Why the script still decides the outcome
Every few months a new generation engine arrives, the sample videos get sharper, and it becomes tempting to conclude that writing no longer matters. In practice the opposite is true. The cheaper generation becomes, the more the script becomes the only real differentiator. Two creators using the same tool on the same afternoon will produce wildly different results, and almost all of that gap is decided before either of them presses generate.
Think of text-to-video as a translation problem. You are translating intent, tone, and sequence into images, motion, and sound. Translation works only when the source is unambiguous. A line like show how our app saves time gives the engine nothing to hold on to. A line like shot three, close-up of a phone screen, calendar at 14:02, one tap on approve, confirmation slides up gives it almost everything. Specificity is not a technical nicety here; it is the input format the system is built to consume.
There is a second reason the script dominates. Video is judged in the first few seconds. Retention curves fall fast, and the cause is rarely a slightly soft frame. It is a slow opening, a missing hook, a promise that never gets paid off, a jump in logic between two shots. Those are writing problems. A flawless render of a boring opening is still a boring video, and no amount of upscaling fixes a structure that never had a point.
The practical takeaway: treat the script as a production document rather than a creative artifact. It should contain runtime, shot beats, on-screen text, voiceover lines, and the emotional target of each beat. If a shot is not described well enough for a stranger to imagine it, the model will not imagine it either.
What text-to-video tools do well — and where they break
A useful mental model is bandwidth. A generation model can hold a handful of constraints at once. Give it one subject, one action, one camera idea, and one lighting mood, and it will usually deliver something usable. Give it four characters, a costume change, a legible logo, and a spoken line, and it will start dropping constraints, usually the ones you care about most.
Where generated footage wins
- Atmosphere and texture: moving clouds, rain on glass, city skylines at dusk, slow pushes through empty interiors, abstract gradients for lower thirds and transitions.
- Scale that would otherwise need permits: aerial-style establishing shots, crowds, industrial interiors, landscapes.
- Concept visualization: pitch teasers, mood films, style tests, and storyboards that move enough to communicate a feeling.
- Repetition and reformatting: the same insert shot rendered for vertical, square, and widescreen deliverables without a second shoot.
- Localization backgrounds: swapping culturally specific street scenes or interiors so a localized cut feels native.
- Placeholder shots: rough visuals that hold a timeline together while you decide whether the real footage is worth shooting.
Where generated footage breaks
- Fingers doing fine work: typing, tying, cooking, playing an instrument, handling small objects.
- Legible text inside the frame: signage, packaging, screens, subtitles burned into the shot.
- Continuity across cuts: the same character, wardrobe, room, and light across ten shots.
- Long unbroken action: anything that needs more than roughly five seconds of coherent physics.
- Brand-critical detail: an exact product silhouette, a specific logo lockup, a real spokesperson.
- Precise sync: a door closing exactly on the beat, a gesture landing on a specific syllable.
Decision criteria
| Shot requirement | Better approach |
|---|---|
| Short, atmospheric, easy to describe | Generate |
| Contains legible words or numbers | Generate the background, add text in the edit |
| Same character in five shots | Lock a reference image and generate per shot, or cast a real person |
| Hand-held product interaction | Shoot for real, generate supporting b-roll |
| Needs a specific beat sync | Shoot or generate separately, cut to the music in the timeline |
| Needs a recognizable person or logo | Real footage, licensed assets, or motion graphics |
The rule of thumb that survives contact with deadlines: generate what is expensive to shoot and cheap to describe, and shoot what is cheap to shoot and expensive to describe.
The script format that makes automated direction work
A script written for humans is not the same as a script written for a pipeline. Novels and essays move through ideas; video moves through moments. Reformatting your writing into moments is the single highest-leverage hour you will spend on any project.
Write in shot beats, not paragraphs
Replace paragraphs with numbered beats, each one a single visual idea. A beat contains a subject, an action, a framing, and a duration. Everything else is decoration.
