Seedance Video Director

A production-quality AI Skill for Seedance 2.5 video prompt generation. It works as a virtual director — you describe your idea, it outputs a complete, copy-paste-ready Seedance 2.5 prompt.

Current version: 0.18.0

Supported Content Modes

🎬 TVC (Television Commercials)

Create brand commercials, product ads, and launch films with professional advertising structure.

• Three durations: 15 seconds / 30 seconds / custom;
• Three audio options: with voiceover / without voiceover / generate two versions simultaneously;
• Two generation methods: direct generation / nine-panel grid storyboard preview;
• 0–3 second ad hook and attention-capture checks;
• 30-second multi-scene progression, shot density, and anti-repetition control;
• Single voiceover truth source with hard stop time and repetitive speaking prevention;
• Contemporary beauty casting, multi-person identity, and seating continuity;
• Realistic force chain for product actions (opening bags, dropping liquid, smearing, pouring);
• Visible evidence chain for food mouthfeel and specific flavors;
• Nine-panel grid storyboard without numbers, timecodes, or labels;
• Generated output post-mortem with frame-by-frame extraction and timecoded transcription;
• Action completion state, transition empty frames, hero reveal, and packshot budget checks.

📱 Vlog (Authentic Smartphone Content)

Create food vlogs, travel vlogs, UGC product reviews, and lifestyle content that looks genuinely recorded on a real phone.

• Five vlog types: solo selfie / boyfriend POV / UGC product review / lifestyle day-in-life / casual home recording;
• Smartphone camera realism: autofocus hunting, exposure shifts, handheld shake, motion blur, imperfect framing;
• Character identity lock: face, hair, clothing, and proportions preserved across all shots;
• Natural dialogue: spontaneous in-the-moment speech, not voiceover narration;
• Scene flow continuity: logical location transitions, forward-only time-of-day progression;
• Food and product interaction: realistic textures, natural tasting reactions, correct product physics;
• Environmental authenticity: locale-specific details, natural background activity;
• Aspect ratio support: 16:9 horizontal / 9:16 vertical (TikTok/Reels) / 4:5 (Instagram);
• Generated output review with identity drift detection, camera realism checks, and localized retry.

🎥 Short Film (Cinematic Narratives)

Create cinematic short films across 10+ genres with film-level production quality.

• Action/Thriller — chase sequences, heist, disaster, survival drama;
• Fantasy/Sci-Fi — dragon flights, kaiju battles, tokusatsu transformation, magic;
• Animated Comedy — 3D physical comedy, animal stories, slapstick;
• Romantic Drama — meet-cutes, emotional moments, relationship stories;
• Historical/Biographical — period pieces, life journeys, cultural narratives;
• Horror/Mystery — body horror, supernatural, tension, unexplained events;
• Dance/Performance — choreographed sequences, music-driven;
• Concept/Art Film — time-travel, abstract, one-take continuous shots;
• War/Military — battlefield sequences, first-person immersive flight;
• Xianxia/Wuxia — martial arts fantasy, sword duels, deadpan comedy.

Key capabilities:

• Full cinematic camera language: shot sizes (EWS to ECU), lens equivalents (16-200mm), dolly, crane, orbit, tracking;
• Detailed character design: physical identity, distinguishing marks, wardrobe per scene, performance direction;
• Multi-character identity lock: exact count, spatial positions, non-interchangeable visual markers;
• Action choreography: chase design, fight choreography, obstacle interaction, physics-based movement;
• World-building: period accuracy, color palettes by genre, environmental consistency, reference image integration;
• Audio design: in-scene dialogue with lip-sync, Seedance 2.5 audio format, cinematic sound design;
• Shot structure: hard cuts with specified count, continuous one-take, match cuts, physical wipe transitions.

Installation

Codex (OpenAI)

# Option 1: Clone directly into Codex skills directory
git clone https://github.com/zero2launch/seedance-video-director.git /tmp/sdv \
  && cp -R /tmp/sdv/seedance-tvc-director ~/.codex/skills/seedance-video-director \
  && rm -rf /tmp/sdv

# Option 2: Keep as project dependency
git clone https://github.com/zero2launch/seedance-video-director.git
# Then instruct Codex: "Read seedance-video-director/seedance-tvc-director/SKILL.md as your skill"

After installation, verify:

ls ~/.codex/skills/seedance-video-director/SKILL.md
# Should show the file

Claude Code (Anthropic)

# Option 1: Clone directly into Claude skills directory
git clone https://github.com/zero2launch/seedance-video-director.git /tmp/sdv \
  && cp -R /tmp/sdv/seedance-tvc-director ~/.claude/skills/seedance-video-director \
  && rm -rf /tmp/sdv

# Option 2: Keep as project dependency
git clone https://github.com/zero2launch/seedance-video-director.git
# Then tell Claude Code: "Read seedance-video-director/seedance-tvc-director/SKILL.md as your skill"

After installation, verify:

ls ~/.claude/skills/seedance-video-director/SKILL.md
# Should show the file

Antigravity / Gemini (Google)

# Option 1: Clone directly into Gemini skills directory
git clone https://github.com/zero2launch/seedance-video-director.git /tmp/sdv \
  && cp -R /tmp/sdv/seedance-tvc-director ~/.gemini/skills/seedance-video-director \
  && rm -rf /tmp/sdv

