Most AI videos fail for the same boring reasons. The subject morphs mid-clip. The camera does something nobody asked for. The product changes color between seconds two and four. The output is technically "a video" and practically unusable.
After looking at tens of thousands of real AI video prompts — the ones that produced clips people actually shipped, and the ones that produced garbage people deleted — a pattern emerges. Great prompts aren't longer or more poetic. They're more structured. They tell the model what changes, how the camera behaves, what must stay locked, and what they refuse to accept.
This is the craft companion to our data report on what 40,000 AI video prompts reveal about what people make. That post covers what creators generate. This one covers how the good ones write it. Five patterns, each with a weak version, a strong version, and why the difference matters.
Pattern 1: Lead With Subject, Action, and Change Over Time
Video is motion. The single biggest difference between prompts that produce living footage and prompts that produce a slow zoom on a photograph is whether you described something happening.
Weak prompts describe a scene. Strong prompts describe a scene that changes.
Weak: A coffee cup on a wooden table in a cafe.
Strong: A steaming coffee cup on a wooden cafe table; steam curls upward and drifts left as morning light slowly brightens across the surface over 5 seconds.
The weak version gives the model a still image and forces it to invent motion — usually a lazy push-in or some ambient jitter. The strong version names the subject (coffee cup), the action (steam curls and drifts), and the change over time (light brightening across the clip). The model now has a beginning and end state to interpolate between, which is exactly what a video model is built to do.
The fix is mechanical. For every prompt, ask: what is the one thing that is different at the end of this clip versus the start? If you can't answer, you're going to get a moving postcard. Bake that change into the sentence. Even a small one — a head turn, a door opening, fog rolling in — gives the model a job to do across the timeline.
Pattern 2: Direct the Camera Like a Cinematographer
If you don't specify the camera, the model picks one for you — and it picks badly, defaulting to a generic dolly-in or a drifting handheld wobble that screams "AI." The best prompts treat the camera as a deliberate creative choice, not an afterthought.
You need three things: shot size (wide, medium, close-up), lens or framing feel (35mm, wide-angle, shallow depth of field), and one motion (slow push-in, orbit, static lock-off). One motion. Not three.
Weak: A car driving down a coastal road, cinematic.
Strong: Wide tracking shot of a vintage convertible on a coastal highway, shot on a 35mm lens with shallow depth of field, camera tracks alongside the car at matching speed, golden hour.
"Cinematic" is a wish, not an instruction. The strong version tells the model the framing (wide tracking), the optical character (35mm, shallow depth of field), and a single coherent move (track alongside at matching speed). That coherence is what reads as professional. Conflicting camera instructions — "orbit while zooming and panning" — are where models fall apart and produce that swimmy, unstable look.
If you're new to thinking in camera terms, our guide on how to write AI video prompts breaks down the vocabulary. The shortcut: imagine you're handing a one-line instruction to a camera operator who will do exactly what you say and nothing more. Be that specific.
Pattern 3: Lock Your Continuity Tokens
This is the pattern that separates hobbyists from people producing usable footage. AI video models drift. Across a few seconds, a face subtly re-renders into a different person, a red logo shifts to orange, a product gains a button it didn't have. Continuity tokens are the specific, repeatable phrases you use to nail those elements down.
A continuity token is a short, distinctive description you commit to and reuse verbatim — for the subject's identity, the product, the color palette, and any branding.
Weak: A woman in a red jacket walks through a city, then we see her closer up.
Strong: A woman with shoulder-length curly black hair and a bright crimson leather jacket walks through a neon-lit city; same crimson jacket and same hairstyle held consistent throughout the clip.
"A woman in a red jacket" is an invitation for the model to reinvent her. "Shoulder-length curly black hair and a bright crimson leather jacket," repeated and explicitly flagged as consistent, gives the model an anchor to hold. When you generate multiple clips for one project, copy those exact tokens into every prompt — never paraphrase them. Paraphrasing is how the character in shot three stops looking like the character in shot one.
For brand work this is non-negotiable. Lock the exact hex-equivalent color name, the logo placement, and the product's defining feature in every single prompt. If your platform supports an image reference or text-to-video with a starting frame, use it — but back it up with locked text tokens, because the description is what carries identity through the motion, not just into the first frame.
