Orelon logoOrelon
料金

Literary Storytelling Techniques for Short AI Video

2026年11月15日 · Orelon Team 著

AI動画テンプレートを見る

着想のためにコミュニティ作品をいくつか閲覧し、任意のテンプレートを開いて Orelon で作成を続けましょう。

Learn how short story craft compresses plot, character, and image into brief runtime, and how to turn those techniques into prompts, shot lists, and AI video edits.

Every weak AI video has the same autopsy report: the images were fine, the sound was fine, and nothing happened. A creator types a lovely description of rain sliding down a window, generates eight clips, strings them together, and the result looks like a perfume advert with the product removed. Rendering was never the problem. Nobody decided what the character wanted.

Short fiction is the cheapest available cure. It is the only literary form that assumes the audience is one breath away from leaving, and it has spent two centuries inventing tricks to make that breath last. Those tricks convert almost perfectly into prompts, shot lists, and edit decisions, because a short story and a short video are both exercises in surviving compression.

What follows is a working method rather than a reading list. You get the craft principles that survive translation into generated footage, a production workflow you can run this afternoon, decision criteria for matching shape to runtime, and the mistakes that flatten otherwise competent work.

Why short fiction teaches faster than a screenplay

A screenplay is an inventory. It lists everything that will eventually exist on set and quietly assumes a hundred collaborators will fill the gaps with performance, production design, and sound. A short story is pressure. It assumes attention is borrowed and must be repaid every few lines.

When you study how short-story writers compress, you learn three transferable moves: cut the approach, enter late, and let one detail carry the exposition. All three are directly useful when you have eleven clips and no room for a training montage.

A quick exercise pays for itself. Take any short story, read it once for pleasure, then read it again with a pen and list only the scenes. Not the themes, not the sentences you admired, just the scenes. Most readers find three to five. Then mark which scenes do double duty, advancing the plot while revealing character. Those are the scenes worth generating. The rest is connective tissue you can replace with a cut.

Do this with three stories and a pattern appears: the writer spends almost no time on arrival, almost all of it on the moment things stop being routine, and then stops. That rhythm is your edit.

Compression: one want, one obstacle, one turn

A minute of video holds one want, one obstacle, and one turn. Not two wants, not a backstory, not a subplot about the brother who never appears.

Two tests keep you honest. The first is the and-then test: if your beat list reads and then this happened, and then that happened, you have written a summary, not a story. The second is the therefore test: each beat should be caused by the previous one, or should interrupt it. Therefore and but create momentum. And then creates a slideshow.

Here is a compressed example. A night-shift cleaner finds a wedding ring inside a vacuum bag and must decide whether the owner's grief is hers to interrupt. That is three beats: discovery, recognition of what the object means, decision. It contains a character, a stake, and a cost, and it needs roughly nine shots.

Translated into generation terms, the turn must be a shot, not a caption. Do not prompt a scene described as a moment of moral reckoning. Prompt the hand closing around the ring, then the ring placed on a desk beside a sleeping stranger, then the empty corridor. The reckoning lives in the sequence, not in the adjective.

When a beat sheet feels thin, the fix is usually subtraction rather than addition. Remove the second character, remove the explanatory scene, remove the establishing sweep of the city. What is left must still make sense, and if it does, you have a film.

The loaded object: foreshadowing you can actually prompt

The oldest rule in short fiction is that a thing introduced with emphasis must eventually matter. This is a gift to AI filmmakers, because objects hold consistency across clips far more reliably than faces, moods, or lighting.

Choose one recurring object and give it three states: pristine, used, ruined. A glass of water that starts full, is knocked over, and later appears as a dried ring on the table. A red scarf worn, dropped, and found by someone else. A paper receipt folded, crushed, smoothed flat again.

Generate the three states as stills first with an AI image generator, lock the look, then animate each state. Working from approved frames removes an entire category of wasted iteration.

There is a second, more practical benefit. When generated faces drift between shots, the object becomes your continuity anchor. Viewers forgive a slightly different jawline. They notice instantly when the scarf changes color. Treat props as the spine of continuity and let performance be the flexible part.

One warning: the object must change hands, place, or condition. A prop that stays identical is decoration. A prop that travels is a plot.

Interior conflict on an exterior screen

Prose can write she decided. Video cannot. You need behavior a viewer can read without narration, and there are three dependable devices.

The hesitation beat: the hand reaches for the handle and stops. The substitution action: the character does something small and wrong instead of the thing they want to do, straightening a picture frame while deciding whether to speak. The withheld reaction: the camera stays on the listener instead of the speaker.

A fourth, more conceptual option borrows from writers who built entire stories around a single image: a mirror, a maze, a locked archive, a corridor that repeats. One surreal establishing shot can carry a theme that would otherwise require dialogue. In generated video this is unusually economical, because a single strange location does more narrative work than five well-lit reaction shots.

When writing prompts, describe behavior, object, and light. Emotion adjectives are nearly useless to a model. He grips the receipt until it tears is a prompt. He feels guilty is a wish.

Character in a single decision

Before generating anything, write two lines. The want, in one sentence. The flaw, in one sentence. If the flaw cannot be expressed as an action the character takes under pressure, rewrite it until it can.

A thirty-second arc has a reliable shape. Shots one and two establish the want in motion, so the viewer sees competence or routine. Shots three through six apply pressure, usually by making the routine impossible. Shots seven and eight contain the decision. Shots nine and ten show the cost, briefly, without explaining it.

Notice the arc is one decision, not a personality change. A courier who is always early chooses, once, to be late. That is a complete story. A courier who learns to be less anxious is a therapy goal, not a plot.

