Black Friday

50% off a whole year of any plan

Get the deal
eroq

blog/prompts·Nov 21, 2026·8 min·by the eroq team

AI video prompt mistakes — 12 that waste takes, and the fix for each

The AI video prompt mistakes behind most wasted takes — two beats in five seconds, keyword soup, three camera moves — each with the broken prompt and the fix.


Most failed AI video takes come from a handful of prompt mistakes rather than from the engine: too many beats for the length, a list of keywords instead of a shot, and instructions the engine was never built to follow. Below are the twelve we see most, each with the prompt that fails, why it fails, and the rewrite that renders.

They hold across engines — Seedance, Kling, Veo, Hailuo, Runway, Wan, Grok Imagine and Motion One all read a prompt the same broad way. Where an engine changes the fix, the note says so, and the per-engine prompt guides at the end go further.

1. Two beats in five seconds

Fails:

She opens the front door, drops her bag, walks to the window and looks out at the storm.

Why: that is four actions. A 5-second clip has room for one, so the engine either rushes all four into a smear or merges them into a single half-gesture.

Renders:

She pushes the front door open against the wind, rain blowing in around her. Slow push-in, cold blue light from the hall.

Put the next beat in the next clip — Extend in Video picks up from the last frame. The help center has the same symptom from the other side: my clip looks like two shots crammed together.

2. Keyword soup

Fails:

cinematic, epic, 8k, masterpiece, ultra detailed, dramatic, beautiful lighting, award-winning, woman, forest

Why: quality words describe nothing the engine can draw, and the two nouns that matter come last, with no action and no camera.

Renders:

A woman in a moss-green cloak walks between giant redwoods at dawn, shafts of light cutting through the mist. Slow tracking shot beside her, soft contre-jour.

Subject, action, place, camera, light — in one sentence or two. Prompt length and what engines ignore explains where the extra words go.

3. Three camera moves in one clip

Fails:

Drone shot over the city, then a crane down to the street, then an orbit around the dancer.

Why: each move needs time to read. Three in one clip become a camera that wobbles between intentions.

Renders:

A dancer spins on a rooftop above the city at dusk. Slow orbit around her at shoulder height, city lights coming on behind.

One move per clip. The camera moves guide helps pick the one that serves the shot.

4. Describing the photo again on image-to-video

Fails (with a start frame set):

A woman with long red hair in a blue dress stands in a sunlit kitchen with white cabinets, and she smiles.

Why: the frame already carries her hair, the dress and the kitchen. Describing them again asks the engine to redraw what it was given, and faces drift.

Renders:

She looks up from the counter and smiles toward the camera; steam rises from the kettle behind her.

With a start frame, write only what moves.

5. Negating inside the prompt

Fails:

A quiet beach at sunset, no people, no boats, no text.

Why: "no people" still puts people in the prompt, and many engines weigh the noun more than the "no".

Renders:

An empty beach at sunset, smooth wet sand reflecting the orange sky.

Describe what is in frame. On engines with an Avoid field in Settings (Veo and Kling have a native one), put what must stay out there instead. Negative prompts explained covers how it works.

6. Dialogue too long for the clip

Fails (6 seconds):

He says "I've been thinking about what you said last night, and honestly I think you were right about everything, so I'm sorry."

Why: that line takes longer to say than the clip lasts. It comes back rushed, cut off, or with lips that stop moving before the words do.

Renders:

He looks down, then back at her: "You were right. I'm sorry."

About one short line every four or five seconds. And check the engine speaks at all: Runway Gen-4.5, Grok Imagine and Hailuo 3 Max render no soundtrack.

7. Lines with no speaker

Fails:

Two friends in a car at night. "Where are we going?" "You'll see."

Why: the engine can't tell who says what, so both lines may come from the same mouth — or from nobody.

Renders:

Two friends in a car at night, dashboard glow on their faces. @Ines asks "Where are we going?" @Sam keeps his eyes on the road: "You'll see."

Attribute every line, ideally to an @mentioned character. On the Seedance engines and Hailuo 3, a character's voice sample then carries their read.

8. Tiny faces at 480p

Fails (480p):

A wide shot of a crowded stadium, the camera finding one fan cheering in the upper stands.

Why: at 480p a face in a wide crowd is a few pixels. It melts, and everyone else melts with it.