Beat 4 — Medium close-up, laptop lid opens on a cluttered desk, morning light from the left, 0:08 to 0:11. On-screen text: Before the first meeting.
That format gives a human editor and a generation model the same starting point. It also exposes weaknesses immediately: if two consecutive beats say the same thing visually, one of them is filler.
Keep narration and on-screen action separate
Mixing voiceover lines into visual descriptions is the most common cause of muddled output. Split the document into two columns or two lists:
- Narration: what the viewer hears, written in spoken language, one idea per sentence.
- Visuals: what the viewer sees, written as shot beats with framing and duration.
When they are separated, you can rewrite a sentence for rhythm without breaking a shot list, and you can replace a shot without rewriting the voice track. It also makes localization straightforward: the visuals rarely change, the narration does.
Do the timing math before you generate
Generation is the slow, expensive part of the process. Do not discover at the assembly stage that your script runs two minutes long. Narration at a comfortable pace lands between about 140 and 160 words per minute. A 60-second video therefore holds roughly 150 spoken words, which is far fewer than most first drafts contain.
Use a simple planning table before you generate anything:
| Runtime | Narration budget | Shot beats | Average shot length |
|---|---|---|---|
| 15 seconds | 30 to 40 words | 3 to 5 | 3 to 4 seconds |
| 30 seconds | 65 to 80 words | 6 to 9 | 3 to 4 seconds |
| 60 seconds | 140 to 155 words | 10 to 16 | 4 to 6 seconds |
| 3 minutes | 420 to 470 words | 30 to 45 | 4 to 6 seconds |
Two details matter here. First, generated clips often work best in the three-to-five second range, which happens to match the natural cut rhythm of short-form video. Second, every shot you plan beyond the budget costs generation time and review attention, so trimming beats early is cheaper than trimming clips later.
Choosing an approach: avatar-led, generative b-roll, or hybrid
There is no universal best configuration. There is a best configuration for your message, your tolerance for unreality, and the amount of control you need over details.
Avatar or presenter-led. A synthetic or recorded presenter speaks to camera, with generated or stock visuals cut in as illustration. This works when the message is explanatory, the brand is personality-driven, and viewers expect a face. It is also the most sensitive to the uncanny valley, so it pays to keep cuts fast and intercut heavily with supporting visuals.
Pure generative b-roll with voiceover. No face, no studio, just sequenced generated visuals under a voice track and music. This is the fastest path from script to publishable video and the easiest to localize. It suits tutorials, listicles, explainers, and product education where the product itself is not the visual subject.
Hybrid. Real footage of the product, the team, or the location, combined with generated atmosphere, transitions, backgrounds, and inserts. This is usually the strongest option for brand work: authenticity where trust matters, generation where cost would otherwise spike.
Screen-led. Recorded software walkthroughs with generated or motion-graphic framing, intros, and recaps. Essential for software, and notably hard for generative engines because interface text must be pixel-accurate.
Three questions decide it for you. Does the audience need to trust a face? Does the message depend on a specific, unalterable visual detail? Is the deadline measured in hours or weeks? If trust matters, use real footage for that part. If exact detail matters, keep it out of the generator. If the deadline is hours, go voiceover plus generated b-roll and accept that atmosphere will carry more weight than specifics.
The production workflow, stage by stage
Stage 1: Lock the script and the runtime
Freeze length before anything else. Read the narration out loud with a timer, and cut until it fits the budget from the table above. Reading aloud catches phrases that look fine on a page and stumble in the mouth. If you cannot record a clean read, the model will struggle with it too.
Stage 2: Build the shot list
Convert every narration sentence into one or two visual beats. Mark each beat as generate, shoot, or graphic. Assign a duration to each. At this point the video exists on paper, and it is far easier to fix a weak middle on paper than in a timeline.
Stage 3: Create a style bible and test frames
Write a short style block that you will paste into every prompt: lens, lighting, palette, grain, era, mood. Keep it to one sentence and do not improvise it halfway through. Then generate three to five test frames on your most important beats. If the test frames do not look like the video you imagined, adjust the style block now, not after thirty clips.
