AI Drama Production

Overview

Use this skill to turn a user's rough idea into a complete AI drama or motion comic production Markdown. Two production modes are available:

• 🎬 Film mode — continuous video generation, shot-based, Seedance/即梦图生视频. Best for AI short films, trailers, and cinematic sequences.
• 🎞️ Motion comic mode — panel-based still-image assembly, voiceover-driven, 即梦出图 + 剪映装配. Best for narrative-driven stories, sci-fi/realistic motion comics, and projects where asset consistency and iteration speed matter more than continuous motion.

Always produce a Markdown deliverable unless the user asks for another format.

Quick Start

1. Determine the production mode:

   • Does the user want continuous video with cinematic camera movement? → Film mode
   • Does the user want a voiceover-driven narrative with high-quality still images and keyframe animation? → Motion comic mode
   • Unsure? Default to film mode for visual spectacle, motion comic mode for story-driven efficiency.

2. Ask only for missing essentials:

   • target length and format: trailer, 30-90s short, 3-minute episode, multi-episode series, motion comic
   • genre, tone, visual style, audience, platform
   • story seed: premise, protagonist, conflict, ending preference
   • constraints: tools, language, aspect ratio, available reference images, safety limits

3. If the user gives little detail, choose reasonable defaults and make them explicit.

4. Follow the appropriate workflow:

   • Film mode -> references/workflow.md
   • Motion comic mode -> references/workflow-motion-comic.md

5. Use the relevant templates:

   • Film mode -> references/templates.md
   • Motion comic mode -> references/templates-motion-comic.md

6. Run the review gates in references/review-checklists.md before finalizing.

7. After delivery, explicitly ask what feedback the user has; create a version 2 if needed.

Output Contract

Film Mode

The final Markdown must include:

• project brief and creative direction
• worldbuilding, theme, story structure, and episode/scene outline
• character bible with visual design and reference image prompts
• scene/location bible with reference image prompts
• style bible with color, lighting, lens, rhythm, sound, and forbidden drift
• director analysis: beat-by-beat "讲戏" using concrete actions, camera movement, light, emotion, and sound
• storyboard table with scene duration, references, camera, action beats, dialogue, audio, and generation prompt
• storyboard panel generation: dual-pass prompt workflow (same prompt -> static panel -> review -> inject as @图片 -> video), panel review checklist, panel-to-video reference injection table
• Seedance-style multimodal prompt pack using clear @图片/@视频/@音频 roles
• production checklist, asset workflow, rough-cut plan, audio plan, iteration plan, and quality/compliance review

Motion Comic Mode

The final Markdown must include:

• project brief and creative direction
• worldbuilding, theme, story structure, and scene outline
• character bible with visual design and reference image prompts
• scene/location bible with reference image prompts
• style bible with color, lighting, cinematic texture, and forbidden drift
• voiceover script with per-line timing breakdown (中文按 3-4 字/秒估算)
• panel-to-voiceover mapping table: each voiceover line maps to one or more image panels
• batch 即梦 image prompts with automatic style prefix and character reference injection
• camera motion assignment: push-in / pull-out / pan / zoom assigned per panel
• 剪映 assembly guide with timeline layout, per-panel duration, keyframe motion, audio tracks
• sound plan: TTS or voice recording, ambience, effects, music, and per-scene emotion-to-music mapping

