AI Video Workflow Guide: From Shot List to Final Cut

Most first attempts at AI video fail for a boring reason: the person wrote a mood instead of a shot. Something like "cinematic woman walking through neon streets, dramatic lighting, ultra detailed" produces a clip that feels like the thumbnail of a film nobody made. The instinct is to blame the tool and go looking for a newer one.

The tool is rarely the problem. Modern video models are good enough that a weak result usually means the input was under-specified, not that the engine couldn't render it. What closes the gap is the same discipline a two-person crew uses on a real shoot: a shot list, a locked look, references that agree with each other, and a review loop that changes one variable at a time. This guide walks through that workflow in practical terms, using examples you can adapt to any contemporary text-to-video or image-to-video system.

Why a shot list beats a clever prompt

A prompt is one instruction. A shot list is a film. If you open the generator first and write prompts one after another, you end up with a pile of clips that look beautiful in isolation and refuse to cut together. Shot one has the subject facing left in warm light; shot four has them facing right in cool light; the room has changed shape twice.

The fix is to decide the sequence before you render anything. A shot list doesn't need to be formal — a table with four columns is enough:

Shot	Job in the edit	Subject action	Camera
1	Establish the space	Nobody on screen, wind moving a curtain	Locked wide, slow drift
2	Introduce the character	She sets down a box, looks at the door	Medium, static
3	Detail that carries meaning	Close on hands peeling tape	Macro, handheld
4	Land the emotion	She exhales, doesn't move	Close-up, slow push-in

That table takes five minutes and saves an afternoon. It also reveals problems early: if two adjacent shots do the same job, one of them is redundant, and no amount of rendering will fix a structural repeat.

The three questions every shot must answer

Before generating, answer these in a single sentence each:

1. Who is on screen, and what do they want in these next few seconds? Not their backstory. Their intent for the duration of the clip.
2. Where is the camera, and why is it there? A locked-off wide says something different from a slow push-in. If you can't justify the position, default to static.
3. What is this shot's job in the edit? Establish, reveal, transition, or land an emotion.

If any answer is vague, the shot isn't ready. Cutting it is often better than fixing it.

Stop writing adjective piles

Compare two prompts for the same moment:

• "Stunning cinematic masterpiece of a mechanic in a garage, dramatic lighting, hyper detailed, 8K."
• "Close-up, slightly low angle: a mechanic in his fifties, oil-stained overalls, silver stubble, wipes his hands on a rag and looks off-frame left. Camera static. Late afternoon sun through a dirty window, hard key from the left, dust in the air."

The second is longer but far easier to render, because every phrase resolves into something visible. "Cinematic" and "stunning" describe your taste, not the frame. They consume prompt space and give the model nothing to draw.

The six-slot prompt structure

Fixed structure beats clever wording. Keep six slots in a consistent order so you can change one at a time and actually learn from the result.

1. Shot type and framing — extreme wide, medium, close-up, over-the-shoulder.
2. Subject and wardrobe — age range, build, specific clothing, one distinguishing detail.
3. Action already in progress — what is happening as the clip begins, not a story arc.
4. Camera behaviour — static, handheld drift, dolly, crane, orbit, with speed.
5. Light and atmosphere — direction, quality, colour temperature, weather, haze.
6. Texture and finish — film stock feel, grain, contrast, palette, softness.

A single sentence per slot is plenty. If you find yourself writing a seventh slot, it probably belongs in the reference set instead.

Handling negatives without planting ideas

Writing "no blur, no extra fingers" often introduces the exact concept you wanted gone, because many models read the noun and skim the negation. Positive substitution works better: describe the condition you want. Instead of "no motion blur", write "crisp, static subject, sharp edges". Instead of "no crowd", write "empty platform, nobody else in frame".

For hard constraints — a costume colour, a logo-free surface, a specific lens — put them in a reference image or a structured control rather than fighting for space in prose. Constraints expressed as pictures are obeyed far more reliably than constraints expressed as sentences.

Reference sets and identity lock

Consistency is where multi-shot projects break. The face in shot one is not the face in shot six, and prompt tuning alone rarely recovers it once drift sets in.

What belongs in the folder

Assemble a small reference set before generating anything:

• One neutral front-facing portrait in even, flat light with no strong expression.
• One three-quarter view so the model learns cheekbones and jawline.
• One full-body shot showing silhouette, posture, and true colour.
• One detail crop — hands, a collar, stitching, a product edge.

Keep these references deliberately boring. Dramatic lighting in a reference becomes a permanent feature of the character, because the model cannot separate lighting from identity. A moody side-lit portrait produces a character who always looks like they're standing in a doorway at midnight, no matter where you place them.

