by Scene Type
Different scene types need different prompt structures. An action scene, an emotional scene, a quiet scene, and a reunion scene each have different centers of gravity — and the AI needs to know what that center is before it can drop into the middle of one effectively.
For action scenes, the center of gravity is physical consequence. Your prompt needs to specify what has already happened, what the body is doing right now, and what the immediate threat is. Don't describe the threat — give the AI the physical sensations that register before understanding kicks in. The character doesn't think "there's a sniper" — they think about the shattered glass, the ringing ear, the wrong angle of their own arm.
For emotional scenes, the center of gravity is a specific thing that has already been said or done that can't be taken back. The prompt needs to arrive at the moment after that thing — not the thing itself. The silence after the sentence. The look that happens when the door closes. If you ask the AI to write the confrontation, you'll get buildup. Ask it to write the five seconds after the worst moment, and you'll get in medias res.
For quiet scenes — a character alone, a moment of realization, something subtle — the center of gravity is a detail that means more than it should. Your prompt needs to give the AI a specific object, action, or sensation that carries weight, and specify that the character is mid-thought rather than beginning to think. She's already been looking at her mother's handwriting for a while. She's three blocks into a walk she didn't decide to take.
For reunion scenes, the center of gravity is usually a recognition beat — the moment a face resolves into someone known. The prompt needs to arrive at mid-recognition, not at the approach. The reader should be inside the scramble of memory and present reality happening simultaneously in the character's mind.
Four Prompt Examples Across Action, Emotional, Quiet, and Reunion Scenes
Here are four prompts you can adapt directly. I've written them the way I'd actually type them — specific, slightly messy, full of the context the AI needs to do the work without hedging.
Action Scene
Write the opening paragraph of a chapter. My protagonist is Sena, a 30s woman, former coast guard, currently working as a private investigator. She has just been thrown against a chain-link fence by someone she thought was an ally — a man named Cotter. The chapter should open in the moment her back hits the fence, before she's processed what's happening. Do not name Cotter in this paragraph. Do not explain why she's here or what led to this. Give me only physical sensation and the first fragment of thought that doesn't make sense yet. High disorientation — the reader should not know where they are or why this is happening. No setup sentences. No weather. Start at the impact.
Why this works: it blocks setup explicitly ("no setup sentences, no weather"), specifies what to withhold (Cotter's name, the backstory), and gives the AI a concrete physical anchor (the fence, the impact). The instruction "the first fragment of thought that doesn't make sense yet" is key — it gives the AI permission to write something partially incoherent, which is exactly what in medias res emotional texture requires.
Emotional Scene
Write the opening of a chapter set in a kitchen. My character is Marcus, 50s, a father. His adult daughter Priya has just told him she's been lying to him about her job for two years — she quit, she's been struggling, and she didn't tell him because she knew he'd be disappointed. She's said everything she's going to say. The chapter opens in the silence after she finishes. Marcus is doing something with his hands — washing a mug, maybe, or just holding it. I want the opening to be in his POV, present tense, and I want the reader to feel the weight of what was said without having it explained. Do not summarize or recap what Priya said. Do not start with Marcus processing — start with him in the middle of already processing. The emotion should be in the physical details, not stated directly.
The phrase "in the middle of already processing" is doing a lot of work here. It tells the AI to skip the moment of receiving information and go straight to the complicated interior that follows. "The emotion should be in the physical details, not stated directly" is the craft instruction that keeps the AI from writing "Marcus felt devastated."
Quiet Scene
Write a chapter opening, third person limited, past tense. Character is Nell, late 20s, in the apartment she shared with her late partner for six years. She's been living here alone for four months. The chapter opens with her noticing something small — a specific, mundane object — that suddenly means something it didn't before. I don't want you to choose something obviously sentimental like a photograph. Choose something unexpected: a kitchen item, a cable, a piece of mail. The opening should feel like Nell is mid-thought, not beginning a thought. She's been standing in this spot for a minute already. The reader should feel the grief without the word grief appearing. Three paragraphs maximum. No backstory. Start with the object.
Notice the constraint "not a photograph" — that's protecting against the AI defaulting to an obvious symbol. Giving the AI permission to choose the object while blocking the predictable choices produces something that feels more discovered than written.
Reunion Scene
Write the opening of a chapter in a busy train station. My protagonist is Felix, 40s. He hasn't seen his older brother in eleven years following a serious falling out — the details of the fallout don't appear in this scene. Felix is waiting on a platform and sees someone he thinks might be his brother, thirty feet away, moving through a crowd. The chapter opens at the moment of half-recognition — Felix isn't sure yet. I want the prose to mirror that uncertainty: the details he notices are partly about this figure and partly about his own memory scrambling to catch up. First person, present tense. Do not tell the reader who this person is. Do not have Felix call out or approach yet. Stay inside the moment of not-quite-knowing. Start mid-observation, not mid-arrival.
"Start mid-observation, not mid-arrival" is the specific instruction that prevents the AI from writing Felix arriving at the platform and scanning the crowd — the classic setup that kills the in medias res effect.
Using a Follow-Up Pass to Seed Retroactive Grounding
Here's something that took me a while to figure out: a fully disorienting opening almost always needs a second pass where you ask the AI to weave in just enough retroactive grounding to keep the reader from bouncing off the page.
This isn't the same as adding setup. Retroactive grounding means the reader gets a detail that makes sense of something they already read — so the disorientation resolves as they move forward rather than stacking. A name that appears in paragraph two. A landmark that suddenly makes the location clear. A line of dialogue that tells us the stakes without explaining them.
The prompts for this pass are simpler, but they need to be specific about what's still unresolved. Something like:
Here is the opening I've drafted [paste opening]. The reader currently doesn't know: the character's name, whether this takes place indoors or outdoors, or who the threat is. I want to keep the threat unknown for now, but I need to seed the character's name and the indoor/outdoor setting naturally into the second and third paragraphs without adding explanation or backstory. The name should arrive through dialogue or a memory fragment, not narration. Revise only paragraphs two and three.
Specifying exactly what's unresolved — and exactly what should remain unresolved — keeps the AI from over-correcting and sliding back into setup mode. It also keeps you in control of the information architecture of your chapter rather than letting the AI decide what readers need to know and when.
One more thing on this follow-up pass: ask for multiple options. "Give me three different ways to introduce the character's name in paragraph two" will produce variations that differ in tone and method, and you'll almost always find that one of them fits the voice of your opening better than anything you'd have landed on alone.
The concrete thing to do right now: take a chapter you've been stuck on — or one where the opening feels flat — and run it through the action scene or emotional scene prompt structure above, adapting the specifics. Don't use the prompts as templates with blanks to fill in. Rewrite them in your own framing, with your actual characters and the specific moment you want to drop into. The level of detail you bring to the prompt is the level of specificity you'll get back. Give the AI a half-formed situation and it'll write something generic. Give it a woman whose back just hit a chain-link fence before she's processed who threw her — and it has something real to work with.