Renders:

A medium close-up of a fan in the upper stands, cheering with both fists raised, the blurred crowd behind her.

Keep 480p for close framing and blocking; render wide crowds at 720p or more.

9. Asking for on-screen text

Fails:

A neon sign spelling OPEN ALL NIGHT above the diner door, and the title "Episode 3" appears at the bottom.

Why: video engines draw letters as shapes, and shapes shift from frame to frame. Titles flicker and misspell themselves.

Renders:

A red neon sign glowing above a diner door at night, rain falling through its light.

Add titles in the Editor, where text stays sharp. If a sign must be legible, make a still on an image engine that writes text, then animate it from a start frame.

10. Writing a cut into one clip

Fails:

Close-up of the letter in her hands, cut to her face as she reads it.

Why: one clip is one continuous shot. "Cut to" comes back as a jump, a morph, or a camera that lurches between the two framings.

Renders — two clips:

Close-up of a handwritten letter trembling slightly in her hands, window light across the paper.

Close-up of her face as she reads, eyes filling, the window light on her cheek.

Line them up in Cinema or side by side in the Editor. On the Seedance engines, Cinema renders the shots of one scene as a single take with real cuts.

11. Real people and famous faces

Fails:

A famous actor and a pop star dancing in a nightclub.

Why: real, identifiable people without their consent are refused by eroq's acceptable-use policy on every engine, uncensored ones included. The request stops before it renders and costs nothing — but it never becomes a clip.

Renders:

@Vera and @Leon dancing in a crowded nightclub under strobing magenta light, handheld, sweat and confetti.

Create your own characters from photos you have the right to use, and cast them by name.

12. Changing everything between takes

Fails: a take is close, so the next prompt changes the subject's clothes, the lens, the light and the camera move at once.

Why: you can't tell which change helped and which hurt, and a new random draw changes things you didn't touch.

Renders: copy the take's Seed from Settings, keep it, and change one phrase. On engines with a seed — the Seedance family, Veo, Wan, Runway — that holds everything else still. On engines without one (Kling, Hailuo, Grok Imagine, Motion One), set Takes to render up to four versions of the same prompt and compare them side by side.

A checklist before you press Generate

  1. One subject, one action, one camera move, one light.
  2. A start frame set? Then describe the motion only.
  3. Dialogue: short, in quotes, after its speaker — and an engine with sound.
  4. No text, no cuts, no real people.
  5. What must stay out goes in Avoid, not in the prompt.
  6. Draft cheap; keep the seed of a take you like.

If a draft is rough, Enhance rewrites it into a director-grade prompt (rewriting a draft needs a plan; Surprise me, which invents a prompt from an empty box, is free). Reading what it writes is a quick way to learn the register.

Prompt guides by engine

When a clip still bends or flickers after a clean prompt, the cause is usually physics rather than wording: why AI video looks warped and flicker and morphing in AI video cover those.

FAQ

How long should an AI video prompt be?

Between 30 and 80 words for most clips: subject, action, place, camera, light, and sound where the engine renders it. Under five seconds, one sentence is enough; past ten seconds, add a beat in order rather than more adjectives.

Why does the engine ignore part of my prompt?

Usually because the prompt asks for more than the clip can hold — several actions, several camera moves — or because the ignored part came last in a long list. Cut to one of each, put the essentials first, and add detail back one phrase at a time with the seed held.

Does Enhance fix these mistakes?

Most of them. Enhance rewrites a draft into a director-grade prompt with one action and one move. It can't make a silent engine speak or a real person allowed, so the checklist above still applies.

Is it free to practise?

eroq is free to start: 40 free credits when you sign up, no card. Seedance 2.0 Mini, at 75 credits for 5 seconds at 480p, is the cheapest way to test a prompt; packs start at $10.

Try the rewrite of your last failed prompt in the Video studio — /studio/video.

Tagsprompt-guidepromptstext-to-video

Shoot it in Video

Everything above happens in the studio: same engines, same settings, rendered on your click.

40 free credits · no card · private mode on every account

Prompts from this article

She opens the front door, drops her bag, walks to the window and looks out at the storm.

Privacy

Private and unrestricted

  • →No filter on eroq's own engines: one written policy instead.
  • →They keep nothing, and nothing trains on your work.
  • →Private mode keeps no file at all.