AI Video Generation Workflow: A Practical Veo-Class Guide

Why generated footage finally earns its place in the timeline

For years, AI video was a demo format. Clips ran four seconds, hands dissolved into texture, and any camera move longer than a slow drift turned into a smear. The output impressed in isolation and fell apart the moment you dropped it next to real footage.

Modern systems behave differently. Larger, better-labelled video datasets and diffusion-transformer architectures now produce takes that hold together long enough to cut: consistent lighting across a shot, plausible weight and momentum, readable faces at medium distance, and enough resolution to survive a 1080p delivery. Text-to-video, image-to-video, and video-to-video are no longer separate novelty toys. They are three modes of one pipeline, and a competent editor switches between them inside a single project.

The people who gain the most from this are not large studios. They are the wide middle of content production: a two-person agency building a product explainer, a solo creator shipping short-form daily, a training team that needs a warehouse walkthrough, a game studio prototyping a trailer. Work that once required a shoot day, a location, and a model release can now be drafted before lunch and refined by evening.

The bottleneck has moved. The question is no longer whether a model can render a shot, because it usually can. The question is whether you can produce twenty shots that feel like they belong to the same film, on a deadline, without spending the week regenerating. That is a directing problem dressed up as a prompting problem, and it responds to process rather than to a longer sentence.

Everything below assumes that framing: the model is a camera with opinions, not a scriptwriter. Your job is to control what enters the frame, what stays consistent, and what gets cut.

Text-to-video, image-to-video, video-to-video: choosing the right mode

Every shot in an AI project should be routed to one of three generation modes before you type anything. Choosing deliberately is the difference between a smooth build and an afternoon of guessing.

Text-to-video

You describe a scene and the model returns a short clip, usually four to eight seconds at a selectable aspect ratio. This mode shines when atmosphere matters more than precision: establishing landscapes, weather, abstract transitions, crowd wides, texture inserts, light moving across a surface. It struggles with precise hand interaction, rendered lettering, choreography that requires exact timing, and anything where the audience must recognise a specific person across multiple shots.

Image-to-video

You supply a still frame and describe the motion. This is the single most reliable technique for continuity, because composition, wardrobe, colour, and framing are already locked before the model touches them. Generate or photograph one strong reference image, then animate it. Storyboards effectively become animatics with almost no rework. Use this mode for characters, product details with visible labels, props, and any shot where a viewer would notice if a jacket changed colour.

Video-to-video and reference conditioning

Some pipelines accept an existing clip plus a style or subject reference. That is useful for restyling live-action footage, extending a shot past its original length, or maintaining a character across takes. Treat any reference file as a first-class production asset: name it clearly, version it, and reuse the same file every time that subject appears. Never overwrite a reference image that already works.

A routing table you can keep open

Shot type	Best mode	Why
Establishing landscape	Text-to-video	Atmosphere matters more than accuracy
Character close-up	Image-to-video from a locked still	Preserves identity and wardrobe
Product detail	Image-to-video with a slow camera move	Full control over composition and label
Crowd or action wide	Text-to-video, short duration	Hides small inconsistencies
Restyle of existing footage	Video-to-video	Keeps motion and timing intact
Dialogue line	Image-to-video plus a separate audio performance	Cleaner control over sync
Abstract transition	Text-to-video	Failure looks intentional

The rule of thumb: the more specific the subject, the more you should start from an image. The more atmospheric the shot, the more a text prompt can carry on its own.

The five-slot prompt framework that survives rendering

A prompt is not a set of magic words. It is a shot description written for someone who has never seen your storyboard. The most reliable prompts contain five pieces of information in roughly this order: subject, action, camera, light, and look. Write them as one flowing sentence, but think of them as five separate decisions you are making on purpose.

Slot one: subject

State who or what, plus two or three concrete visual details. Age range, wardrobe, material, colour, condition. A woman in her thirties in an olive raincoat, a brushed steel kettle with a chipped enamel handle, a delivery van with faded blue paint. Abstract praise such as stunning or cinematic carries almost no visual information. Colour, texture, and material do.

Slot two: action

One continuous verb phrase. One. Walking slowly toward the camera. Steam rising from a cup. A cyclist coasting through a puddle. Two or three actions in a single prompt force the model to interpolate between them, and the middle of the clip usually turns to soup. If you need a person to pick something up and then turn, that is two shots.

Slot three: camera

Shot size, angle, and movement. Slow push in from a medium shot. Handheld tracking behind the subject. Locked-off wide. Slow arc to the right. Choose one movement, not three. Camera language is where beginners overreach most; a static shot that looks good beats a complicated move that wobbles.

Slot four: light and time of day

Golden hour, overcast diffusion, practical neon at night, hard noon sun, single window source with soft falloff. Light direction should match between adjacent shots, so write it down and reuse the same phrase across the whole scene.