Stage 4: Generate motion in priority order
Generate your hero shots first, the ones that carry the message. Then fill in supporting shots. Keep prompts to one subject, one action, one camera move. Reuse the same style block verbatim across every prompt so the footage feels like it came from one production rather than one afternoon.
Stage 5: Produce the audio bed
Record or generate the voiceover before you lock visuals. Voice timing is far less flexible than image timing: you can stretch a shot, you cannot stretch a syllable without it sounding wrong. Once the voice is locked, cut the music to its rhythm rather than the other way around.
Stage 6: Assemble, caption, and version
Lay the voice track down first, then place shots against it. Generate captions from the final audio, not from the original script, so the text matches what is actually spoken. Export each aspect ratio separately rather than cropping a finished master, and check the safe zones for captions on vertical.
Stage 7: Review and version properly
Watch the cut once with sound, once without, once at double speed, and once on a phone. Each pass catches a different class of problem. Keep the previous export until the new one is approved; version discipline saves more time than any prompt trick.
Consistency: the hardest problem in AI video
Ask anyone who has produced more than a handful of AI-assisted videos what actually consumes their time, and the answer is rarely generation. It is consistency: keeping a character, room, wardrobe, and light looking like they belong to the same world across a dozen separate clips.
Four habits fix most of it.
Write a character sheet and reuse it word for word. Not a paragraph of personality, but a physical description: age range, hair, clothing, distinguishing features, posture. Paste it into every prompt that includes the character. Paraphrasing it is the same as changing the character.
Lock a reference frame. Generate one frame you are happy with and use it as the visual anchor for every shot in that scene. Reference-based workflows exist precisely because text alone drifts.
Limit the variables you change between shots. If the wardrobe changes, keep the location and light. If the location changes, keep the wardrobe and the lens. Changing everything at once makes drift look like chaos.
Cut around weak continuity. If two shots refuse to match, insert a close-up, a graphic, or a hard cut on a beat. Editors have hidden continuity problems for a century; you can use the same trick.
Also accept a practical limit: if a sequence needs six shots of the same person doing different things, that sequence is a candidate for real footage. Fighting a model for hours can cost more than twenty minutes with a phone camera.
Audio, pacing, and the rhythm of the edit
Viewers forgive soft visuals far more readily than bad sound. Treat the audio as the spine of the piece and build outward.
Voice direction matters more than voice quality. Write for the mouth: short sentences, concrete nouns, no nested clauses. Where the voice lands, place the cut; where it pauses, place the transition. Narration that sits between roughly 140 and 160 words per minute feels natural for explainers, and faster than that starts to feel like a list being read aloud.
Music should be chosen after the voice, not before, because its job is to fill the gaps the voice leaves. Aim for a bed that you notice when it stops and not while it plays. If you can hear specific instruments competing with consonants, the music is too loud or too busy.
Sound design is the cheapest production value available. A soft whoosh on a transition, a click on a button press, a room tone under a scene, a low thump on a title card: these take minutes to add and change how finished the video feels. Generated visuals often arrive silent, which makes an otherwise good cut feel like a slideshow.
Pacing follows one rule: the cut should arrive slightly before the viewer wants it. If you find yourself waiting for a shot to end, that shot is too long by about a third. Short-form edits typically live in the two-to-five second range, with a longer hold only when the frame contains something genuinely worth reading or studying.
Quality control checklist and common mistakes
Pre-publish checklist
Run this list on every export:
- Hook: does something happen in the first two seconds?
- Clarity: can a stranger say what the video is about after one viewing?
- Audio: are voice levels consistent, and is the music ducking under speech?
- Captions: do they match the spoken words, and are they inside safe zones on vertical?
- Legibility: is any on-screen text readable on a phone at arm's length?
- Continuity: do consecutive shots feel like the same production?
