Course resource
AI Video Prompt Book
Getting usable clips out of a video generator, and knowing when not to bother.
The prompt structure
[SHOT TYPE] of [SUBJECT] [ACTION], [SETTING], [LIGHTING],
[CAMERA MOVEMENT], [STYLE], [DURATION AND PACING]
Weak: a person walking in a city
Working:
Medium tracking shot of a woman in a grey raincoat walking briskly
along a wet pavement, neon-lit street at night, reflections in puddles,
camera follows from her left at walking pace, cinematic, shallow depth
of field, single continuous 5-second shot, no cuts
Camera movement is the highest-leverage term
Naming the movement is what makes output look filmed rather than generated.
| Term | Effect |
|---|---|
| Static / locked off | Calm, observational |
| Slow push in | Building attention |
| Pull back / reveal | Context, surprise |
| Tracking shot | Energy, following |
| Pan left/right | Surveying a space |
| Handheld | Immediacy, documentary |
| Drone / aerial | Scale |
| Orbit | Showcasing an object |
Say "single continuous shot, no cuts" on almost everything. Generators will invent cuts, and an unintended cut mid-clip is the most common reason a take is unusable.
What still fails
Be realistic about what you will get. Current generators are strong on short, simple, physically plausible shots and weak on everything else.
- Hands doing complex things. Gripping, manipulating objects, typing.
- Multiple people interacting. Two subjects is much harder than one.
- Continuity across clips. The same character in two separate generations will not match. Use reference images or character features where supported.
- Text in frame. Add it afterwards in an editor.
- Long shots. Quality degrades with length. Generate short and cut together.
- Precise choreography. You cannot direct a specific sequence of actions reliably.
The working method: generate many short clips, discard most, edit the survivors together. Do not try to generate the finished video.
Prompts by purpose
B-roll
[SHOT TYPE] of [SUBJECT], [SETTING], [LIGHTING], slow [MOVEMENT],
no people's faces, 4 seconds, single continuous shot, cinematic,
neutral colour grade for editing
Product
Slow orbit around [PRODUCT] on [SURFACE], soft studio lighting from
above and left, seamless [COLOUR] background, shallow depth of field,
macro detail on [FEATURE], 5 seconds, no text, no hands
Establishing shot
Wide aerial of [LOCATION] at [TIME OF DAY], [WEATHER], slow forward
drone movement, natural colour, 6 seconds, single continuous shot
Talking-head avatar
Script: [YOUR SCRIPT]
Framing: medium close-up, eyes to camera
Background: [PLAIN / OFFICE / BLURRED]
Delivery: [CONVERSATIONAL / AUTHORITATIVE], moderate pace
For avatars, the script matters far more than the visual prompt. Write for the ear: short sentences, no subordinate clauses, punctuation where you want breath.
The workflow that produces usable video
- Storyboard first, in text. List every shot you need before generating anything.
- Generate 5–10 takes per shot. Assume a 1-in-8 hit rate.
- Judge on the first second. If the first second is wrong, the rest will be.
- Watch finalists at full speed twice. Artifacts hide in motion — hands, background people, cloth, reflections.
- Cut in a real editor. The generator makes clips; the edit makes the video.
- Add text, logos and captions afterwards. Never generate them.
- Grade the whole thing together so clips from different generations match.
Cost control
Video generation is the most expensive thing in this course by an order of magnitude.
- Storyboard before generating. Most waste is generating shots the edit never needed.
- Generate at low resolution to find the take, then regenerate the winner at full quality where the tool supports it.
- Set a budget per project before you start, and track it.
- Ask whether stock footage would be faster and cheaper. Often it is.
Before you publish
- Watched at full speed, twice, looking for artifacts
- No unintended cuts mid-clip
- Hands and background people check out
- No real person's likeness without written permission
- Licence tier permits this use
- Disclosed where the audience would reasonably care — especially anything that looks like footage of a real event
- Generation records kept
When not to use it
- When you need a specific real place, product or person shown accurately.
- When continuity across many shots matters.
- When stock footage would do — it is cheaper, faster and more reliable.
- When it must be evidence of anything.