Slot five: look

Lens feel, grain, palette, texture. Shallow depth of field, 35mm grain, muted teal and amber grade, documentary realism, high-contrast monochrome. This slot controls mood more than any other, and it is the one people forget when a clip technically works but feels wrong.

A complete example

A woman in her thirties in an olive raincoat, walking slowly toward the camera through a wet market at dusk, slow push in from a medium shot, warm practical lights reflecting off puddles, shallow depth of field, muted teal and amber grade.

Every clause in that sentence is something a camera operator could act on. Nothing is abstract.

Negative lists and continuity anchors

State what you do not want: no text overlays, no extra limbs, no lens flare, no watermark. Keep the list short and stable, because a long negative list starts fighting the positive prompt.

Then freeze your continuity anchors and repeat them verbatim in every prompt for a sequence: the same wardrobe description, the same location phrase, the same time of day, the same lens language. Small wording changes produce large visual changes. If a phrase must not move, copy and paste it rather than retyping it from memory.

Iterating without drifting

Change one variable per generation. If you alter camera and light at the same time, you learn nothing about which change caused the improvement. Keep a simple log with the prompt version, the seed if the tool exposes one, and a one-line verdict: usable, close, or discard. After twenty clips that log is worth more than any tutorial you will read, because it describes your project rather than someone else's.

Planning the piece: beats, shot list, and a continuity bible

Write beats, not shots

A beat is a change in information or emotion, not a camera setup. The problem appears. The first attempt fails. The turn happens. The resolution lands. A thirty to sixty second piece usually needs four to eight beats. Write them as present-tense sentences before you think about visuals at all. When the beats are stable, decide how many shots each one needs, and resist the urge to give every beat five.

Build a shot list as a spreadsheet

Six columns are enough: shot ID, beat, description, generation mode, target duration, and status. Add an audio column if the piece has dialogue. Keep the list open while you generate and mark shots in progress, usable, or parked. In the edit you will live in this file, and ten minutes of setup saves an hour of scrolling through an unlabelled folder of clips.

Keep a one-page continuity bible

This is the document that separates a coherent piece from a pile of pretty clips. It lists wardrobe descriptions, location phrases, time of day, colour notes, and the filenames of reference stills. Copy exact strings from the bible into prompts instead of paraphrasing. Small wording changes cause large visual changes, so the bible exists to freeze anything that must not move.

The production loop, stage by stage

Stage one: beat sheet and script

Finish the words first. For narration-driven work, write the voice track and read it aloud with a timer. For visual-only pieces, write the beat sheet and one line per shot. Deciding the story after generating footage is how projects die.

Stage two: storyboard stills

Generate a still for every shot before generating a single second of video. Stills are fast, cheap to iterate, and trivially easy to rearrange. When the still sequence reads clearly without any motion, you know the video edit will read clearly too. This one step removes most of the frustration people associate with AI video, because you discover structural problems while they cost seconds rather than hours.

Stage three: clip generation

Ask for four to eight seconds per clip, not fifteen. Short takes render faster, fail cheaply, and cut better. Generate two or three variants of each shot using the same prompt with a single varied element, then choose in the timeline rather than in the browser tab. Batching matters here: write all prompts first, then generate all clips, then review them together. Context switching is the hidden cost of AI production.

Stage four: assembly

Bring clips into an editor and cut on action. Resist using a full clip simply because it exists; a beautiful shot that stalls pacing is still a bad shot. Overlap audio and picture slightly at cut points so the transitions breathe, and let the sound of the next scene arrive a few frames before the picture does.

Stage five: sound design

Room tone under every scene, a music bed ducked under narration, foley on any visible contact. Replace synthetic ambience whenever it fights the edit; nothing signals amateur work faster than audio that cuts hard between shots.

Stage six: grade and delivery

Apply one grade across all shots so they sit in the same world. Matching black levels and colour temperature does more for perceived coherence than any single regeneration. Then normalise loudness, check the piece on a phone speaker, and export separate masters for each aspect ratio you publish in.

Sound design is the invisible quality multiplier

Viewers forgive a slightly soft image far more readily than they forgive bad sound. If generated footage is silent or carries thin synthetic ambience, the piece reads as a demo no matter how good the frames look.

Start with room tone. Every location has a floor of background noise, and laying the same tone under all shots from one scene glues them together. When the scene changes, change the tone, even subtly.

Add foley for anything that touches. Footsteps, a mug landing on a table, fabric moving. You do not need perfect recordings; you need a plausible transient. Libraries of small household sounds cover most of it.

Duck music under dialogue rather than lowering the whole bed. A two to four decibel dip under speech keeps energy without masking words. If a model generates dialogue with lip sync, use it for short, front-facing lines only, and watch the alignment at half speed. That is where artefacts hide: a slight drift between vowel shapes and sound, a jaw that keeps moving after the line ends.

