How to Build Coherent AI Video Stories With Director Prompts

Two creators type the same sentence into the same generative video model. One gets a gorgeous but meaningless eight-second clip. The other gets a scene that feels like it belongs to a film. The model is identical. The difference is direction — the deliberate, repeatable craft of deciding what the camera sees, in what order, and why.

That craft is what this guide is about. Not a single tool, not a magic prompt, but a workflow you can run again and again: plan like a director, generate like a cinematographer, assemble like an editor. If you have ever generated a beautiful shot and then realized you cannot connect it to anything else, this is the missing layer.

Why Coherence Breaks Before Quality Does

Most AI video problems are not rendering problems. They are continuity problems. A character's jacket changes color between shots. The sun jumps from left to right. A wide shot establishes a rainy street, then the close-up has dry pavement. The audience may not consciously notice, but they stop believing the scene. Once belief breaks, no amount of resolution or cinematic bokeh can repair it.

Coherence operates on four independent levels, and each one fails differently:

Character coherence — the same face, body, wardrobe, and age across every shot.
Spatial coherence — consistent geography, screen direction, and eyelines so viewers can build a mental map.
Tonal coherence — the same palette, contrast, grain, and lens language throughout a sequence.
Narrative coherence — each shot adds information, escalates tension, or pays off something set up earlier.

Generative models are excellent at the first frame and indifferent to the second. They do not know that the character blinked in the previous shot, or that a door was on the left. Direction is the process of supplying that memory externally — in documents, references, prompt structure, and edit decisions — so the model only has to solve one problem at a time.

This is why a storyboard-first workflow consistently outperforms a prompt-first one. The storyboard is not decoration; it is the model's context.

Think Like a Director: Shots, Beats, and Intent

A director does not think in "cool images." A director thinks in coverage: what the audience needs to see, feel, and understand at this exact moment, and what they should be denied for now.

From script beats to shot list

Start with a scene written as beats, not dialogue. A beat is a change: someone decides something, learns something, or loses something. A three-beat scene might be: a courier enters a silent archive (beat one), notices a drawer already open (beat two), hears footsteps behind her (beat three).

Now translate each beat into one to three shots. The rule of thumb: one beat, one visual idea. If a shot is trying to communicate two ideas, split it.

The twelve-shot spine of a short scene

For a 30–45 second sequence, this structure is remarkably durable:

Establishing wide — where and when.
Medium entrance — who and their state.
Insert detail — the object that matters.
Reaction close-up — how it lands emotionally.
Over-the-shoulder — relationship between character and object.
Movement shot — character commits to an action.
Obstacle shot — something blocks or complicates.
Escalation — tighter framing, faster motion.
Turning point — a reveal, often a push-in or a hard cut.
Consequence — the character's new state.
Breath — a quiet wider shot to let the audience settle.
Exit — a final image that closes the loop opened in shot one.

You will not always use twelve. But having the spine prevents the most common failure in AI video: a sequence of unrelated pretty shots with no escalation.

Coverage ratios that keep editing flexible

Generate more than you need, but generate it strategically. A useful ratio for AI work is roughly 50% of shots at the intended framing, 25% tighter, and 25% wider or from an alternate angle. This gives you material to fix pacing in the edit without regenerating an entire scene — the single biggest time sink in AI video production.

Build a Style Bible Before You Generate Anything

If the shot list is the skeleton, the style bible is the skin. Create one document per project and treat it as law. Every prompt you write later should be traceable back to it.

What belongs in the style bible

Palette — three to five named colors with hex values if you have them. "Amber streetlight, wet slate, oxidized copper" is far more useful than "moody."
Lighting philosophy — motivated practical light, high-key, single-source, or hard-edged noir.
Lens language — 24mm for geography, 50mm for neutrality, 85mm for intimacy. State it so your prompts stay consistent.
Texture and grain — clean digital, 16mm grain, anamorphic flares, or a soft analog haze.
Movement vocabulary — locked-off, slow dolly, handheld drift, crane reveal. Pick two or three per project and stay inside them.
Reference frames — six to ten still images that represent the target look.

Character sheets that survive camera changes

For each recurring character, document physical specifics that a model cannot invent consistently: hair length and part, eye color, distinguishing marks, wardrobe layers, and one silhouette-defining accessory. Then generate a character sheet — a single image or a small grid showing the character in neutral light, from front, three-quarter, and profile.

Use that sheet as an image reference for every shot the character appears in. When a model supports reference images or subject conditioning, the sheet becomes your continuity anchor. When it does not, restate the character's description word-for-word in every prompt rather than paraphrasing. Small wording changes produce large face changes.

Locations, props, and wardrobe continuity

Create one establishing reference image per location and reuse it as a visual anchor. Note the direction of the light source, the position of major furniture, and the time of day. Then write a short continuity log as you generate: which shots exist, what time of day they are, which wardrobe state the character is in, and which props are present. This log catches contradictions before the edit does.

