Text-to-Video Prompt Structure: Text and JSON

Most weak video prompts fail in the same way: everything is mixed into one sentence, and the model cannot tell which words describe the scene and which describe the camera. A structured prompt fixes that by giving each instruction one job and a fixed place.

This guide covers the seven parts, the order that works, and how to write the same prompt in two formats: a continuous text block and JSON.

The seven parts

Dialogue, background audio and a list of things to avoid are optional extras that sit after these seven.

Writing the scene

The scene is the only part written as full sentences. Describe it as a sequence of events in the order they happen: the character enters, does one thing, then another. Name objects that must appear and state what must stay hidden. Models follow “he holds the letter so only the blank back of the paper faces the camera” far better than “mysterious letter”.

Keep it to one continuous action that fits the length of the clip. If you need a second location or a time jump, that is a second shot and a second prompt.

Text version: one continuous block

For models with a single prompt box, write the parts in this order: style, scene, then camera, lighting and lens as short comma-separated labels. Put dialogue and audio in a second paragraph.

[Visuals] Cinematic 4K, Photorealistic. A weathered lighthouse keeper climbs the last steps of a spiral staircase, carrying a brass lantern. He stops at the window, wipes the fogged glass with his sleeve and watches a storm roll in over the sea. Medium Shot, Rule of Thirds, Slow Dolly In, Slight Low Angle. Low Key, Candlelight, Volumetric Lighting (God Rays). 35mm Anamorphic, f/2.8, Film Grain.

[Dialogue/Audio] Keeper: "She is early this year." (spoken in a low, gravelly, unhurried voice, with quiet dread, after he wipes the glass) Audio: wind against the tower, distant waves, creaking wood, NO MUSIC.

[Avoid] text, watermark, background music, singing, visible flames outside the lantern.

The labels in square brackets are for you and for models that respect them. If your model reads them aloud or renders them as text on screen, delete them and keep the order.

JSON version: one field per job

JSON carries the same information with each part in a named field. Settings that apply to the whole clip go at the top. Each shot is one entry in the scene list, so a second shot can reuse the same style and audio.

{
  "global_parameters": {
    "style_reference": "Cinematic 4K, Photorealistic",
    "audio_ambiance": "wind against the tower, distant waves, creaking wood, NO MUSIC"
  },
  "negative_prompt": [
    "text",
    "watermark",
    "background music",
    "singing",
    "visible flames outside the lantern"
  ],
  "scene": [
    {
      "cinematic_prompt": {
        "visual_details": "A weathered lighthouse keeper climbs the last steps of a spiral staircase, carrying a brass lantern. He stops at the window, wipes the fogged glass with his sleeve and watches a storm roll in over the sea.",
        "camera": {
          "composition": "Medium Shot, Rule of Thirds",
          "motion": "Slow Dolly In",
          "angle": "Slight Low Angle"
        },
        "lighting": "Low Key, Candlelight, Volumetric Lighting (God Rays)",
        "technical_lens": "35mm Anamorphic, f/2.8, Film Grain"
      },
      "dialogue": {
        "speaker": "Keeper",
        "text": "She is early this year.",
        "emotion": "quiet dread",
        "voice_style": "low, gravelly, unhurried",
        "timing_trigger": "after he wipes the glass"
      }
    }
  ]
}

When JSON helps

• Dialogue: the spoken line is separate from the description, so the model is less likely to show it as on-screen text or have the wrong character say it.
• Audio: “no music” is stated once as a setting and repeated in the negative list, which is more reliable than a remark at the end of a paragraph.
• Negative prompts: each unwanted item is its own entry.
• Series of shots: change the scene entry and keep everything else, and the clips match.

JSON is not a requirement and not every model reads it better than text. Field names are not standardized across models, so check what yours expects. When a model takes only plain text, the text version above gives you the same structure.

Common mistakes

• Two camera moves in one shot. Pick one. “Slow dolly in” and “orbit” cancel each other out.
• Lighting that contradicts itself. “High key” and “low key” cannot both be true.
• Adjectives instead of events. “Epic, stunning, dramatic” tells the model nothing it can show. Describe what happens.
• Negatives inside the scene. Saying “no cigarettes” in the description can make cigarettes appear. Describe what is there, and move what must not appear to the avoid list.

Build one yourself

The Prompt Builder follows this structure. Fill in the scene, pick the options, and switch between the text and JSON tabs to copy either format. For the vocabulary, see the shot types and cinematic looks libraries.