Describe the video you want and let a language model break it into a numbered shot list, with one prompt per shot written for the video model that will render it. Then run each prompt in the same workspace, watch the clips, and tighten the wording before you cut them together.
A video model is only as good as the sentence you hand it, and most first prompts are one vague line that leaves the camera, the light and the ending to chance. The fix is to treat the prompt like a shot list: ask a language model to split your idea into shots, write each shot the way your video model reads best, then render the clips one by one and cut them together. Picasso IA keeps both halves in one workspace.
Writing a video prompt is a different skill from writing an image prompt. A still image needs a subject and a look. A clip needs a subject, an action, a camera move, a starting point, an ending point and often a sound, all inside a few seconds. Most people forget at least two of those, and the video model fills the gap with its own guess. A language model keeps the checklist and asks what the camera is doing.
Picasso IA lists 502 models in total, including 75 large language models and 121 text-to-video models, and they all run in the same toolkit. That matters because the handoff is the whole point: you draft the shot list with a chat model, paste a shot into a video model, and carry a wrong clip back to the chat model as evidence. New accounts get free credits to try it, and every generation, text or video, spends credits at a rate that depends on the model and the settings, so the pricing page is the place for numbers.
The other reason is where mistakes get caught. A rendered clip is an expensive place to discover that you never said what the camera does; a numbered shot list on screen shows that shot three has two actions before a single frame is rendered.
A useful shot prompt has the same parts every time, in roughly the same order: the subject, what the subject does, the camera, the setting and the light, the visual style, and the sound if the model makes sound. A fixed order makes a shot list easy to scan and easy to fix.
Here is one shot written that way: a wide shot, slow push-in, of a red fox standing at the edge of a misty pine forest at dawn, soft side light, the fox lifts its head and turns to look left, quiet birdsong in the distance. It has one subject, one action, one camera move and one light source. The story comes from the next shot, which is its own prompt.
When you ask the language model for the list, say what you want back. Tell it the video model you will use and the clip length it allows, ask for numbered shots with one action each, and ask it to describe the main subject identically in every shot. Ask for visible things, not feelings: a prompt that says the scene feels lonely gives the model nothing to draw, while a single lantern on an empty pier does.
Tip: copy the subject description word for word into every shot. Each clip is generated on its own, with no memory of the one before it, so a fox described as small and russet in shot one and as a fox in shot two can come back as two different animals.
The same shot text does not land the same way on every model, because the models accept different inputs and have different limits. Here is how four of the text-to-video models in the catalog differ, from the details listed on their model pages.
| Model | What it accepts | How to shape the prompt | Watch out for |
|---|---|---|---|
| Veo 3.1 | Clips of 4, 6 or 8 seconds, start and end frames, up to 3 reference images, audio on by default | One shot per prompt: scene, camera and ambient sound in a few sentences | Reference images only work at 16:9 and 8 seconds |
| Kling v3 Video | Up to 15 seconds, up to 6 shots in one run, prompt up to 2500 characters, audio off by default | A short overall prompt plus one line per shot, each with its own duration | Shot durations must add up exactly to the total length |
| Seedance 2.0 | Up to 9 reference images, 3 reference videos, 3 audio clips, audio on by default | Tag references as [Image1] in the prompt and put spoken lines in double quotes | Reference images cannot be combined with start and end frames |
| Wan 2.7 T2V | Up to 15 seconds, 1080p, optional audio file, negative prompt, prompt expansion on by default | Plain, specific sentences; turn expansion off if you want your exact wording kept | Expansion rewrites your prompt and adds waiting time |
| Any model with a start image | A still photo or image as the first frame | Describe only what moves and how the camera behaves | The clip can drift away from the still as it plays |
Two habits follow from the table. Tell the language model which row you are writing for, because a prompt for Kling v3 Video should come back as shots with durations while one for Veo 3.1 should be a single tight paragraph. And test one shot on two models before committing to a whole list. The Picasso IA vs Google Veo and Picasso IA vs Kling AI pages go deeper on how those two families differ.
