youngstudent2/create-handdrawn-mv

Hand-Drawn Music Video & Character MV Production

Translate the rhythm, camera language, and character performance of a reference video into actionable storyboards, then produce assets and a final video. Design each shot as a unit of "what changes." Do not compensate for insufficient character performance by adding transitions or scaling static illustrations.

1. Confirm Task & Evidence

• Carry forward the character, source, style, duration, and revision requests the user has already provided; proceed directly within the scope of what's already authorized. Only ask questions that genuinely block production.
• Check existing projects and final videos; preserve versions already approved. Confirm that assets, audio tracks, and tools are actually usable before committing to delivery.
• When receiving problem frames or feedback about anatomy, fingers, or transitions, first read repair-anatomy-and-transitions.md, map screenshots to shot and pose switch points, and distinguish between original art errors, compositing ghosting, and motion discontinuity.
• Actually watch the user's reference video, combined with frame-extraction checks for beat, shot scale, composition, movement, layers, and transitions. Titles, thumbnails, and text descriptions are not substitutes for watching; when unable to read them, state this clearly and proceed with the portions you can work on.
• Extract a small set of clear identity anchor points from the character source. Distinguish between the character design source and the video style source, and avoid transplanting character appearances from the style reference.
• Record the music source and clip interval; do not claim the referenced audio track as original or assume commercial license. Do not add approval processes based on hypothetical risks.
• Limit tool research to specific gaps: check local capabilities, consult official documentation when necessary, run short tests, then choose a tool. Do not replace actual verification with "I already know this tool."

2. Establish Performance First, Then Asset Quantity

Read storyboard-and-art.md. Copy and rewrite the checklist from storyboard.example.json.

• Plan shots based on music and narrative; do not mechanically apply a fixed number of shots. First write out the timeline, emotional development, and action changes clearly.
• For each character performance shot, plan at least two distinct key art poses A/B with visible differences in action or expression; for more complex actions, add keyframes or in-betweens.
• If the user requires "two images per shot," set requirements.all_shots_need_character to true, and ensure all shots have at least two independent character images, including effect shots.
• min_character_keyframes indicates the minimum number of images per character shot; unique_assets defaults to true. When the user explicitly allows static holds, asset reuse, or different quantities, configure per task and document the counting criteria.
• Distinguish between shot count, independent character image count, and video frame count. Flipping, cropping, recoloring, scaling, or repeated exports of the same image do not constitute new poses. 19 shots × 2 images = 38 images is the scale of the Cookiya pink-dress case, not a template for all 30-second videos.
• For each shot, write down the A→B action, shot scale, camera, background, foreground, effect purpose, transition, visual focus, and acceptable cropping. Two images "looking different" does not mean the action is valid.
• Run python scripts/check_project.py project.json --plan-only to check the structure, duration, and planned asset quantity before starting batch generation.

3. Create Paired Key Art & Layered Assets

• Use the current environment's built-in image generation tools and follow their skill. First review reference images; prioritize generating A using approved identity anchor points, then edit B using A and the same identity anchor points. For each submission, clearly state what to keep and what to change.
• Lock identity, clothing, art style, lighting, canvas format, and camera angle; make expressions, eye lines, gestures, or body silhouettes produce purposeful differences. Do not independently generate A/B randomly and then forcibly combine them.
• Output character assets with real transparent backgrounds, leaving motion space for hair, hands, skirt hems, and shoes. When a full body shot is required, missing feet or clipped hair is not acceptable; chest-up natural cropping should be handled in compositing using a frame or close-up.
• Produce background, character, props, and foreground separately; use image generation for natural images, and programmatic drawing for precise circles, text, masks, and particles.
• Name files as shot_01_A.png / shot_01_B.png, and record the actual prompt and source. Keep approved versions; repair problematic areas locally; do not easily redo everything.
• Check each image for character consistency, left/right accessories, fingers, arm joints, eyes, clothing, and transparent edges. Quantity or hash checks cannot replace these visual checks.
• For full-body actions, first verify shoulder-hip alignment, joint directions, center of gravity, and support foot; for hands, first verify palm orientation, thumb side, knuckle alignment, and contact relationships. Do not forcibly spread fingers that should naturally be occluded just to show all five. After repairing A, still edit B from A and re-check motion continuity.
• Parallel work is only allowed when permitted by current rules; if splitting work, assign complete A/B pairs to the same executor, unifying anchor points, naming, and handoff checklists.

