The official prompting guides from Google, OpenAI, Runway, Kling, Luma, LTX and MiniMax agree on a few things, such as concrete description and motion-only prompts for image-to-video, and openly contradict each other on negative phrasing, dialogue quotation marks and prompt shape. This article puts the guides side by side in one table, quotes what each actually says, and rewrites one shot for each model's rules.
Method. Vendor guides and API references read November 18, 2026; every prompting rule is linked to the vendor page it comes from, and the rewritten prompts are our untested illustrations. Vinc makes an AI film studio.
Which guides we read, and which models they cover
The official video-model prompting guides disagree more than most "ultimate prompt guide" posts admit. We read the official pages from seven vendors on 18 November 2026:
Its example uses inline "No hands visible. No text overlays."
API lists supported camera-motion strings
Start and end keyframes; no written advice found
Not covered
LTX (LTX-2.5)
Shot, scene, action, characters, camera, audio, as one present-tense paragraph
About 4–8 sentences for a single shot
No API field; the ComfyUI template applies a default
11 camera phrases, plus an API camera_motion setting with 8 values
Describe what should happen; one continuous take
In quotation marks; state language and accent
MiniMax (Hailuo 2.3)
[command] brackets for camera moves
Up to 2,000 characters
No field in the request schema
15 bracket commands, up to 3 per bracket
Not in the page we read
Not in the page we read
Where the guides agree
Concrete beats abstract. Runway says to replace words like "beautiful" with visible facts. Kling says to "Avoid abstract terms", and LTX asks writers to "Express emotion through physical cues, not abstract labels."
In image-to-video, describe the motion, not the picture. Google's heading is "Prompt for motion only", and it warns "Redundant prompts confuse the model". Runway says image-to-video prompts "focus almost exclusively on motion", and LTX says the prompt "describes what should happen". Kling and Sora treat the image as the first frame and the text as what comes next. Runway and Google both ask for a high-quality source image.
One generation, one focused moment. Sora: "Each shot should have one clear camera move and one clear subject action." Google warns that chaining events in a short clip "often leads to muddled or incomplete videos". Runway warns against conflicting instructions such as "dramatic shadows with soft, even lighting".
Standard film vocabulary works. Every guide uses ordinary terms like pan, tilt, dolly and close-up. Camera Language for AI Video Prompts maps each term to the guides that list it. The guides differ in how many movement terms they spell out:
Camera-movement terms listed in each vendor guide
Runway Gen-4.518
MiniMax Hailuo 2.315
Google Veo / Omni12
LTX-2.511
Counts from the guides read 18 November 2026: Runway Camera Terms page (moves), MiniMax Hailuo API reference (bracket commands), Google video generation prompt guide (movements), LTX-2.5 prompting guide (camera phrases, not all of them moves).
Timing can be written in. Runway supports timestamp prompting ("[00:03 through 00:04] black eyes squint"), Omni accepts a timecode such as "[0-5s]", and Kling uses lines like "X–Y s".
Where they differ, and why it matters
Negative phrasing. This is the sharpest split. Runway says "Models respond better to descriptions of what you want to happen rather than what you don't want to happen." Veo moves exclusions into a separate field, while Omni Flash ("No dialogue") and Luma ("No text overlays.") write them into the prompt.
Quotation marks around dialogue. Google says to put speech after a colon and "avoid using quotation marks", so the words aren't rendered as on-screen text: A woman says: My name is Clara. LTX says the opposite: "Place spoken dialogue in quotation marks." Kling also uses quotation marks.
Shape of the prompt. LTX wants "a single flowing paragraph" and warns: "Don't paste a prompt written for another video model (e.g. Kling or Seedance) into LTX unchanged". Kling's own examples use exactly that: bracketed headers, @-tags and timed shot lists.
Cuts. Omni Flash "will try to create a video with a few different shots" unless you ask for "a single continuous shot". Runway suggests adding "Continuous, seamless shot" when an unwanted cut appears.
Camera control outside the prompt. MiniMax uses bracket commands such as [Pan left], and LTX's API has a camera_motion setting. Kling 4.0's text-to-video API has no camera parameter, so camera direction goes in the text.
One shot, rewritten for each guide
The base idea is ours: an old fisherman mends a net on a wooden pier at dawn. A gull lands beside him. He looks at it and says "Not today." The rewrites are our illustrations, applying only the rules cited above. None is quoted from a vendor guide or tested against the model.
Google Veo 3.1 (audio in separate sentences, speech after a colon, unwanted items in the negative field):
Medium shot, eye level, of an elderly fisherman in a grey wool sweater mending a green net on a weathered wooden pier at dawn. Soft golden backlight, light mist over the water. A seagull lands on the post beside him; he looks up at it. The fisherman says: Not today.
Audio: gentle waves against the pilings, a gull's call, the creak of the pier.
Negative prompt: music, text, subtitles
Medium shot of an elderly fisherman mending a green net on a weathered wooden pier at dawn. A seagull lands on the post beside him and he slowly raises his head to look at it. Soft golden backlight, light mist on calm water, muted natural colours, the camera holds steady at eye level.
Kling 4.0 (core sentence, then camera, timing, audio; quoted dialogue; one short exclusion):
An elderly fisherman mends a green net on a wooden pier at dawn, calm and absorbed.
Camera: medium shot, eye level, slow push-in.
0–3 s: he works the net. 3–5 s: a seagull lands on the post beside him and he looks up.
Audio: waves and a single gull call. In a dry, low voice, he says: "Not today."
No subtitles.
LTX-2.5 (one present-tense paragraph, emotion as physical cues, quoted dialogue):
A medium shot frames an elderly fisherman in a grey wool sweater on a weathered wooden pier at dawn, mist drifting over flat water behind him. His thick fingers pull a green net taut and knot it with practised speed. A seagull lands on the post beside him with a flap of wings. He pauses, raises his eyes without moving his head, and the corner of his mouth tightens. He says quietly, in a gravelly English voice, "Not today." The camera pushes in slowly toward his face as soft waves lap against the pilings.
MiniMax Hailuo 2.3 (bracket command for the camera move):
An elderly fisherman mends a green net on a wooden pier at dawn, soft mist on the water. A seagull lands on the post beside him and he looks up at it. [Push in]
The content barely changes; the syntax decides whether "Not today" is spoken, burned in as a caption, or dropped.
How we compiled this
Every rule here comes from the vendor pages below, read on 18 November 2026; we used no third-party guides and did not test the rewrites. Guides change quickly (the Sora 2 API shut down on 24 September and the Veo 3.1 preview in the Gemini API on 22 October), so check the date on any page before relying on it.
AI Music for Film: Suno, Udio and Other Generators on Licences, Ownership, Lawsuits and Cue Sheets
Whether an AI-generated track can go in your film depends on the plan you made it on, whether the terms exclude film and TV, and whether anyone can own it at all. This guide compares the published terms of Suno, Udio, ElevenLabs Music, Stable Audio and AIVA, sets out the label lawsuits as reported, and explains what US copyright law and North American PRO rules mean for your cue sheet.