YGtemple/video-skills-collection

AI Short Drama & Animation Prompt Library

This is a "Film Language Translator": converting vague feelings like "very sad, very tense, cinematic" into actionable camera, lighting, expression, action, and sound instructions that models can execute. The core method breaks down film and TV creation into 61 "language libraries" with approximately 4,000 entries, which can be combined like ordering from a menu. Materials are sourced from high-vote tutorials like Qianqiu | AI Dream Factory, optimized for a fixed toolchain: Doubao/DeepSeek for writing · Seedream 5.0 Pro for image generation · Seedance 2.5 for video generation.

First, read this file to establish a global view; when you encounter a specific step, only read the specific volume that matches, do not load everything at once.

Two Unbreakable Rules

1. Assets First: Always lock down the character three-view/character sheet, scene top-view + nine-grid, and prop images before generating video. If the face, clothing, or scene is not locked, multi-shot videos will inevitably change faces and collapse.
2. Bind with Reference Images, Don't Describe with Words: For video generation, use Seedance 2.5's universal reference/multi-image reference. Feed the character image, scene image, and staging image together, and @ the corresponding images in the prompt. After the first name mention, use only names, do not repeat the @.

Nine-Step Full Workflow (Landscape 16:9)

1. Script/Outline: Use Doubao/DeepSeek to generate the outline, episode breakdown, character bios, and dialogue → Read 01
2. Assets First: Use Seedream 5.0 Pro to generate character three-views, scene top-views + nine-grid, and prop images → Read 03 (characters), 05 (scenes)
3. Storyboard: Break down into a storyboard table (shot number/timecode/shot size/character/action/emotion/scene/image generation prompt/video generation prompt/dialogue+SFX) with each shot ≤ 5 seconds → Read 02
4. Seedream Image Generation: Generate the first frame of each shot per the storyboard, landscape 16:9. Generate at least 4 candidates for each visual asset and pick the most stable one.
5. Seedance Video Generation: Use multi-image reference to bind characters + scene + staging. Apply a second-by-second timeline. Single segments 5–15 seconds, key shots up to 30 seconds → For camera/action words, read 07/08/09
6. Segment Transition: Use a 1-second technical full-shot staging image and action connection points to bridge segments. Maintain motion direction/screen position/eye line continuity. Cut at the moment of action (turning, opening doors, drawing swords) → Read 02/06
7. Sound: Dialogue is generated directly by Seedance 2.5 (standard Mandarin Chinese, Chinese lip-sync, no TTS). BGM is selected by the assistant and added during the editing phase → Read 10
8. CapCut/Cut Editing: Color grading, remove duplicate frames, unify BGM/SFX. Export as H.264 Baseline + AAC + faststart (ensures direct playback on Windows).
9. Cover/Poster: Write the title and copy directly into the Seedream image generation prompt. The model renders it in one step. DO NOT add text locally with PIL/PS. Use uniform 16:9, ≥ 4 candidates → Read 12

Volume Routing (Read only the corresponding volume based on the user's current need)

What the user wants to do	Which volume to read
Establish a global view, check the toolchain, check the usage of 61 language libraries, overall strategy	references/00-Overview-and-Toolchain-Map.md
Write story outline/episodes/character bios/full script/dialogue/narrative structure	references/01-Script-and-Story-Creation.md
Break down the script into a storyboard table, second-by-second timeline, shot衔接 and transitions	references/02-Storyboard-and-Shot-Transition.md
Character three-views, character sheet, face shapes/features, clothing/fabric, hairstyle/hair, eyes, non-human characters	references/03-Character-Assets.md
Micro-expressions, body language, habitual actions, lip-sync for dialogue, performance direction	references/04-Performance-Micro-Expressions-and-Dialogue.md
Scene top-views, nine-grid environment, architectural spaces, weather, props, atmosphere	references/05-Scene-Assets-and-Environment-Library.md
Multi-character staging, group blocking, virtual production positioning	references/06-Staging-and-Multi-Character-Positioning.md
Push/pull/pan/tilt, camera movement curves, composition, shot sizes, focal lengths, long takes, subjective shots	references/07-Camera-Language-and-Composition.md
Action scenes, second-by-second action, physical forces, weapon clashes	references/08-Action-and-Fight-Design.md
Xianxia magic, hand seals, Dharma manifestations/god forms, light effects/particles	references/09-Effects-and-Xianxia-Magic.md
Dialogue dubbing, BGM selection, sound effects/Foley, mixing, sound-image relationship	references/10-Sound-Design-and-Music.md
Reverse engineering image/video prompts, nine-dimension prompt writing, universal instructions, negative prompts, consistency	references/11-Reverse-Engineering-and-Prompt-Engineering.md
Cover posters, dance/advertising/e-commerce videos and other types	references/12-Cover-Posters-and-Other-Types.md
Full reproduction of a real finished film's frame-by-frame prompts (model: intangible cultural heritage short film Kaiwu)	references/13-Case-Reproduction-Kaiwu-ICH-Short-Film.md