4. Composite Shots with Performance and Rhythm

Read compositing.md, select tools based on the project, and do not default to a specific editor.

• For collage pieces requiring programmable perspective, mesh deformation, and GLSL transitions, use ModernGL/EGL compositing with FFmpeg encoding; simple tasks can use existing editors or FFmpeg directly. First render a representative shot and the most difficult transition.
• Prioritize expressing motion, then layer camera work, depth of field, prop responses, and effects. Shot scale and composition should also vary; continuously pushing and pulling the same illustration does not constitute sufficient performance.
• Use the same canvas and the same combined crop box for A/B, keeping the head, torso, or designated anchor point aligned. Transition between poses according to action and music; do not apply the same slow fade to all shots.
• For hands and bodies with clearly different contours, default to exposing only one pose per frame; when differences are large, add in-betweens or change the action design. Check for unintended A→B→A→B backtracking; do not use a "breathing" cycle to repeatedly replay full gestures. For a standing character, maintain the support foot's relationship with the ground; do not fake a turn using rotation, scaling, or mesh wobble of the whole character.
• Use a consistent premultiplied alpha convention when blending transparent assets, and check for black borders and double contours. Two key poses plus compositing should be honestly called keyframe collage / limited animation, not frame-by-frame hand-drawn.
• Place strong kaleidoscope, recursive frames, distortion, and similar effects in the background or on independent layers; when the action needs to be clearly visible, draw the main character after these effects.
• Foreground ribbons, particles, and motion blur should avoid faces, hands, and key props. Effects should have rhythm and purpose, and also give the audience time to see the character.
• Transition direction should follow action or eye lines, with intensity varying with the music; avoid applying equally intense effects at every cut point.
• During rendering, record which assets are actually drawn and their effective visible frame count (if measurable). Assets existing in a folder does not mean they entered the final video; being included in the draw call does not mean they are ultimately visible.

5. Validate and Deliver in Phases

1. Check the full A/B asset overview; fix repeated poses, missing images, wrong hands, wrong accessories, and unusable crops.
2. Render low-resolution short clips covering action changes, the most complex effects, transitions, and end cards. Watch the motion, not just screenshots.
3. Render the final video at the target resolution; verify duration, resolution, frame rate, audio track, start/end, and full decode.
4. Run python scripts/check_project.py project.json --video output/final.mp4 --usage output/usage.json --decode --contact-sheet output/keyframes.jpg --report output/qa.json. Requires Pillow; video checks require FFmpeg/ffprobe. Parameters and limitations are in validation.md.
5. Manually watch the entire video, especially checking whether A/B is actually distinguishable, whether the character is occluded, whether the beat and lip-sync are in sync, and whether pose switches cause jumping. Passing automated checks does not mean visual quality is approved.
6. Deliver the final MP4, an asset overview, and a reusable project (assets, timeline, code/project files, sources, and dependency documentation). Save according to the current environment's persistence rules, and provide a playable final video. Respect scope when the user requests only parts.

During rework, additionally extract problem poses from the final encoded video plus at least three frames before and after each modification cut point. Check motion order, foot positions, hand ghosting, text, and cropping. When the user requests to delete bottom text from the end card, delete the corresponding text layer or draw call and check at multiple end card timestamps; do not cover it by cropping the character or overlaying content.

Report the actual completed counts, specific improvements, inspection results, and any remaining limitations. Tools that were not run, images that were not generated, and videos that were not watched cannot be reported as completed.

Reusing the Cookiya Case

Only read cookiya-case.md when the task involves Cookiya or requests reuse of this project. Preserve the production lessons from it; do not impose that character, color palette, shot count, or music onto a new task.