The Five Layers of a Shot Prompt

Most prompts fail because they blend five different jobs into one sentence. Separate them into layers, and both consistency and controllability improve dramatically.

Layer 1 — Subject and action

Who is in frame and what are they physically doing right now. Use present-tense, observable verbs: she lifts the latch, he turns toward the window. Avoid interior states the camera cannot see: she feels betrayed. Translate emotion into behaviour — a clenched jaw, a half-step back, a hand closing around a strap.

Layer 2 — Framing and camera

Shot size, angle, and lens. "Medium close-up, eye level, 85mm, shallow depth of field" gives the model a much narrower target than "cinematic shot." If you want a specific move, describe it as a camera instruction: slow push in, lateral tracking left, static locked-off frame.

Layer 3 — Light and color

Name the source and the quality. Single warm practical lamp from frame right, deep shadow on the opposite cheek, cool ambient spill from a window. This layer does more for perceived production value than any other, and it is the one most often omitted.

Layer 4 — Motion and pacing

AI video models interpret motion cues differently. Being explicit about speed and restraint helps: subtle movement only, cloth shifts gently, no camera shake. If you want energy, state the direction and tempo: quick whip pan right, motion blur, 0.5 seconds.

Layer 5 — Continuity anchors

Every prompt should end with a short block that pins the things most likely to drift: hair, wardrobe, key prop, time of day, palette, grain. Repeating this block verbatim across every shot in a scene is the single highest-leverage habit in AI video work.

A full example, assembled:

Medium close-up, eye level, 85mm, shallow depth of field. A woman in her thirties with shoulder-length dark hair tucked behind her left ear, wearing a charcoal wool coat over a rust-colored scarf, lifts a brass latch with her right hand. Single warm practical lamp from frame right, deep shadow on the opposite cheek, cool ambient spill from a window behind her. Subtle movement only, cloth shifts gently. Rain on glass, wet slate and oxidized copper palette, 16mm grain.

Note how much of the prompt is shared with the next shot in the scene. Only the subject-and-action and framing layers change. That is the point.

Choosing Tools by Shot Type, Not by Hype

The generative video field moves quickly, and the honest answer is that no single model wins every shot. Professional AI workflows are usually multi-model. What matters is a clear rule for which tool handles which job.

Dialogue and performance shots

Performance is the hardest problem: lip sync, micro-expression, subtle eyeline. Look for models with strong identity preservation, reference-image conditioning, and reliable mouth-shape accuracy. Where a model struggles with speech, generating a silent performance and layering audio separately in post is often more convincing than forcing a talking-head generation.

Stylized and illustrative sequences

For animation-adjacent looks, painterly styles, or graphic sequences, choose models that hold a stylized render without collapsing into realism. Test consistency the same way every time: generate the same character in three different framings and compare the face and line weight. If the style drifts across framings, that model is a poor anchor for the project.

Action and motion-heavy shots

Physical motion, cloth simulation, and camera movement are a separate skill set. Look for strong temporal stability and the ability to respect an explicit camera instruction. Short generations stitched together frequently beat one long generation, because you can hide the hardest transitions on a cut.

A simple selection matrix

Shot need	Priority quality	Practical tactic
Character close-up	Identity stability	Reference image plus locked description block
Establishing wide	Spatial plausibility	Generate multiple takes, choose for geography
Action beat	Motion clarity	Short clips, cut on movement
Stylized sequence	Style retention	Test across three framings first
Insert detail	Texture realism	Macro framing, minimal subject motion

Build the matrix once per project, then stop re-litigating tool choices mid-scene. Switching models mid-sequence is one of the fastest ways to destroy tonal coherence.

Worked Example: A Three-Scene Short From Scratch

To make the workflow concrete, here is how a 60-second piece comes together end to end.

Premise. A night-shift lighthouse keeper finds a message in a bottle that is dated tomorrow.

Pre-production. Write the premise as three scenes of three beats each. Scene one: routine, discovery, hesitation. Scene two: reading the message, realization, a decision. Scene three: action at the shore, the reveal, the quiet aftermath. Draft a style bible: cold blue ambience, one warm practical source, wet surfaces, 35mm grain, slow dolly and locked-off frames only.

Assets. Generate a character sheet for the keeper: weathered face, grey stubble, dark knit cap, olive raincoat. Generate two location references — the lamp room interior and a stony shore at dawn. Save both as anchors.

Shot list. Roughly 14 shots, allocated: 7 at intended framing, 4 tighter inserts, 3 wider alternates. Every prompt ends with the same continuity anchor block naming palette, grain, and wardrobe.

Generation. Generate the lamp room wide first, because it sets geography. Then coverage of the keeper. Then inserts: hands on a brass rail, the bottle on wet stone, the paper unfolding. Do not generate the climactic shot first; you will have no context for it and will likely regenerate it later anyway.

