How to write a video prompt that works

Do it yourself

1. [Shot size and angle] of [one subject, one line of description], [one action], in [place, time of day].[One camera move, with its speed]. [Lens and focus].[Light: where it comes from, how big, how hard]. [Style or film look, if any].Sound: [the place's ambience]. [Effects, marked "off screen" if their source must stay out of shot].[Speaker] says: [one short line, no quotation marks] Image to video: delete everything the still already shows. Keep the camera move, the action and the sound.Omni Flash: end with "One continuous shot, no scene cuts."Open models: follow the maker's own grammar (LTX: one paragraph under 200 words; H3: its three fields).

   Each bracket is a decision. Leave one out and the model makes it for you, usually with its most average answer.

2. 1. Text to video, establishing shot. Veo 3.1 Lite, 16:9, 720p, 4 seconds ($0.20)Wide establishing shot of a small bakery on a cobbled street at dawn, light rain. The camera holds still. Deep focus. Cool blue light from the sky, warm light spilling from the shop window. Documentary realism. Sound: rain on the awning and a tram passing in the distance. 2. Image to video, product push-in. Veo 3.1 Lite, 16:9, 720p, 4 seconds ($0.20). The approved still is frame one.The camera pushes in slowly towards the tin. A thin wisp of steam rises behind it. Nothing else moves. Sound: quiet room tone and, off screen, the clink of a spoon on a saucer. 3. Talking head with one line. Veo 3.1 Fast, 9:16, 720p, 6 seconds ($0.60)Medium close-up of a woman in her sixties at a market stall, eye level, soft overcast daylight. She holds up a jar of honey and says: Picked this morning, from the top field. Sound: market chatter far behind her.

   Example 2 uses the off-screen wording that kept the hand out of our Lite clip. Example 3 follows the colon rule and sits on Fast, where our dialogue went through.

3. For the target video, at 0.00 seconds into the target video, <Picture 1> (from [Shot 1]) is fully referenced. integrated_multimodal_description: [Shot 1] Live-action, cinematic, the matte black tea tin shown in <Picture 1> stands on the grey slate shelf, keeping its label, its shape and the soft window light from the left. The camera pushes in with small amplitude at slow speed towards the tin as a thin wisp of steam rises behind it. overall_soundscape: Quiet room tone. The clink of a spoon on a saucer comes from off screen. non_diegetic_music: N/A

   Same shot, different grammar. H3's guide wants the alignment line first, then three named fields, with every camera move written as type, amplitude and speed.

4. # check_prompt.py: read a video prompt back to you the way the makers' guides would.# It sends nothing and costs nothing. Run: python3 check_prompt.py "your prompt here"import re, sys prompt = sys.argv[1]words = prompt.split()notes = [] # Google's Agent Platform guide: quotation marks can make Veo draw the words on screen.if re.search(r"[\"“”]", prompt):    notes.append("Quotation marks. For speech, write: She says: the line. Keep quotes only for words that must appear.") # Runway and Luma: negation can backfire. Describe what is there instead.for w in ("no ", "not ", "don't ", "without ", "never "):    if w in prompt.lower():        notes.append(f"Negation ('{w.strip()}'). Describe the result you want instead.") # One camera move per short clip.moves = re.findall(r"\b(push(?:es)? in|pull(?:s)? out|pans?|tilts?|orbits?|arcs?|dolly|dollies|zooms?|tracks?|cranes?)\b", prompt.lower())if len(set(moves)) > 1:    notes.append(f"More than one camera move ({', '.join(sorted(set(moves)))}). Pick one.") # Our own finding: Veo stages whatever makes a sound. A spoon set down brought a hand.if re.search(r"\b(set down|clink|pour|stir|knock|tap|click)\w*", prompt.lower()) and "off screen" not in prompt.lower():    notes.append("A sound with a maker. Veo may show the hand that makes it. Add 'off screen' or move it to the edit.") # Google and Runway: one moment per short clip.if prompt.lower().count(" then ") >= 2:    notes.append("A chain of events. Short clips do one moment; split it into shots.") # LTX asks for under 200 words; long prompts collect contradictions.if len(words) > 200:    notes.append(f"{len(words)} words. LTX asks for under 200, and Veo caps prompts at 1,024 tokens.") print(f"{len(words)} words.")print("\n".join("- " + n for n in notes) if notes else "Nothing to flag. Now read it aloud as a shot list.")

   Show all 36 lines

   It needs no key and sends nothing. On a deliberately bad prompt it flagged five problems; on the three examples above it flagged none.

The words, in plain language

Shot size

How much of the scene the frame holds, such as a wide shot, a medium shot or a close-up.

Camera move

What the camera itself does during the clip, such as a slow push-in, a pan or holding still.

Motion prompt

The words you send with a still. They describe what moves, because the picture already shows the rest.

Positive phrasing

Describing what is there instead of what is absent. Runway and Luma both warn that negation can backfire.

Prompt rewriting

A second model expanding your words before the video model reads them. Wan calls it prompt extension; Luma calls it Enhance.

Timestamp prompting

Writing each beat against a time, such as [00:02] the lid lifts. Runway, Wan and MiniMax H3 each document a form of it.

Sound cue

Words for a sound. Veo tends to put whatever makes the sound into the picture.

Why it matters

Elsewhere