Video Prompting Overview
Turn a user’s intent into a strong, model-compliant video prompt by routing to the correct model guide and applying its formatting/tokens.
Model-specific guidance lives in references/models/. This file is the entry point: pick the model, ask the minimum clarifying questions, then draft the prompt in that model’s expected format.
Model Index Ovi: references/models/ovi/prompting.md Sora (Sora 2): references/models/sora/prompting.md Veo 3 / 3.1: references/models/veo3/prompting.md Wan 2.2: references/models/wan22/prompting.md LTX-2: references/models/ltx2/prompting.md
To add a new model later: create references/models/
Workflow Step 1 — 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:
Text-to-video (t2v), or Image-to-video (i2v)
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).
Step 2 — Load the model reference and follow its format
Open the model’s prompting.md from the Model Index and follow its rules strictly (tokens, audio formatting, parameter constraints, recommended structure).
Step 3 — Draft the prompt as a coherent clip
Default structure (adapt to the model’s style and required sections):
Subject(s): who/what, distinctive details Setting: where/when, lighting, mood Action progression: what changes over time (start → beat → beat → end) Camera: framing/movement only if it matters Dialogue/audio: only if the model supports it, using the model’s exact format
Avoid keyword soup. Prefer a single, well-described shot unless the user explicitly wants multiple cuts/shots.
Step 4 — Output
Default: output only the final prompt text.
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.