Order of operations that avoids drift

1. Generate or photograph the references first, using a still-image tool if you don't have photos.
2. Write a short style note — wardrobe, palette, lens feel, film texture — and reuse it word for word across every prompt in the project.
3. Render the hero shot first: the most demanding framing. Iterate until it's right.
4. Use that approved frame as the reference for adjacent shots, not the original portrait. It carries the grade and the wardrobe in one image.
5. Re-check the whole set every five or six shots. Drift accumulates silently, and the moment it becomes visible it's expensive to undo.

For product work, add one rule: never describe a product in words when you can show it as an image. Text descriptions of packaging produce approximations, and approximations look wrong to anyone who knows the brand — especially on labels, where letterforms are the entire point. If you need a clean base frame to start from, the AI image generator is a faster route than fighting a video prompt.

Camera language that survives generation

Generated motion is most convincing when the camera has a reason to move. A short list of moves that hold up well:

• Static with subject motion. The safest option and often the most cinematic.
• Slow push-in. Reads as attention. Use it when a character decides something.
• Lateral truck. Good for reveals and for hiding background instability.
• Handheld drift. Adds documentary texture, but a little goes a long way; heavy shake ages badly.
• Orbit. Impressive, expensive in iterations, reserve it for hero moments.

Two things to avoid: fast whip pans, which most models resolve as mush, and multi-move shots such as "push in, then orbit, then tilt up". One move per shot. If the story needs two moves, use two shots and cut between them. A sequence of simple, confident frames will beat one clever shot surrounded by broken ones every time.

Compose for the frame you will actually deliver. Deciding between horizontal and vertical after you've perfected a composition means cropping away the work you just paid for in iterations.

Designing for sound from the first line

Models that generate audio alongside picture change how you write prompts. Specify the sonic environment in the same breath as the visuals:

• Room tone — "small tiled bathroom, slight reverb, a tap dripping off-frame".
• Dialogue delivery — give the performer a verb and a physical action: "she says it flatly, without turning her head".
• Effects — name the source, not the abstract sound: "boots on gravel", not "footstep sound".

Keep spoken lines under roughly twelve words, and make sure the clip has time to hold them. A line delivered inside a four-second shot leaves no room to breathe, and breathing is what makes dialogue feel performed rather than recited. When a line matters, give it its own shot with extra seconds of tail.

Plan captions from the start rather than retrofitting them later. If a shot depends on text being legible, treat that as a reason to overlay titles in the edit instead of asking the model to draw letterforms inside the frame — generated typography remains the least reliable element in any pipeline.

The three-pass review loop

Reviewing a clip in one pass leads to over-correcting: you start adjusting grain on a shot whose silhouette was never right. Split the review into three passes, each with exactly one question.

Pass one — shape and motion. Watch small, or squint. Does the silhouette read? Is the motion plausible? If the shape is wrong, stop and regenerate immediately. No amount of face polish saves a shot with a broken outline.

Pass two — face, hands, and props. Watch at full size and pause on the frames where the subject turns or the hands move. Check teeth, eyes, fingers, and anything the character touches. Objects that change shape between frames are a strong signal to simplify the action rather than re-roll the prompt.

Pass three — grade and texture. Only now adjust contrast, colour, and grain. Grading an unstable shot wastes effort, and grain applied over a broken frame simply makes the break harder to spot.

One rule ties the loop together: fix one variable per regeneration. Changing three things at once teaches you nothing about which one worked, and you'll repeat the mistake on the next project.

Troubleshooting: symptom, cause, fix

What you see	Likely cause	Practical fix
Face changes between shots	Reference set varies in light or angle	Rebuild references in flat, even light; reuse the style note verbatim
Subject drifts toward the frame edge	Camera move conflicts with action	Make the camera static, let the subject carry the motion
Hands melt during a gesture	Too much simultaneous action	Cut to a detail shot instead of describing the whole gesture
Background warps during a pan	Large lateral camera move	Swap to a truck move or a static frame with foreground occlusion
Dialogue sounds recited	Line too long for the clip length	Shorten to one clause, add physical action, extend duration
Product label unreadable	Typography generated inside the model	Show the product in a wider shot, overlay the label in the edit
Clip feels flat despite good framing	No light direction specified	Name the key light source and its angle

Keep this table next to your shot list. Most wasted renders fall into one of these rows, and naming the cause takes thirty seconds.

A worked scene, shot by shot

Here is how the workflow looks end to end for a thirty-second opening sequence about a woman moving into an empty apartment.