Pattern 4: Match the Shot to Platform and Duration
A prompt that's great for a 12-second YouTube hero is wrong for a 4-second TikTok hook, and the difference isn't just aspect ratio. The best prompts are designed backward from where the video will live.
Three decisions get made before you write a word of description: aspect ratio (9:16 vertical for feeds, 16:9 for YouTube and landing pages), duration (and therefore how much can actually happen), and pacing (one calm beat for a short loop, a clear arc for a longer clip).
Weak: An energetic montage of a fitness product with lots of quick cuts and text, for social media.
Strong: 9:16 vertical, single continuous 5-second shot: a runner laces up bright orange sneakers and pushes off frame-left into a sprint, fast-paced, punchy, designed as a TikTok hook with the action landing in the first 2 seconds.
Asking for "lots of quick cuts" inside a single short generation is asking for a mess — most models produce one continuous shot per generation, so the request fights the tool. The strong version respects the format: vertical, one shot, an action engineered to hit in the first two seconds where the platform demands it. You'll often get a better result by generating several clean single-shot clips to this spec and cutting them together than by trying to cram an edit into one prompt.
Duration drives how much change you can ask for, too. In four seconds, one clear action lands. In twelve, you can stage a small arc. Asking for a three-act story in four seconds just smears everything together.
Pattern 5: Constrain With Negatives and a Clear Output Spec
The final pattern is the one almost nobody uses, which is exactly why it's an edge. Telling the model what you don't want is often more powerful than piling on more of what you do. Pair that with an explicit output spec and you stop leaving the unglamorous decisions to chance.
Two moves: negatives (the artifacts and clichés you refuse — warped hands, text gibberish, extra limbs, flickering, the unwanted slow zoom) and an output spec (frame rate feel, lighting, mood, and aspect ratio stated plainly at the end).
Weak: A chef plating a dish in a restaurant kitchen.
Strong: A chef precisely plating a dish in a warm restaurant kitchen; medium shot, soft key light from the left, calm and deliberate pacing, 16:9. Avoid: distorted hands, extra fingers, floating utensils, on-screen text, fast camera movement.
The negative list does real work. Hands are where video models embarrass themselves, so naming "distorted hands, extra fingers" tells the model to spend effort there. "Avoid on-screen text" kills the gibberish lettering models love to hallucinate. And closing with the output spec — shot size, lighting direction, pacing, aspect ratio — means you're not hoping the model guesses your intent; you've stated it.
Keep your negative list tight and relevant. Ten generic negatives dilute the signal. Three or four that target this prompt's likely failure points sharpen it. Different models have different weak spots, so it pays to know which one you're using — our AI model strengths map breaks down where each model excels and where it tends to break.
How to Combine All Five Into One Prompt
These patterns aren't a menu — the best prompts stack all five. Here's the order they naturally fall into:
- Subject + action + change ("a chef plates a dish; steam rises as she sets the final garnish")
- Camera ("medium shot, 50mm, slow push-in")
- Continuity tokens ("same chef in a white double-breasted jacket throughout")
- Platform + duration spec ("16:9, 8 seconds, calm pacing")
- Negatives + output ("warm key light from the left. Avoid: distorted hands, on-screen text")
Read top to bottom, that's a single coherent instruction a model can execute confidently. Each clause answers a question the model would otherwise answer for itself — and "for itself" is where bad AI video comes from.
You don't have to start from a blank page every time, either. A library of copyable prompt templates gives you proven skeletons for common shot types; you swap in your subject and tokens and you're already running all five patterns without thinking about it.
Your Next Step
Pick one prompt you've written that produced a disappointing clip. Run it through the five patterns: Does it name a change over time? Does it direct one clear camera move? Are your continuity tokens locked and repeated? Is it specced to a real platform and duration? Does it tell the model what to avoid?
Fix the two weakest answers and regenerate. That single edit pass is usually the difference between a clip you delete and a clip you ship.
When you're ready to put the patterns to work, open text-to-video in the app and write your first prompt the structured way — subject, camera, tokens, spec, negatives. And if you want the data behind what's actually working at scale, read the companion analysis of what 40,000 AI video prompts reveal. Craft plus evidence is how you stop guessing and start directing.