Start with the delivery spec, not the tool
Most stalled AI video projects begin in a browser tab rather than in a brief. Someone types a prompt, gets an arresting result, and only then tries to build a story around it. That approach is fine for testing a tool and disastrous for delivery, because nothing in the process knows what the finished piece has to be: how long it runs, where it plays, who approves it, or what sound it needs.
The delivery spec is a one-page document written before the first generation. It contains the aspect ratio, the runtime, the placement, the tone, the audio plan, the approval owner, and any legal constraints such as whether real people, brand marks, or licensed music may appear. It is short, and it prevents the most expensive habit in this discipline: generating clips that cannot be used because the frame or the duration was decided too late.
Frame shape decides composition
A vertical frame pushes subjects close to camera and makes wide establishing shots nearly pointless; it rewards faces, hands, and texture. A wide cinematic ratio gives room for landscape and negative space and punishes small subjects. Square frames are forgiving for product work and awkward for lateral movement.
If you need several formats, choose one strategy and commit. Either generate natively for each aspect ratio, or finish a master and reframe downstream. Native generation usually looks better because composition is designed for the frame; reframing is cheaper and keeps a single visual language. Both approaches work. Drifting between them mid-project does not.
Runtime decides shot count
Divide the target runtime by your comfortable shot length, then multiply by the number of attempts each shot will need. A thirty-second piece built from three-second shots is roughly ten shots. At four attempts per shot, that is forty generations before review, sound design, and colour. Do this arithmetic in ten minutes at the start and the schedule stops being a surprise.
Placement decides pacing and sound
A product page video can be slow and detail-focused. A social feed clip needs motion inside the first second. A conference loop can be abstract, silent, and ambient. Write the placement at the top of the brief, because it changes camera behaviour, cut rhythm, and how much audio work is required.
Build a shot list that generation can execute
The shot list is the document that stops forty generations from becoming forty orphan clips. It is not paperwork; it is the spine of the edit.
One idea per row
Every row should be describable in a single sentence. If a row needs three sentences, it is three shots. Engines handle one clear action far better than a compound one, and editors appreciate the clean cut points.
A workable row reads like this: macro of condensation forming on a brushed steel surface, locked-off camera, soft directional light, three seconds. Subject, action, camera, light, duration. Nothing else.
Columns worth maintaining
- Shot number and scene
- One-sentence description
- Duration target
- Camera behaviour: locked, dolly, orbit, handheld, crane
- Subject action
- Continuity dependencies: character reference, location plate, wardrobe, prop
- Method: text-to-video or image-to-video
- Risk level
- Status: not started, in review, approved, replaced
That last column is the one teams forget. Without a status field nobody knows whether a shot is finished, and the project drifts quietly for days.
Risk scoring saves schedule time
Mark each shot low, medium, or high risk based on how hard it is to render. Hands interacting with objects, legible text on screens, complex crowd motion, and fast camera whips are high risk. Landscapes, weather, atmosphere, abstract textures, and slow reveals are low risk.
Then schedule by risk. Do the fragile shots first, while there is still time to redesign them. Leaving them for the final session guarantees a panicked scramble or a compromised edit.
Plan the cut before the shots
Sketch the sequence with rectangles and arrows. Knowing that shot four cuts on movement into shot five tells you exactly how shot four must end. Motion rarely lands on the frame you wanted by accident, so design the exit.
Test engines per shot, not per project
Different shots have different requirements. A realistic dialogue close-up, a wide landscape, and a fast action beat rarely perform best on the same engine. Treating one tool as universal means compromising half the sequence.
What to test
Build a short list based on what your project actually needs, then test each candidate against it:
- Photoreal faces at close range, held for four seconds or more
- Subject motion with a locked camera
- Camera motion with a static subject
- Physical interaction such as pouring, folding, or opening
- Legible text on signage, screens, and packaging
- Stylised or illustrative looks
- Duration limits and native resolution
- Input types: text only, single image, first and last frame, video extension
A simple test protocol
Generate the same three test shots on every candidate: a dialogue close-up, a moving-camera wide, and a hand interaction. Keep prompts identical across tools. Score each on identity stability, geometry stability, texture stability, and motion plausibility. Two hours of structured testing saves weeks of guessing, and the results are specific to your subject matter rather than to a general ranking.
A worked example
A sixty-second brand film might break down like this. The opening landscape uses whichever engine handles foliage and haze best. The product macro uses the tool with the strongest texture detail at high resolution. The founder dialogue shot uses the engine that holds facial identity longest, generated from a locked reference frame. The end card with the logo is composited in an editor rather than generated, because rendered text is still unreliable at speed.
Four shots, three engines, one compositing step. That is normal. Multi-engine work costs a little coordination and saves enormous time.