Core Rules

• Treat story as the source of value. Spend more effort on premise, conflict, reversals, emotion, and audience retention than on tool trivia.
• Build visual consistency early. Create character and scene reference prompts before writing final video prompts.
• Treat the first confirmed core location as a reusable production anchor; design later shots around it when consistency matters.
• Write video prompts as narrative film direction, not keyword piles.
• Describe what changes in the shot. Do not repeat static details already visible in reference images.
• Make every storyboard row image-actionable: who/what, where, camera viewpoint, shot size, expression, and the relationship to the previous/next shot.
• Keep each continuous shot within beat density: about 1 action beat per 2.5 seconds. For 5 seconds, 1-2 beats is usually enough.
• For 10-15 second generations, use time ranges such as 0-3s, 3-7s, 7-12s, 12-15s.
• Reserve the first and last 0.5 seconds for setup and natural settling; avoid key action or dialogue there.
• Label every @ reference by purpose, such as "以 @图片1 中的女主为主角" or "参考 @视频1 的运镜节奏".
• Respect platform limits when writing per-shot prompts: images up to 9, videos up to 3, audio up to 3, mixed input total up to 12, video reference/audio total durations up to 15s each, generation length 4-15s.
• Avoid negative wording in prompts. Replace "不要切镜" with "全程一镜到底", and "不要说话" with "角色保持沉默".
• Avoid real-person face reference material, copyrighted IP dependence, political sensitivity, sexualized minors, explicit sexual content, graphic violence, and unsafe imitation.
• Before video generation, rough-cut the still storyboard images to check shot size, perspective, dialogue timing, and scene continuity.
• Plan sound like performance: dialogue/voiceover sets timing, ambience and effects carry scenes without music, and music should mark genre, suspense, reveal, or climax.
• One prompt, two passes — static panel first, video second. Each per-shot video prompt serves two roles. First pass: run it through a text-to-image model to generate a static storyboard panel — omit time-coded segments (0-3s, 3-7s), motion descriptions, and camera movement; keep only the keyframe composition, character placement, lighting, and expression. Second pass: after reviewing the panel, inject it as an @图片 reference back into the same prompt and run it through the video model. The video prompt now only describes what changes (motion, camera movement, dialogue, sound) since static details are already visible in the panel. This catches composition problems at image-generation cost instead of video-generation cost, and gives the video model an exact visual anchor.

Motion Comic Core Rules

• Voiceover is the timing anchor. Every panel is driven by a voiceover line; panel duration equals voiceover line duration + 0.5s transition buffer.
• One script line maps to one or two panels. If a voiceover line exceeds 15 words (中文), split into two panels.
• All image prompts in a motion comic must share a unified style prefix. Build the prefix once from the style bible and prepend it to every panel prompt.
• Every character reference image must be referenced in every panel where that character appears. Do not rely on the model "remembering" the character.
• Limit 即梦 image prompt reference images to 3-4 per generation: 1 style reference + 1-2 character references + 1 scene reference.
• Use only 4 camera motions: push-in (进入内心/强调), pull-out (释放/远去/结束), pan (场景过渡/环境展示), zoom (强调细节/反转). Assign by the panel's emotional function, not by visual variety.
• Maintain color temperature consistency between adjacent panels. If panel N is warm indoor, panel N+1 should not jump to cold outdoor without a transition panel.
• If a character has costume or state changes across the story, generate a new character reference image for each variant state and label it clearly.
• Rough-cut still panels in 剪映 before generating all images: lay out placeholder images in timeline, check pacing with scratch voiceover, then replace with final generated images.
• Sound design for motion comic is simpler than film but more deliberate: voiceover TTS, scene ambience, key sound effects, and background music. Each serves a different track.

Reference Loading

Load these only when needed:

Film Mode

• references/workflow.md: full end-to-end film production process
• references/templates.md: intake questions, Markdown skeleton, character/scene/prompt templates
• references/production-practice.md: practical image/video/audio/editing tactics from finished AI short production

Motion Comic Mode

• references/workflow-motion-comic.md: full end-to-end motion comic production process
• references/templates-motion-comic.md: panel mapping, batch prompt, motion assignment, and 剪映 assembly templates

Both Modes

• references/review-checklists.md: scoring gates and repair strategies

Source-Derived Principles

The workflow is distilled from six Feishu Wiki sources about AIGC video setting design, Seedance 2.0 multimodal usage, AI short drama production shifts, Agent + Skill storyboard systems, a viral AI video production postmortem, and the award-winning Jimeng short "梦镜" production breakdown. Preserve these learned principles:

• Start with four foundations: world, character, story, style.
• Convert scripts into visible physical actions before generating prompts.
• Use a producer/director/art-designer/storyboarder mental model even when one person does all work.
• Use reference images to lock identity, costume, scene layout, color, and atmosphere.
• Use short 10-15 second scenes as testable units, often with one reversal or audience-retention hook.
• Let generation results feed back into the next script iteration; do not force a fully frozen long script too early.
• Use AI for structure, but manually audit storyboard shot size, viewpoint, expression, and adjacent-shot continuity.
• Combine tools pragmatically: text-to-image drafts, high-quality image iteration, Photoshop/layering/inpainting, first/last frames, image upscaling, frame interpolation, still-image rough cuts, voice/audio passes, and final editing all matter.