Assembly. Cut to a scratch music bed, drop in placeholder dialogue, and find the rhythm before polishing anything. Most scenes lose one or two shots at this stage. That is normal and it is why coverage matters.

Polish. Fix colour across all clips in one grade, add sound design (wind, gulls, distant water, the creak of the lamp mechanism), and layer in the score. The score often does more for continuity than the grade does.

The Edit Is Where Coherence Is Won

Editing is not the cleanup phase for AI video. It is the phase where most continuity illusions are actually created.

Match cuts, eyelines, and screen direction

Keep characters looking in consistent directions. If a character looks frame right in shot A, their object of attention should be frame left in shot B. Breaking this rule disorients viewers even when they cannot say why. Use match cuts on shape, motion, or colour to bridge shots that were generated in completely different sessions.

Cut on movement

When two clips do not match perfectly, cut on the peak of an action — a hand swinging, a head turning, a door closing. The motion masks the discontinuity. This is one of the oldest tricks in film editing and it works just as well on generated footage.

Sound as continuity glue

A continuous ambient bed across a sequence quietly tells the audience that these shots belong together. Lay a room tone or environmental loop under the entire scene, then place discrete sounds on cuts. Even a 300-millisecond audio overlap across a hard cut can make two mismatched clips feel like one space.

Colour grading across heterogeneous clips

If you generated shots with different tools, they will have different contrast curves and colour temperatures. Rather than matching each clip individually by eye, apply a single look — a curve, a slight desaturation, a grain layer — over the whole sequence. A unifying grade is often more effective than per-shot correction because it creates a deliberate visual signature.

Nine Mistakes That Kill the Illusion

Rewriting the character description between shots. Paraphrasing changes faces. Copy and paste instead.
Changing lens language mid-scene. If the scene is 85mm intimate, do not insert a 16mm wide unless it is motivated.
Forgetting the light direction. Two shots with opposite key light read as two different places.
Generating the hero shot first. You will not know what it needs to connect to.
Using one long generation where three short ones would cut better. Long clips are harder to control and harder to fix.
Ignoring screen direction. Eyelines and movement directions are the audience's map.
Grading each clip separately. Look for a unified look instead.
Skipping sound design. Silence exposes discontinuity; ambience hides it.
No continuity log. Without a written record, you will contradict yourself by shot thirty.

Iterating Without Drift: Version Control for Prompts

Consistency collapses during revision, not during initial generation. The moment you start "just tweaking" prompts, drift begins. A few habits prevent it.

Version your prompt blocks. Keep the continuity anchor block in a separate text file and paste it in. Never retype it.
Change one layer at a time. If a shot is wrong, decide whether the problem is framing, light, motion, or subject — then change only that layer.
Freeze approved shots. Once a shot is approved, stop regenerating it, even if a new model tempts you. Regenerating approved shots is the most common cause of late-stage inconsistency.
Name files by scene, shot, and take. s02_sh07_take3 keeps your edit organised and your sanity intact.
Log what changed. One line per iteration — what you altered and what it fixed — turns guesswork into a repeatable process.

FAQ

Do I need a script before generating AI video?

You need beats, not a screenplay. A one-page outline of three scenes with three beats each is enough to begin. The structured document matters more than its length because it defines what each shot is for.

How do I keep a character's face consistent across shots?

Generate a character sheet, use it as a reference image wherever the tool supports it, and reuse an identical description block word-for-word in every prompt. Consistency comes from repetition, not from better adjectives.

How long should each generated clip be?

Shorter than you think. Five to ten seconds per shot is usually plenty for narrative work, and shorter clips are easier to control, easier to fix, and easier to cut. Let the edit create the sense of duration.

Can I mix multiple video models in one project?

Yes, and most experienced creators do. The catch is tonal drift. Choose one look, apply one grade, and keep a continuous sound bed so the joins disappear.

What is the fastest way to improve my results?

Write a shot list before you write prompts. Nearly every coherence problem in AI video traces back to generating images before deciding what the sequence needs to say.

How many takes should I generate per shot?

Three to five for key shots, one or two for inserts. Generate the establishing shot and the emotional turning point most heavily — those are the two shots audiences actually remember.

A Practical Checklist to Start Today

Before your next prompt, do these seven things in order. It takes about ninety minutes on a first project and much less afterwards.

Write three scenes, each with three beats.
Convert the beats into a shot list with framing noted per shot.
Draft a one-page style bible: palette, lighting, lens, grain, movement.
Generate one character sheet and one location reference per major setting.
Write a continuity anchor block and reuse it verbatim.
Generate in coverage order — establishing, then coverage, then inserts, then the hero shot.
Edit to a scratch track, add ambience before effects, and apply a single unified grade.

The tools will keep changing. Models will get better at faces, motion, and duration. What will not change is the underlying discipline: decide what the audience needs to see, supply the model with memory it does not have, and cut with intent. Direct your story first, generate second, and the coherence problem largely solves itself.