Document the mapping
Keep a short note in the shot list recording which engine produced each approved shot, plus the prompt, seed, and settings. When a stakeholder asks for a variation six weeks later, you will not be reverse-engineering your own work.
Look development: approve stills before motion
Generating motion before the imagery looks right is the most expensive habit in AI video production. Stills are fast, cheap to compare side by side, and easy to judge. Motion is slow and hard to evaluate in isolation.
Character sheets
If a person appears in more than two shots, build a character sheet: neutral expression, three-quarter angle, profile, and a full-body standing pose in consistent light. Approve one version and stop exploring. Every subsequent shot in that scene starts from one of those frames instead of from a text description.
Location plates
A location plate is a clean, empty frame of the place: a kitchen counter, a workshop bench, a street corner, a control room. It fixes architecture, palette, and light direction. When you place a character into that space, the plate becomes the first frame and the environment stops morphing between shots.
Lighting language
Write your lighting rules in plain words and reuse them word for word. Warm window light from camera left, soft shadows, visible dust in the beam is a rule. If one prompt says soft window light and the next says diffused daylight from the side, expect two different rooms. Consistent wording produces consistent pixels far more reliably than any seed number.
Get approval at the still stage
Stakeholders judge stills quickly and honestly. Once visuals are approved as frames, motion generation becomes execution rather than exploration, which removes the most common source of late-stage rework. It also gives you a reference set you can hand to another artist without a long conversation.
Prompt architecture for motion
Video prompts fail differently from image prompts. A still can be beautiful and static; a clip must stay coherent while things change inside the frame. That means specifying action and camera behaviour, not only appearance.
The five-part prompt
A reliable structure: subject identity, subject action, camera behaviour, lighting and atmosphere, style reference. For example: a ceramicist in a linen apron lifts a wet bowl off the wheel and turns it to inspect the rim, medium close-up, slow handheld drift to the right, warm window light with soft shadows and dust in the air, documentary footage with shallow depth of field.
Each part does one job. Identity says who. Action says what changes. Camera says how the frame moves. Lighting sets mood and keeps scenes matched. Style keeps the render consistent across a sequence.
One action per prompt
When a prompt stacks two simultaneous actions, engines often merge them into an unnatural hybrid motion. If a character must walk and pick something up, decide which happens in this clip and which happens in the next.
Describe speed and weight
Motion ambiguity produces mushy output. Slowly, deliberately, with weight, quick snap: each phrase changes the result measurably. If a movement reads too fast, adjust that word in the next attempt instead of rewriting the whole prompt.
Negative prompts aimed at real failures
Negatives work best when they target what you actually see. If hands deform, add negatives for extra fingers and warped limbs. If faces drift, add negatives for morphing features. If textures crawl, add negatives for flicker and grain inconsistency. Keep the list short, because a long negative list starts suppressing legitimate detail and flattens the image.
Iterate composition before detail
Generate at the lowest acceptable resolution, pick the best composition, then regenerate that composition at final quality with the same prompt and settings. This single habit cuts rendering time dramatically and improves output, because you are choosing from more options at the stage where options are cheap.
Naming and versions
Save prompt, seed, engine, and settings for every approved shot in a plain text file beside the footage, and use a convention such as sc02_sh04_v03_approved. Version numbers are not bureaucracy; they are how you compare and roll back when an experiment goes wrong.
Continuity systems that hold a sequence together
Continuity is the difference between a sequence and a pile of clips. It comes from systems, not luck.
Reuse references instead of regenerating
Generate a clean character reference once, then reuse that exact frame as the input for every shot in the scene. Regenerating the same person from text each time is the fastest route to a cast of near-identical strangers.
Wording discipline
Keep wardrobe, hair, and lighting descriptions identical across every prompt in a scene. Copy and paste rather than retyping. Small phrasing changes create visible drift, and drift compounds shot by shot until the sequence feels assembled from different productions.
Wardrobe and prop tracking
Note what changes between scenes and when. If a jacket comes off, mark the shot where it happens and make sure no later shot in the same scene shows it back on. Props are just as fragile: a mug in the left hand in one clip and the right hand in the next reads as an error even when the viewer cannot name it.
Scene-level style locks
Write a two-line style lock for each scene covering lens language and grade direction, for example 35mm equivalent, shallow depth of field, slightly cool shadows, gentle highlight rolloff. Paste it into every prompt for that scene. It costs nothing and prevents the visual jump that makes generated sequences feel stitched together.
When continuity breaks anyway
If a shot refuses to match after three or four attempts, change the shot design rather than the settings. A wider frame, a different angle, or a cutaway to a prop solves continuity problems that no prompt rewrite will fix. Redesigning is faster than fighting.
Assembly, sound design, and finishing
Generated footage without sound feels like a mood board. Audio is the fastest way to make decent imagery read as intentional, and it is usually the difference between a demo and a delivered piece.