Dialogue economy works the same way. What a character refuses to say defines them more sharply than what they announce. If you must use words, keep them to roughly one line per fifteen seconds, and consider placing them as a text insert rather than in a mouth, because lip sync remains the most fragile element in generated footage. Silence with environmental sound reads as confidence. A poorly synced close-up reads as a mistake.

Four structures that survive generation constraints

Some shapes are naturally friendly to clip-based production, because they need few locations and tolerate slight visual inconsistency.

In medias res

Start after the decision has already been made. A woman is already packing. We learn why across four shots, and the final shot returns to the suitcase. This saves an entire setup sequence, which is often the weakest generated material because nothing is at stake yet.

The loop

The last shot rhymes with the first, but something has changed. A man enters a lift, descends, and exits into the same lobby, except the plant is now dead. Loops are cheap to produce because first and last shots can share a location, and they feel intentional even when generation is imperfect.

The frame

An older character opens the story and closes it, and the middle is the memory. This is the easiest way to add scale to a small runtime, and it lets you use two distinct visual registers, which masks continuity gaps between generations.

The turn that re-reads the opening

The final shot changes the meaning of the first. The friendly neighbor was never friendly. This costs one additional shot and pays off the entire piece. It only works if the opening contains a specific detail worth reinterpreting, which is why the loaded object matters so much.

A repeatable workflow from page to final cut

The one-page reduction. Write the story in prose first, badly and quickly, in under four hundred words. Then reduce it to ten sentences, one per shot. If a sentence cannot be photographed, it is not a shot.

The shot bible. For each shot, note four things: who is in frame, what changes, camera distance, and light direction. Keep style descriptions out of this document. Style belongs in one block at the top so it can be repeated verbatim in every prompt.

Stills before motion. Generate two or three frames per shot as images. This is far cheaper than iterating on video and it exposes composition problems early. Approve the frames, then animate only the approved ones with the AI video generator. Anything with a precise prop or a precise composition belongs in this pipeline.

Prompt the beat, not the beauty. Each video prompt should name the action and the camera behavior, in that order. A hand lifts the ring toward window light, slow push in, shallow depth of field beats a cinematic emotional close-up of grief. Reusable structures for this live in the prompt library, and starting from a template is often faster than facing a blank field.

Cut to the sentence. In the edit, every shot must earn its place against the ten-sentence script. If a clip is beautiful but neither advances nor interrupts, delete it. Short narrative video is mostly a deletion discipline.

Direct prompting, for contrast, suits motion, weather, crowds, and texture, where exact framing matters less than energy. Use it for transitions and atmosphere, not for the beat that carries the decision.

Mistakes that flatten short narrative video

The most common error is writing a mood instead of an event. Melancholy streets, neon reflections, slow-motion rain: these are ingredients, not stories. Ask what happened, not what it felt like.

Second is character inflation. Two characters is a crowd in a forty-second piece. If a third appears, they are probably a prop with a face.

Third is aesthetic prompting with no beat. If every prompt describes the look and none describes the change, you will get a fashion film without a plot.

Fourth is skipping continuity anchors. Decide in advance which three elements must stay identical across every clip, usually one garment, one object, one location detail.

Fifth is letting the twist be the whole story. A reversal without a character who wanted something is a magic trick.

Sixth is over-length. A story that needs sixty seconds stretched to three minutes reads as padding, and generated padding is especially visible because the model has no new information to work with.

Seventh is treating music as structure. A track can carry a sequence that already works; it cannot rescue one that does not.

Matching story shape to runtime

For six to ten seconds, use a single image plus a reversal. One shot, one turn, no dialogue.

For fifteen to thirty seconds, use the decision story: establish a routine, break it, show the choice, end on the cost.

For forty-five to sixty seconds, use in medias res with one turn and one loaded object. This is the sweet spot for most social platforms.

For ninety seconds to three minutes, use the frame or the loop. Both allow a change of visual register, which keeps attention while forgiving minor continuity drift.

If you are adapting a text you did not write, confirm the rights position for your territory and platform before you generate anything, and keep a record of where each element came from. If you are adapting your own prose, cut it by seventy percent before writing a single prompt. The remaining thirty percent is your film.

FAQ

How many shots should a one-minute AI video contain? Eight to twelve. Fewer feels static; more forces clips so short that each action becomes unreadable.

Do I need a full script before generating? No, but you need a ten-sentence beat sheet. Generating without one is the single most common cause of unusable footage.

Can AI keep a character consistent across shots? Partially. Consistency improves with a reference still, fixed wardrobe, and a recurring object. Design your story so that perfect facial continuity is not load-bearing.

Should I include dialogue? Only if you can commit to close-ups and short lines. Otherwise design for silence, environmental sound, text inserts, or a brief voiceover.

What if a generated clip does not match the beat? Change the beat or change the clip, but do not keep both. Beautiful footage that does not serve the sentence is the fastest way to lose a viewer.

Is it better to animate stills or prompt video directly? Stills first for anything with a precise prop or composition; direct prompts for motion, weather, and crowd energy.

How do I know the story is finished? When the last shot would mean something different if the first shot were missing. That is the test, and it rarely fails.

Build the story before the shot list

Every good minute of generated video contains a decision, a cost, and one object that changes hands. That is the whole inheritance from the short story, and it costs nothing to apply. Write the ten sentences. Name the want and the flaw. Choose the object. Then let the visuals follow the structure instead of hoping structure emerges from the visuals.

Start with a single scene on the Orelon video generator, and if you want more craft breakdowns before committing to a shoot, the Orelon blog has practical guides on pacing, prompting, and editing. The cinematic ideas are yours. Give them a shape worth watching.