Video Prompting

Video Prompting

Overview

Turn a user’s intent into either:

• a strong, model-compliant video prompt, or
• a strong image-model prompt for a character sheet that will later support image-to-video consistency.

Model-specific video guidance lives in references/models/. Character-sheet guidance lives in references/workflows/character-sheets.md. This file is the entry point: route to the right path, ask the minimum clarifying questions, then draft the prompt in the expected format.

Model Index

• Ovi: references/models/ovi/prompting.md
• Veo 3 / 3.1: references/models/veo3/prompting.md
• Wan 2.2: references/models/wan22/prompting.md
• Wan Animate 2: references/models/wan-animate-2/prompting.md
• Seedance 2.0: references/models/seedance2/prompting.md
• Seedance 2.5: references/models/seedance2-5/prompting.md
• MiniMax H3: references/models/minimax-h3/prompting.md
• LTX-2: references/models/ltx2/prompting.md
• LTX-2.3: references/models/ltx2-3/prompting.md
• LTX-2.5: references/models/ltx2-5/prompting.md

Workflow Index

• Character sheets for consistent characters: references/workflows/character-sheets.md

To add a new model later: create references/models/<model>/prompting.md, then add it to this index.

To add a new workflow later: create references/workflows/<workflow>.md, then add it to the Workflow Index.

Global video-prompt rules

These rules apply to every video model reference:

• Never include the model name, model version, duration, aspect ratio, resolution, or API/control parameter names in the final prompt text unless the selected model's required prompt schema explicitly includes timing. MiniMax H3 alignment instructions and shot timestamps are such an exception.
• Otherwise, use duration only as internal planning context for how many action beats the prompt can support.
• If the user asks for parameters or the model requires them, provide them outside the prompt in a separate recommended-parameters line.
• For image-to-video, treat the image as the visual anchor. Do not describe the image in depth unless the user asks for an image analysis or a detail must change. Focus the prompt on motion, camera, emotion/performance, and audio.

Workflow

Step 1 — Route the request

Decide whether the user wants:

• a video-generation prompt, or
• a character-sheet prompt for an image model

Route to the character-sheet workflow when the user wants a reusable reference sheet, turnaround, expression sheet, costume sheet, photographic identity sheet, or a consistent-character starting point for a longer image-to-video project.

If the user is asking for both, do them in this order:

1. Character sheet
2. Scene still / anchor frame
3. Video prompt

Step 2 — If it is a video prompt, identify the model and input mode

If the user did not name a model, ask which model they are using (or offer supported options from the Model Index).

Then confirm the input mode. Start with text-to-video (t2v) or image-to-video (i2v), and use any additional keyframe or full-reference modes supported by the selected model guide.

For MiniMax H3, distinguish T2VA, I2VA, first-and-last-frame-to-video (FL2VA), last-frame-to-video (L2VA), and full-reference mode. Ask for the effective duration when a final-frame alignment or timed cuts require it.

For Seedance 2.5, distinguish text-to-video, image-to-video, reference-to-video, edit, and extend. Ask for the effective duration when a multi-beat timeline is needed, and ask for the intended ending state on long or continuity-sensitive shots. Recommend the shortest duration that fits the action instead of defaulting every request to the model's maximum.

If i2v: ask the user to share the image (optional, but it will help you generate a better prompt). Use the image as an anchor according to the chosen model’s guidance (e.g., keep identity/wardrobe/composition stable; focus your text on motion/camera/what changes).

If the chosen model has versions, duration constraints, or required parameters, ask the minimum questions needed to select the right format (see the model guide). For LTX-2.3 specifically: default to 10 seconds as the external duration setting when duration is missing, ask if the user wants shorter or longer, and scale motion complexity to match that duration. Do not write the duration into the prompt itself.

For LTX-2.5 specifically: distinguish a continuous single shot from a native multi-shot scene, screenplay-style dialogue, Dub-It speech replacement, and Video Editing IC-LoRA. When the user asks for settings, no fixed length is required, and the interface supports it, recommend automatic duration as an external setting. Use a fixed external duration when last-frame conditioning is supplied. Do not write duration or setting names into the prompt itself.

Step 3 — Load the correct reference and follow its format

For video prompts: open the model’s prompting.md from the Model Index and follow its rules strictly.

For character sheets: open references/workflows/character-sheets.md and follow its structure strictly. Treat this as an image-model prompt, not a video-model prompt.

Step 4 — Draft the prompt in the right form

Draft the prompt using the structure and constraints from the markdown file you selected in Step 3.

For video prompts: follow the chosen model’s prompting.md exactly, including its preferred section order, dialogue/audio format, and any shot-structure guidance. Before returning a video prompt, remove any prompt-internal references to model name/version, clip length, aspect ratio, resolution, or generation settings except timing required by the selected model's prompt schema.

For character sheets: follow references/workflows/character-sheets.md exactly, including layout, consistency constraints, and expression-row guidance.

Step 5 — Output

Default: output only the final prompt text. Default formatting: output prompts as a single line with no line breaks unless the user explicitly requests multiline formatting or the selected model guide requires a structured multiline schema. MiniMax H3 is such an exception: preserve its required field names, line order, and blank-line separation. Seedance 2.5 is another exception for complex timed or reference-driven shots: preserve the guide's production-note sections and timeline line breaks. For LTX-2.5, preserve multiline screenplay formatting when dialogue or beat clarity benefits from it; keep single-shot, image-to-video, and multi-shot prose as one paragraph by default.

If the user asks for options: provide 2–3 distinct prompt variants, each fully self-contained and compliant with the model’s formatting.

If the model uses required API parameters (e.g., duration/size), include a short “Recommended parameters” line only when the user has specified them or explicitly asks for them.

If the user wants the full consistency workflow, after the character-sheet prompt also provide:

• one prompt for a first scene still that uses the character sheet as reference, and
• one prompt for the follow-on image-to-video shot