How to Query the 61 Language Libraries

When unsure which "language library" to use, first read Section 4 ("Quick Reference for 61 Language Library Usage") of 00-Overview-and-Toolchain-Map.md. One sentence will locate the library, then refer to the volume indicated in the table to retrieve specific entries. Recommended combination order: Set emotion → Select camera → Add lighting/color → Add sound/performance → Finally check negative prompts.

Fixed Toolchain and Parameters (User's established preferences, to be followed by default)

• Image Generation: Seedream 5.0 Pro, highest quality, no downgrade; landscape 16:9; ≥ 4 candidates for each visual asset; when locking characters/scenes, upload reference images, prepend prompt with "Based on the reference character", reference metric ~68.
• Video Generation: Seedance 2.5 highest quality, no downgrade; landscape 16:9; key shots up to 30 seconds.
• Model Tiering for Credit Saving: Use Seedance 2.0/2.0-fast for drafts, camera positioning, and storyboard rhythm. Reserve 2.5 only for key shots, emotional close-ups, action climaxes, and the final delivery.
• Dialogue: Standard Mandarin Chinese, Chinese lip-sync, generated directly by the video model, no TTS, no Japanese.
• Subtitles: Do not add subtitles or watermarks to videos. The prompt should explicitly state "no subtitles." The user will add subtitles later.
• Cover: Write the title/copy into the image generation prompt; the model renders it in one step. Do not add text locally.
• BGM: Selected by the assistant and added during the editing phase. Only dialogue and ambient sound remain during generation.
• Export: H.264 Baseline + AAC + faststart.

Prompt Output Standards (Must be executed before each video generation)

1. Before generation, fully display the detailed prompt for that shot. The character/scene images provided by the user must be referenced in the prompt using @image. Mandatory use of image references; prohibit pure text generation from scratch.
2. By default, directly use the approved assets + prompt to generate video. Do not produce keyframes/frame-by-frame fixes/first and last frames (unless the user explicitly requests).
3. Storyboard prompts do not require user confirmation for each one; you can directly generate video. The user only reviews the final video output and approves before moving on.
4. Image-to-video prompt three-part structure: Global setting layer (style/lighting/texture/character anchoring) → Character layer (@character image, appearance, personality, emotion, subtext) → Storyboard layer (second-by-second timecode + camera + action + dialogue + sound effects).
5. Total formula: Style + Character (@image) + Scene (@image) + Storyboard (time + shot number + picture + action + SFX + dialogue).

Pitfall Avoidance Checklist

• Do not skip "assets first" and go directly to video generation.
• Do not cram too much information into a single reference image (zoom in on faces for larger proportions).
• Do not use empty words like "cinematic/atmospheric/premium feel." Translate them into actionable camera, lighting, and emotion words.
• Do not use version 2.5 for all drafts.
• Do not add BGM independently to each segment.
• Do not add subtitles to the video.
• Do not add text locally for covers.
• Do not produce vertical screen videos.
• Do not use TTS for dialogue.
• Do not forget faststart when exporting.
• At the end of every prompt, add negative prompts according to Volume 11 (no deformed hands, no facial distortion, no extra fingers, no clothing penetration, no background flickering, lip-sync).

How to Use

• Paste the entire "system prompt" from the volume into the Doubao/DeepSeek dialog box, and append the user's creative seed at the end. Doubao, DeepSeek, and GPT all use the same large language model dialog interface.
• The volumes are large (several are 60,000–170,000 words). When you need a specific prompt, search within the corresponding volume using keywords (e.g., "three-view", "Hitchcock zoom", "Dharma manifestation", "second-by-second timeline"). Do not read an entire volume straight through.