# Option 2: Keep as project dependency
git clone https://github.com/zero2launch/seedance-video-director.git
# Then instruct the agent: "Read seedance-video-director/seedance-tvc-director/SKILL.md as your skill"

After installation, verify:

ls ~/.gemini/skills/seedance-video-director/SKILL.md
# Should show the file

ChatGPT (Manual)

Since ChatGPT doesn't have a file system, paste the contents of SKILL.md into:

• Custom Instructions (Settings → Custom Instructions → System Prompt), or
• The beginning of a conversation ("Here is your skill: [paste SKILL.md content]")

Any Other Agent

Place the seedance-tvc-director folder into the Skills directory supported by your agent platform. The Skill will use whatever image generation capabilities the agent already provides.

Usage

Skill invocation syntax differs by platform:

• Codex: type / to open the skill picker, then select seedance-video-director
• Claude Code / Antigravity / Gemini: type $seedance-video-director

The examples below use the $ prefix. Replace with / if you are using Codex.

TVC

Use $seedance-video-director to make a 30-second TVC for this product.
Choose the audio option with voiceover, and the generation method as direct generation.
Here is the product image and selling points; please give me the complete prompt first.

Vlog

Use $seedance-video-director to create a 30-second Tokyo food vlog.
Solo selfie style, 16:9 horizontal.
The creator discovers a hidden dessert café and tries a strawberry cream dessert.
Give me the complete Seedance 2.5 prompt.

Short Film

Use $seedance-video-director to create a 30-second survival drama short film.
Two climbers on a collapsing bridge in the Himalayas. 13 hard cut shots.
Include flashback memories. One rescue moment at the climax.
Give me the complete Seedance 2.5 prompt with full character design.

If no configuration is specified, the Skill will first determine the content mode (TVC, Vlog, or Short Film), then confirm the relevant options:

TVC options:

1. Duration: 15 seconds, 30 seconds, or custom seconds;
2. Audio: With voiceover, without voiceover, or generate two complete versions simultaneously;
3. Generation method: direct generation or nine-panel grid.

Vlog options:

1. Duration: 15 seconds, 30 seconds, or custom seconds;
2. Vlog type: solo selfie, boyfriend POV, UGC product review, lifestyle day-in-life, or casual home recording;
3. Aspect ratio: 16:9 horizontal, 9:16 vertical, or 4:5.

Short Film options:

1. Duration: 15 seconds, 30 seconds, or custom seconds;
2. Genre: action/thriller, fantasy/sci-fi, animated comedy, romantic drama, historical/biographical, horror/mystery, dance/performance, concept/art film, war/military, nostalgic/slice-of-life, or xianxia/wuxia;
3. Shot structure: number of hard cuts, continuous one-take, or user-specified;
4. Aspect ratio: 16:9 horizontal (default), 9:16 vertical, or 4:5.

Iterating with Generated Output

After generation, hand the video back to the Skill for review and retry:

Use $seedance-video-director to review this generated video.
Keep the successful product, characters, and voiceover.
Fix only the main issues based on the timecodes.

Provide simultaneously:

• Generation platform and version;
• Configuration (duration, audio, generation method);
• Prompts actually used and reference image responsibilities;
• Original generated video or accessible link;
• Accurate timecodes of the problems you identified;
• Successful parts you want to retain.

The Skill will separate successful layers from failed layers, then output retained items, the single main correction layer, and retry instructions.

Project Structure

seedance-tvc-director/
├── SKILL.md                          ← Main routing file (the "brain")
├── agents/openai.yaml                ← OpenAI agent config
├── evals/evals.json                  ← 41 regression test cases
└── references/
    ├── [12 TVC reference files]      ← Ad hooks, structure, voiceover, casting, QC...
    ├── [9 Vlog reference files]      ← Camera realism, dialogue, scene flow, examples...
    └── [9 Short Film reference files] ← Camera language, genres, choreography, world-building...

The Skill reads SKILL.md first to determine the content mode, then reads relevant reference files on demand.

Contributing

You are welcome to use this Skill to generate cases and submit post-mortem materials via GitHub Issues. Suggested title format:

[Case Feedback] Content mode / Product or genre / Duration / Generation method

The Issue body should ideally include:

• Generation platform and version
• Duration and audio mode
• Generation method
• Product category or genre
• Prompt or relevant snippets
• Video link
• Problem timecodes
• Parts to retain
• Results you want to improve

Please do not submit confidential briefs, unauthorized client materials, personal privacy, or videos that cannot be made public.

Boundaries

• This is a director and prompt compilation Skill; it does not guarantee generation quality, parameter limits, or commercial viability on any platform;
• Exact logos, small packaging text, prices, CTAs, disclaimers, and broadcast-grade voiceovers should be completed in post-production;
• Unverified efficacies, data, certifications, and absolute claims must be retained as [TBD: ...];
• Seedance 2.5's actual capabilities may vary over time and across platforms; verify with the current platform before use.

License

Code and documentation use the MIT License. Third-party brands, products, and trademarks mentioned in the repository are not covered by this license.