Break a single comprehensible power display into continuous states, then hand the motion between adjacent states to video generation. Deliver the plan, keyframes, or finished film according to the user's request for this instance, without automatically expanding the scope.
Take over from the current progress
- Has an approved style image: use it directly as the master template, without redoing a broad style selection.
- Has approved continuous frames: review inter-frame contradictions, then produce the video, without regenerating images.
- Only wants the plan, Skill, prompts, or keyframes: complete the corresponding part, without automatically submitting a paid video task.
- When the user says "continue after selection," stop at the selection point; if already selected or already authorized, do not confirm again.
- When non-critical preferences are missing and autonomous decision-making is allowed, proceed with reasonable defaults. Rainy night, eight frames, 25 seconds, knife, and soundtrack are not fixed requirements.
Can receive novel-distill event cards, or directly use the user's concept, without needing to install an extra Skill. Distinguish original facts F, audiovisual treatment P, and added adaptation A; domains, manifestations, and weapons in reference videos must not be back-fed as novel facts. When there is no original text, call it a visual concept interpretation.
Tool selection: platform-agnostic
This Skill fixes the production method, without designating a default finished-film platform. Image generation, image-to-video, and editing/mixing can be done by different tools or on the same platform.
- When the user specifies a tool or an existing project exists, prioritize using it and verify its actual capabilities; do not automatically choose Pexo just because historical cases used it.
- When unspecified, choose based on currently available tools, reference-image constraint capability, duration, downloadable output, and budget, and briefly explain the reason. If a usable plan already exists and the user allows autonomous decision-making, there is no need to set up a separate confirmation step for platform selection.
- When there is no callable generation tool, deliver standalone keyframes, a general segmentation brief, and prompts, explaining the missing capabilities; do not describe prepared materials as an already generated video.
- First verify the current interface or UI, then map the general brief to specific parameters. A platform name does not equal model capability; do not assume every platform supports multi-image reference, first/last frames, or native sound effects.
1. Lock the same-frame relationship and continuity
Use a brief continuity table to record existing information:
| Item | Content to lock |
|---|---|
| Character | Face, age, hairstyle, robe silhouette, material, identifying object |
| Manifestation | By default it is one's own enlarged manifestation, with the same face and outfit; if the user specifies a deity/aberrant form, record its corresponding relationship |
| Same-frame space | True body in the foreground, manifestation behind and taller than buildings; the small figure's silhouette and the giant manifestation's face are both readable |
| Body and picture coordinates | Anatomically which hand holds what, and where the start and end points are in the frame |
| Scene | Architectural landmarks, light sources, weather, camera positions, and action axis |
| Material evolution | E.g., silver-gray phantom → skin-toned clothing material → edge dissipation; do not impose it if the user does not need it |
| Attack | Start → charge → release → settle, with count, direction, force target, and environmental response |
Same-action mapping, not a horizontal mirror. The physical body's anatomical right arm corresponds to the dharma-form's right arm; do not write an ambiguous "mirror pose" that causes left-right flipping. The dharma-form must not become a second opponent in the distance; a separate large-face close-up cannot replace the required key frame with both in the same frame.
By default, use a single camera position from the approved master or a slight push-in to reduce continuity difficulty; when the user wants multiple camera angles, specify the cuts and axis clearly, and do not turn single-camera into a style prohibition.
2. Design the state changes first, then generate independent frames
Read references/keyframes.md. Record in each frame the physical body's pose, the dharma-form's pose/opacity, weapon state, environmental response, differences from the previous frame, and actions that have not yet occurred. The differences should be visible in the image.
The common structure is "gathering momentum → domain first opening → initial appearance → unfolding → solidifying → gathering the blade and raising the hand → synchronized release → aftermath", and it can be merged or split. When the span between gathering the blade and raising the hand is too large, add a transition state instead of stuffing more requirements into the same prompt.