Cut for rhythm first
Bring approved selects into the editor, ignore colour and gaps, and cut for pacing. Most generated clips are slightly the wrong length, so trim to the beat rather than stretching to fill. If a shot cannot survive a trim, it was never the right shot.
Three sound layers
- Ambience: room tone, wind, traffic, machine hum. One looping bed under the whole edit glues shots together.
- Effects: footsteps, fabric, metal, water, clicks. Place them on the action frames, because they anchor generated motion to physical reality.
- Voice: recorded or synthesised narration. Narration can be synthetic; on-camera dialogue usually should not be, since lip sync and emotional nuance remain the weakest link.
Upscale only what survives the edit
Sharpening artefacts multiply when applied to footage you later cut. For flickering clips, mild temporal denoise helps. For geometry wobble, no filter helps and the shot should be regenerated or replaced. Sequence-level decisions come first; per-clip polish comes last.
Match colour and grain across the timeline
Apply one grade across the edit, then match individual shots inside it. A light, consistent film grain over everything hides small differences in render character better than shot-by-shot correction. Aim for the sequence to feel unified rather than for each clip to be flawless in isolation.
The quality gate before anything reaches the timeline
Run every clip through the same checklist. Consistency matters more than speed here, because the checklist is what protects the edit from a weak shot that slipped in unnoticed.
- Does the shot satisfy the shot list, or is it merely attractive?
- Does the face hold identity for the full duration?
- Do hands and limbs stay anatomically plausible through motion?
- Do background elements stay stable, or warp at the frame edges?
- Is the camera move continuous, without an unexplained jump?
- Do colour and lighting match the neighbouring shots?
- Is the clip long enough to cut without stretching?
- Does it conflict with brand, legal, or talent rules?
Any clip that fails the same check twice should be redesigned, not regenerated a third time. Repeating a failed approach is the most common way a working day disappears.
Hold a short review at the end of every session and update the status column immediately. Projects drift when the document and the folder disagree, and the disagreement is usually only a few hours old when it starts.
Hybrid production and realistic effort planning
Generated video is not always the right answer, and pretending otherwise wastes time. Practical shooting wins in predictable situations:
- Hands interacting precisely with objects
- Legible text on packaging, screens, or signage
- Real people speaking to camera
- Product detail where colour accuracy matters
- Anything requiring a specific real location for authenticity or legal reasons
Generated footage wins for scale, atmosphere, impossible camera moves, period settings, weather, and abstract transitions. A hybrid edit is normal practice, and audiences judge coherence rather than provenance. Mixing a practical product shot with generated atmosphere is usually faster, cheaper, and better than forcing one method to cover everything.
Effort and role planning
Estimate attempts per shot, multiply by shot count, and add contingency of roughly half again. Track generation volume the way you would track any variable cost, and review it weekly rather than at the end.
On a small team, three responsibilities matter even when one person holds all of them. A director owns the shot list and final approval. A reference owner maintains character sheets, plates, and style language. An editor assembles selects and manages sound. When those roles blur, projects generate endlessly and nobody decides when a shot is done.
FAQ: practical answers for first projects
How long should each generated clip be?
Shorter is safer. Three to eight seconds gives editing flexibility and reduces drift. Longer shots are possible, but they work best when built from several generations joined at natural cut points.
Should I start with text-to-video or image-to-video?
Use text-to-video to explore mood, concept, and framing options. Use image-to-video whenever continuity matters: character close-ups, product shots, or any frame that must match something already approved.
How do I keep a character consistent across shots?
Lock a reference frame, reuse it for every shot in that scene, keep wardrobe and lighting wording identical, and use any subject-reuse feature your tool offers. Regenerating the person from text each time guarantees inconsistency.
What is the biggest quality gain for the least effort?
Sound design. Well-placed ambience and effects make average generated footage read as professional far more reliably than another round of regeneration.
Do I need a powerful local machine?
Not necessarily. Most production work happens in browser-based tools. Local hardware matters mainly for upscaling, compositing, and colour work, where a mid-range workstation is usually enough.
How many variations should I generate per shot?
Three to six during exploration, one to three once the composition is approved. More variations only help when you compare them against a clear checklist; otherwise you accumulate indecision instead of options.
What if a shot never looks right?
Change the design, not the settings. A different angle, a wider frame, a cutaway, or a practical pickup solves most persistent failures faster than another prompt rewrite.
Can generated footage be used commercially?
That depends on the terms of each tool you use and your jurisdiction. Read the licence for every engine in your stack, keep records of prompts and source references, and confirm the rules for real people or brand elements that appear on screen.
Where should a beginner start?
Start with a single ten-second scene, three shots maximum, image-to-video only, with one character reference. Finish it completely, including sound, before scaling up. The lessons from a finished ten seconds are worth more than an unfinished hour.