This is the route from a loose idea to a folder of clips in order, with a chat model and a video model, both inside the toolkit.
The most common mistake is putting too many actions in one shot. A clip of a few seconds can hold one clear movement, maybe two. When a prompt asks a character to walk in, pick up a cup, smile and wave, the model compresses all of it into a blur of half-finished gestures. Split it into shots; a language model will do that without complaint if you tell it the rule.
The second is mixing camera moves that fight each other, such as a fast zoom in with a slow pull back. Pick one move per shot and name it plainly: slow push-in, static wide, pan left, low tracking shot. The third is writing exclusions into the main prompt, because saying no people or no text there often plants the very thing you wanted gone. Where the model has a negative prompt field, as Veo 3.1, Kling v3 Video and Wan 2.7 T2V do, put plain nouns there instead. For an exact look, make a still of the shot first, as the AI storyboards page describes, and use it as the start frame.
A language model can write a good prompt, but it cannot make a video model obey it. These are the places where this workflow stops being enough on its own:
A better prompt raises the odds of a usable clip but never makes one certain. When a shot shows people, use only your own face and voice or fictional characters, never a real, identifiable person. If the finished video is going to a platform, check that platform's current requirements for length, format and disclosure before you export.
It is a language model used as a writing partner for video prompts. You give it an idea and the limits of your video model, and it returns a numbered list of shots, each with a subject, an action, a camera move and a light. On Picasso IA the chat model and the video model sit in the same toolkit, so a prompt goes straight from the reply into the generator.
There is no single best, so try two. Picasso IA lists 75 large language models, including Claude Sonnet 5, Gemini 3 Flash and GPT-5.4. Give each the same idea, render the first shot from each reply on the same video model, and keep the one that came closer to what you pictured.
Usually the content stays the same and the format changes. Veo 3.1 reads one shot per prompt and accepts start and end frames. Kling v3 Video can take up to 6 shots with their own durations in one run. Seedance 2.0 expects reference tags like [Image1] and quoted dialogue. Ask the chat model to rewrite the same shot list for whichever model you picked.
Long enough to cover the subject, the action, the camera, the setting and the light, and no longer. For most shots that is a few sentences. Kling v3 Video caps a prompt at 2500 characters, which you will rarely approach. Wan 2.7 T2V can expand a short prompt for you; turn that off when you want your exact wording kept.
On some models. Kling v3 Video lets you define up to 6 shots, each with its own prompt and duration, inside one clip of up to 15 seconds, and the durations have to add up to the total. Most other models make one shot per run, so you render each shot separately and join the clips afterward in an editor.
Do two things together. Describe the character with identical words in every shot, and give the model reference images where it accepts them: Veo 3.1 takes up to 3 and Seedance 2.0 up to 9. Expect some drift anyway, especially in clothing, and check each clip against the first.
Yes, when the model has one. Veo 3.1, Kling v3 Video and Wan 2.7 T2V each have a negative prompt field, and it works best with plain nouns: text, watermark, extra people, blur. Keep exclusions out of the main prompt, where naming the unwanted thing can make it more likely to appear.
Yes, with differences between models. Seedance 2.0 generates dialogue, effects and music in one pass and expects spoken lines inside double quotes. Veo 3.1 has audio on by default and responds to a described soundscape. Kling v3 Video has audio off by default, so turn it on if you want it. Voices should be fictional or your own.
New accounts get free credits, which is enough to test a shot list on a model or two. After that every generation costs credits, and the amount depends on the model and the settings you choose, such as length and resolution. Current plans are on the pricing page, which is the only place to trust for numbers.
Yes. Veo 3.1, Kling v3 Video and Seedance 2.0 accept a start image, and the prompt then describes only the motion and the camera, not the scene already in the picture. Use photos you took yourself, or fictional art, never a real person who has not agreed to it. The turn photos into videos page covers the animation side.
Your idea is already a shot list waiting to be written. Open the Picasso IA toolkit, draft the first three shots with a chat model, and render the first clip to see what the wording needs.
Choose your language