Finally, listen to the whole piece on a single phone speaker before publishing. Most short-form video is watched exactly that way.

Quality control checklist before you publish

Run this pass on the finished timeline, not on individual clips.

• Continuity: does light direction match between adjacent shots?
• Motion: does any clip begin or end with a stall, a snap, or a sudden speed change?
• Faces and hands: any drift when watched at half speed?
• Text in frame: any garbled lettering? Blur it, replace it, or cut the shot.
• Backgrounds: any warped architecture, floating objects, or reflections that behave like liquid?
• Audio: room tone present throughout, and no hard cuts in ambience between shots?
• Loudness: consistent across the piece and within normal platform range?
• Captions: burned in or uploaded, line breaks checked, no overlap with faces?
• Aspect ratios: correct master exported for each destination?
• Rights: every reference image, voice, and music track licensed, generated, or cleared?

Two minutes here prevents a re-upload, and re-uploads cost reach.

Mistakes that quietly burn your render budget

Overloading a single prompt. Three subjects, two actions, and a camera move in one sentence produce mush. Split the idea into separate shots and let the edit create the relationship between them.

Choosing the aspect ratio late. Generating widescreen material for a vertical edit means cropping away the composition you carefully designed. Set the frame first and keep it consistent through the whole project.

Regenerating instead of diagnosing. If a shot keeps failing, the problem is usually structural: too many subjects, impossible physics, or a camera move the model cannot sustain. Change the shot, not the seed.

Chasing one perfect take. Ten ordinary clips that cut well beat one flawless clip that does not fit anywhere. Judge clips in context.

Ignoring reference assets. Save reference stills, prompt text, and seeds in one folder per project. Future you will need them, and so will anyone who picks up the work.

Skipping sound. Silent generated footage is a demo. Sound design is what makes an audience believe they are watching a film.

Forgetting the half-speed check. Watch every clip once at half speed specifically for hands, eyes, teeth, and background text. Most problems are visible there and invisible at full speed, which means your audience might see them on a large screen.

Treating the first pass as final. Nobody writes a good edit on the first timeline. Expect to cut twenty to thirty percent of what you generated, and plan for it emotionally before you start.

Time, iteration, and what to promise a client

Estimate generously. A thirty second finished piece commonly needs sixty to ninety seconds of generated footage, and each shot may need two to four attempts before one holds up. That is normal, not failure, and it should be inside your schedule from the beginning.

Batch your work by stage rather than by shot. Write all prompts, then generate all stills, then generate all clips. The savings come from staying in one mode of thinking for an extended stretch instead of jumping between writing, generating, and reviewing.

For recurring characters or a brand look, build a small library: three reference stills, one locked prompt template, one grade preset, one loudness target. Reuse is what turns a one-off experiment into a production line that can be quoted confidently.

Keep a consistent project folder: brief, beat sheet, shot list, stills, clips, audio, exports. Two minutes of filing saves twenty minutes of searching next month, and it makes handing a project to a collaborator almost painless.

When scoping client work, be honest about the split. Generated visuals are excellent for atmosphere, inserts, concept sequences, and short-form social. For a hero performance with a named actor, treat the model as a previz tool and shoot the real thing. Clients accept that boundary when you explain it early; they do not accept it when you explain it after delivery.

Frequently asked questions

Do I need a powerful computer to work this way?

Usually not. Most capable video models run in the cloud through a browser, so a mid-range laptop and a stable connection are enough. Local hardware only matters if you plan to run open-weight models yourself or do heavy editing with many timeline effects.

How long should a single generated clip be?

Four to eight seconds. Longer clips drift, physics degrades with speed, and you rarely need more before a cut anyway. Generate several short takes and cut between them rather than asking for one long one.

How do I keep the same character across shots?

Generate or photograph one strong reference image, then use image-to-video for every appearance. Lock the wardrobe, hair, and lighting wording in your prompt and never paraphrase those lines. Save the reference file and reuse the identical version every time.

Can I use generated footage commercially?

It depends on the terms of the specific model you use and on your local rules. Read the licence, keep records of what you generated and when, and be cautious with prompts that reference real people, brands, or protected characters. If a client project depends on it, get the usage terms confirmed in writing before production starts.

Why does my footage look fine in the preview but wrong in the edit?

Previews are usually watched in isolation at full speed on a small window. In an edit your clips sit next to each other, where mismatched light direction, colour temperature, and motion speed become obvious. That is why the grade and the QC pass matter as much as generation.

What is the fastest way to get better at this?

Finish one thirty second piece every week: scripted, generated, edited, scored, published, and reviewed. Volume of completed work teaches more than any settings menu, because it forces you to make the decisions that actually determine quality.

Should I write prompts first or storyboard first?

Storyboard first. Prompts are shot descriptions, so writing them before you know the shot means rewriting them immediately. Once the still sequence reads clearly, prompt writing becomes a mechanical translation of decisions you have already made.