JSON PROMPT
GENERATOR.

A shot description written as key-value pairs instead of a paragraph: subject, action, environment, camera, lighting, audio, timed beats and an avoid list, each in its own field. You paste it into Veo, Flow, Sora, Runway or Kling exactly as you would a prose prompt. The structure is for you as much as the model: it makes a missing decision visible before you spend a generation on it.

None of them publish a JSON prompt schema. The model reads the braces and keys as text, the same as any other prompt. JSON still earns its place for two reasons: every key is a reminder to decide something you would otherwise leave to chance, and across a batch of clips, keeping the same keys and changing one value is the cleanest way to see what that value actually does.

Subject and action first, because they carry the meaning. Then environment, camera (angle, movement, lens), lighting and color palette, and audio if the model generates sound, as Veo 3 does. Add timed beats only when the clip has more than one movement, and an avoid list for things the model keeps inventing. If you are animating a still, describe the starting frame, and fix the seed image before you touch the prompt: in our own A/B test, changing only the seed image, with the prompt held identical, flipped the result.

For text models, and for extraction, production workflows and anything you need to be repeatable, usually yes. Key-value structure removes the ambiguity that prose carries, and these models were trained on a great deal of structured data — schemas, configs, APIs — so the shape itself is familiar to them. The gain is mostly consistency across runs rather than a smarter answer on any single run.

Exploratory and creative work. Structure narrows the space of possible outputs, which is exactly what you want for data extraction and exactly what you do not want when you are asking the model to surprise you. If you are brainstorming, write prose and let it wander.

No, and the distinction matters. This tool formats the prompt you send. Structured outputs, offered through the model APIs, enforce a schema at generation time so the response is guaranteed to parse. They solve different halves of the problem and work well together: use a structured prompt to say what you want, and structured outputs to guarantee the shape of what comes back.

Yes. Switch to Text (LLM) mode; the output is plain text you can paste into any of them, or into a system message via an API. In our own use Claude handles deeply nested structures and long constraint lists particularly well, but nothing here is model-specific.

JSON when something downstream will parse the response. XML when the prompt itself is long and needs clearly separated sections, which models follow well. Markdown when a human will read the output. Plain text when the task is simple enough that structure adds nothing — over-formatting a trivial request just spends tokens.

Free, no signup, and nothing is stored or sent anywhere. The prompt is assembled in your browser from the fields you fill in.