Brief. She is carrying the first box into a flat with no furniture. The sequence should feel like a held breath, not a montage of excitement. Delivery: vertical, for social.

Shot 1 — establish. Wide, locked off, camera inside the empty room looking toward the door. Afternoon light through a bare window, dust visible. No subject for the first second, then she enters frame carrying a box. Prompt slot four says static; everything else is light and texture.

Shot 2 — introduce. Medium, static, eye level. She sets the box down, straightens, and looks around the room without speaking. Room tone: a distant street, a slight echo because the flat is empty.

Shot 3 — detail. Macro on hands peeling packing tape, shallow depth of field, the box edge filling the frame. This is the cheapest shot to render and often the one that sells the sequence.

Shot 4 — land it. Close-up, slow push-in. She exhales, sits on the floor, and doesn't move. The push-in is the only camera move in the whole sequence, which is exactly why it registers.

Four shots, one camera move, one wardrobe note reused verbatim. The references are a flat-lit portrait, a three-quarter view, and a full-body frame in the same jacket. If shot three fails twice, cut it — the sequence still works with three shots, and that decision costs nothing.

Managing iterations without waste

Render in small batches of three or four variants with one variable changed across them, rather than one at a time. This gives you a comparison instead of a coin flip, and it stops the slow drift into tweaking a single clip for an hour.

Set a hard rule per shot: if four attempts don't produce something usable, the shot is wrong, not the prompt. Simplify the action, reduce the number of subjects, or replace it with a different shot that does the same job. Knowing when to abandon a shot is the single biggest time saver in AI video work.

Name files from the first render using a consistent pattern such as scene-shot-take, and keep your approved prompts in one document. The second project always goes faster than the first, but only if the first left notes behind. Browsing finished examples is also a fast way to calibrate what a given engine handles well — the prompt library is useful for that, and video templates give you a running start when the deadline is close.

Mistakes that quietly ruin a sequence

• Describing a story instead of a shot. "She realises he's gone" is a performance note, not a frame.
• Too many subjects. Two characters in a wide shot is fine; five is a lottery.
• Ignoring the first frame. The opening frame sets the tone for everything after it; a weak first frame poisons the clip.
• Mixing aspect ratios mid-project. Decide delivery format before generating, always.
• Letting the style note drift. Paraphrasing your own wardrobe description between sessions is the most common cause of a character who changes across shots.
• Chasing one perfect clip. Two good shots cut together beat one flawless shot nine times out of ten.
• Rendering before the brief exists. A clear sentence costs nothing. A wasted batch costs an afternoon.
• Forgetting sound until the end. Dialogue timing affects shot length, and shot length affects everything upstream.

FAQ

Do I need film experience to get good results? No, but you need shot discipline. Learn three things — framing, light direction, and one camera move per shot — and most of the quality gap closes on its own.

How long should a generated clip be? As short as the edit allows. Four to six seconds covers most cuts and leaves less room for drift. Generate longer only for slow atmospheric material where duration is the point.

Why does my character change between shots? Usually because the references vary in lighting or angle, or because the written description drifted between sessions. Freeze the style note and reuse it word for word.

Can I fix a single detail without regenerating everything? Sometimes, using inpainting or a tighter reference crop. More often it's faster to regenerate the shot with one variable changed and keep whichever take lands.

Should I generate audio separately? Generate it natively when dialogue or ambience is part of the performance. Layer music and detailed effects in the edit, where you can time them precisely to the cut.

How many takes should I expect per shot? Plan on three to six for a simple shot and up to a dozen for anything involving hands, text, or two interacting subjects. Batch the attempts rather than rendering one at a time.

What about vertical formats? Decide before you generate. Reframing a horizontal shot into vertical crops away the composition you spent iterations perfecting, and the model cannot help you recover it.

Is a more expensive model always better? No. The right question is which engine suits this shot — identity stability for character work, clean image-to-video for product accuracy, speed and tolerance for small artefacts for social variations, native audio for dialogue. Run a ten-second test of the same prompt on two candidates before committing a project; the alternatives hub lays out the trade-offs plainly.

Put the workflow to work

The gap between a lucky clip and a finished sequence is process: a shot list, a fixed description, references that agree with each other, and a review loop that changes one thing at a time. Pick a single thirty-second scene, write four or five shots, and run the whole loop end to end. That exercise teaches more than any list of settings.

When you're ready to practise, start in the Orelon AI video generator, borrow structures from the prompt library, and reach for video templates when you need momentum. The Orelon blog goes deeper on the techniques above, and keeping your best prompts in one document means the second project is always faster than the first.