Use the currently available image generation/editing tools; when the user specifies a platform or model, confirm the actual callable capabilities. If the interface does not return a version, record the channel and "version undisclosed"; do not falsely claim a specific model such as GPT-Image 2.
One independent image at a time, maintaining the aspect ratio. Usually pass in "approved master + previous approved frame": the master maintains identity, and the previous frame maintains state. If the previous frame has errors, fix them first; do not propagate errors, and do not throw in all images at once. Early frames should clearly suppress the climax in the master: before the manifestation appears, there must not already be a giant face; before release, there must not already be a complete blade light.
After each frame is generated, look at the image before continuing. If only the dharma-form's arm is wrong, make a targeted fix and preserve the correct physical body, face, clothing, and scene. Reasonable differences in local texture are allowed, but major contradictions in identity, hand side, and action stage must be resolved first.
3. Assign inter-frame motion to video segments
Read references/video-and-qa.md. Segment according to duration capability and action complexity, with adjacent segments sharing one approved boundary state. A shared image does not guarantee a natural continuation after generation.
For each segment, write the start state, end state, process reference, main action, camera movement, correspondence between both parties, environmental forces, sound beats, and ending pose. Images fix appearance and state; video prompts describe motion between states, not a slideshow of still images.
If there are dedicated first/last frame fields, use them according to the interface; ordinary image_list can only be called ordered references. When there is no last-frame constraint, check the actual segment end, and if necessary use the approved segment end as the first frame of the next segment. Submitting independent clips at the same time is only suitable when the boundaries are clear and the tool supports the corresponding constraints.
Only read the optional adaptation note when Pexo is actually used this time references/pexo.md. It is not a default entry point or a required dependency; other tools execute according to the general brief and current capabilities.
4. Acceptance, correction, and compositing
Look at the actual material first, then decide whether to regenerate. Prioritize fixing same-source relationships, hand side and direction, repeated attacks, action regression, and major identity drift. Judge material deviations against the originally intended goal; a piece of self-check text from the platform cannot directly trigger a paid retry.
Same frame does not equal synchronized. The physical body has already sheathed the blade while the dharma-form is still raising the blade is a key error. Check before and after release and every connection point; use reasonable cropping, matching cut points, and sound-effect bridging to handle small differences, and do not use reverse playback or long dissolves to hide incorrect poses.
Record the original and revised versions, and do not overwrite the only successful asset. Address clear defects in each round; reserve retries and later-stage allowance. User-given limits must not be raised without authorization, and deviations from the assistant's own cost targets must also be reported. Count multi-segment revisions item by item in consumption, and do not mistakenly write "one round" as a free retry.
Arrange sound as buildup, release, and aftermath, retaining effective ambient and action sound effects. If the score is too long, trim and fade it out first; only consider regenerating if the content is unsuitable. "Replace the audio track" must not be claimed as "overlay and retain." If the platform cannot mix audio, external post-production may be completed under existing authorization, but the division of labor must be stated truthfully.
Local technical check example:
python3 scripts/audit_video.py /absolute/path/final.mp4 --cuts 6.04,12.08,17.12 --out /absolute/path/qa
Cut points must be filled with actual timeline times. Requires Python 3, ffmpeg, and ffprobe. The script outputs media parameters, a full overview, frame grabs at edit points, and black-frame and volume detection; it does not judge character sync, perceived sound, or whether there is speech, so visual inspection and appropriate audio checks are still required.
Delivery
Deliver independent keyframes and/or a downloadable final cut according to scope, retaining the status table, prompts, and production records. Record the actual tools/confirmable models, timeline, retries, consumption, and residual deviations. For the final cut, also record the source assets and output specifications; do not describe upscaling as native high resolution, and do not treat the platform's "complete" as acceptance.
Finally, directly present the work and important limitations. When a real case is needed, read references/worked-example.md; historical prices and results do not constitute promises of success rate, budget, or follower growth.