Treat prompts as executable cinematography blocking, not as a shot wish list. First confirm the story, the footage, and the space allowed by the first frame, then design the visual content, character actions, camera, and lighting.
Choose a working mode
- Plan: output the creative direction, the division of labor between real/AI footage, risks, and footage still needed.
- Write: output the storyboard, per-shot generation packages, or model prompts.
- Partial revision: revise only the shots specified by the user, preserving shot numbers and unauthorized changes.
- Diagnose: first analyze the generated video or first and last frames, and report only the causes and revision plan; do not rewrite the entire film unless the user requests it.
- Full production: execute the following workflow in order and pause for confirmation at key checkpoints.
Mandatory principles
- Make camera movement serve the story subject, character actions, and information reveal; do not move the camera for showmanship.
- Treat the input first frame as the spatial fact of the current generated segment. First analyze the visual content, then write the character actions and camera movement.
- Do not let grass, roads, doors, windows, people, buildings, or occluders appear out of nowhere if they do not exist in the first frame and have no visible path of entry.
- Arrange only one primary camera movement per generation, with at most one additional connecting action that is directionally continuous and physically reachable.
- First write the camera start point, trajectory, speed, subject being followed, and end point, then write the lens, lighting, and mood.
- For image-to-video, prioritize describing "which elements move, how they move, how the camera moves, and where it finally stops"; do not reinvent the image.
- Lighting must have an explainable source. First determine the ambient exposure, then shape the subject; lock exposure and white balance.
- By default, split cross-location sequences into two generated segments, connected with a sound bridge, action/shape match, or edit; do not disguise a location change as a long take.
- Reference images are the basis for character identity and costume continuity. Only create an additional character setting sheet when reference images are missing or the user requests it.
- For authorization, likeness, religious, and folk custom issues, only provide clear risk reminders; do not presume to determine on the user's behalf that authorization has been obtained.
Full workflow
1. Establish project constraints
Record the objective, audience, platform, total duration, aspect ratio, narrative throughline, mood, real footage, reference characters, image model, video model, and per-generation duration. Both Seedance 2.0 and Kling can use first and last frames; by default, use an end frame only when there is a genuine need for end-point control.
Read the full project workflow-and-gates.md, and copy project-brief.md。
2. Formulate a footage strategy
Prioritize having real authorized footage carry landforms, snow mountains, lakes, canyons, villages, animals, and humanistic documentary content; have AI carry fixed-character drama, actions that are difficult to reshoot, and controlled transitions. Record watermarks, resolution, black bars, sources, and character and cultural risks.
3. Audit each first frame
Before writing any visual content prompt or camera movement, read source-frame-spatial-audit.mdin full, and fill in the spatial fields in shot-manifest.yaml .
Must answer:
- What already exists in the foreground, midground, and background?
- Where are the subject, entrances and exits, visible paths, and occluders?
- What are the camera's current height, angle, direction, and reachable endpoint?
- Which image content is allowed to occur naturally, and which would cause out-of-thin-air generation?
- Within the specified duration, can the character actions and camera movement be completed physically?
When key first frames are missing, first list the scene images to be supplemented; do not use text to force the model to create unknown space.
4. Write the storyboard and generation package
Write in the order of "narrative goal → time-segmented image content → shot size and composition → character actions → camera blocking → lighting → sound/dialogue → end frame". The image content must be time-segmented, but each time segment describes state changes within the same continuous space and must not be written as multiple edited shots.
Read completely before camera movement camera-motion-library.md; read completely before lighting motivated-lighting-library.md; read when cross-shot continuity is involved transitions-and-continuity.md。
5. Adapt to the generation model
Read completely model-recipes.md. Preserve accurate English terminology in Chinese descriptions. Keep Midjourney prompts within the character count specified by the user or model; for Seedance 2.0 and Kling, the single-segment duration is based on the current task parameters, and do not assume it is always 15 seconds.
Use three rounds of iteration:
- Test only the subject action and one main camera movement.
- Add shot size, speed, and composition landing point.
- Then add focal length, light, and foreground that actually exists in the first frame.
When messy shots appear, first delete the second trajectory and irrelevant style words; do not keep piling on adjectives.
6. Quality check and revision
Use failure-catalog.md to perform eight categories of checks on the generated results: space, action, camera movement, characters, light, continuity, technical, and risk. Run:
python3 scripts/lint_prompt.py prompt.txt --model seedance --max-chars 0
python3 scripts/validate_timeline.py storyboard.md --max-shot-seconds 15
python3 scripts/lint_shot_manifest.py shot-manifest.yaml
When fixing a local shot, keep the shot number stable; update only the timecode and the explicitly affected dependencies.
Key confirmation points
Full production pauses by default at the following nodes:
- Creative direction and narrative throughline.
- Division of labor between real/AI materials and the risk list.
- First-frame space audit and the first three test shots.
- All storyboards and per-shot generation packages.
- Quality check of generated results and final editing suggestions.
When the user explicitly requests batch execution, confirmations may be reduced, but the first-frame space audit must not be skipped.
Output contract
For a single-shot revision, output only that shot; for a full project, prioritize outputting:
00-project-brief.md01-asset-register.csv02-visual-reference.md03-storyboard.md04-shot-manifest.yaml05-generation-prompts.md06-generation-log.csv07-qc-report.md
For the Tibet project, it is also necessary to read tibet-reference.md。