- Brand: logo, colours, and fonts appear where they should, and nowhere they should not.
- Payoff: does the ending deliver what the opening promised?
- Deliverables: correct aspect ratio, duration, and file naming for each platform.
Mistakes that cost the most time
- Generating before the script is final. Every script change invalidates footage you already made.
- Overloading prompts. Five constraints produce one that sticks; one constraint produces a usable clip.
- Ignoring the reference frame. Consistency problems are almost always solved before generation, not after.
- Writing narration that reads well but speaks badly. Always read aloud.
- Adding captions from the script instead of the audio. They will drift.
- Rendering one master and cropping it. Vertical, square, and widescreen need separate framing decisions.
- Chasing perfection on a shot nobody notices. Spend the saved hours on the hook.
- Publishing without watching on a phone with the sound off. Most of your audience will do exactly that.
Building a repeatable pipeline for a team
Once one video works, the temptation is to treat it as a one-off. The value is in turning it into a pipeline that produces the tenth video as easily as the first.
Standardize the document. One template with sections for runtime, narration, shot beats, style block, and deliverables. New projects start from the template, not from a blank page.
Standardize naming. Project, version, aspect ratio, date. Naming sounds trivial until three people are looking for the right file at 11 p.m.
Standardize the style block. Teams drift because everyone rewrites the style sentence slightly differently. Keep the approved versions in one place and reference them.
Define review gates. Script approval, test frames, first assembly, final export. Fewer gates means more rework; more gates means slower delivery. Three gates is usually the sweet spot: script, look, and final.
Separate generation from editing. Whoever generates clips should not also be cutting the timeline if you can avoid it. Generation is a focused, prompt-level task; editing is a rhythm-level task, and switching between them wastes both.
Plan for localization from the start. Keep narration in a separate file, avoid text baked into visuals, and leave a little extra room in each shot for languages that run longer than English.
FAQ
How long should a text-to-video project take?
A 60-second explainer with 12 to 16 generated shots is typically an afternoon of generation and editing once the script is final, and a day or two if you are building a new style. The script stage is where most of the calendar time hides, because every revision there is free and every revision later is expensive.
Do I still need an editor if I use AI tools?
Yes. Generation replaces camera work and some motion design. It does not replace pacing, sound balance, captions, or the judgement about which shot is too long. Editing is the part of the process that most determines whether the video feels professional.
Which is better: one long prompt or many short shots?
Many short shots. A single long prompt asks the model to solve a sequence in one pass, and sequence is exactly where generation tends to break. Short clips cut together give you control, and cutting is something you already know how to do.
How do I keep a character consistent across shots?
Write a fixed character description and reuse it verbatim, generate a reference frame and anchor every shot to it, and change only one variable at a time. Where consistency still fails, cut around it or use real footage for that sequence.
What should I shoot for real instead of generating?
Anything with a real person's face that needs to build trust, anything with legible text or an exact product shape, and anything involving hands doing precise work. Generate the atmosphere and the inserts around that footage.
How do I write a hook that actually works?
Say the most interesting true thing first. Avoid introductions, throat-clearing, and logos at the top of a short video. If the first two seconds could be removed without losing meaning, they should be removed.
Should I generate my own voiceover?
Use synthetic speech when you need speed, multiple languages, or a consistent narrator across many videos. Record a human voice when warmth, humour, or authority is the point. Either way, write for the mouth, not the page.
What is the most common reason an AI video underperforms?
A vague script rendered beautifully. The footage looks fine and the video still fails, because nothing in it was specific enough to remember or to act on.
How do I handle captions for multiple platforms?
Generate captions from the final audio, style them once, and keep them inside the safe zone for each aspect ratio. Then review them by hand; auto-captions handle names, brands, and technical terms poorly.
How much should I plan before generating anything?
Enough that a stranger could draw your video from the shot list. That usually means runtime, narration, beats, durations, and a style block. It takes an hour and saves several.
