video-generation

Video Generation

Generate AI video using Veo (MCP tools) or Sora (direct API script). This skill covers provider selection, generation modes, defaults, and the draft-to-final workflow.

For full tool references, model IDs, camera reliability data, and negative prompt blocks, see references/provider-details.md.

Provider Selection

Choose before writing prompts.

Need	Provider	Why
Native audio, 4K output, style/asset references, video extension	Veo	Veo 3.1 has the richer media-control surface
1080p production, text-heavy scenes, clean draft/final model split	Sora	sora-2 drafts + sora-2-pro finals is the most reliable loop
High-value hero shot, uncertain which provider wins	Both	Generate with both, pick winner from review evidence

Decision rules:

1. Veo when: native audio matters, you need 4K, you have style-anchor images for cross-clip consistency, or you need to extend an existing clip.
2. Sora when: 1080p is sufficient, the scene has on-screen text, or the draft-to-pro upgrade path saves iteration time.
3. Both when: the shot is expensive to reshoot and the extra cost of a bakeoff is justified.

Generation Modes

Text-to-Video

Prompt describes the scene. Both providers support this.

• Veo: mcp__veo__generate_video -- supports number_of_videos (1-4) for multi-take
• Sora: sora_direct.py create-and-poll -- single generation per call

Image-to-Video (I2V)

Animate a static image. Source image quality is critical.

• Veo: mcp__veo__animate_image -- absolute path to source image required
• Sora: sora_direct.py create-and-poll --input-reference -- auto-resizes/crops source to match output size

Video Extension

Continue an existing clip with new content.

• Veo: mcp__veo__extend_video_clip -- Veo 3.1 only, continues from last second
• Sora: sora_direct.py extend --id <video_id> -- extends by prompt

Styled Generation (Veo only)

Generate video with 1-3 reference images for visual consistency.

• mcp__veo__generate_video_with_style -- Veo 3.1 only
• Reference types: "asset" (preserves composition) or "style" (preserves palette/grain/lighting)

Default Settings

Veo Defaults

Parameter	Draft	Final
Model	veo-3.1-fast-generate-preview	veo-3.1-generate-preview
Resolution	720p	1080p
Duration	6s	8s
Aspect ratio	16:9	16:9
Negative prompt	Always include default block (see references)	Always include default block

Sora Defaults

Parameter	Draft	Final
Model	sora-2	sora-2-pro
Resolution	1280x720	1920x1080 (or 1080x1920 portrait)
Duration	4s	4s (or 8s for extended scenes)
Draft count	2	1

Draft-to-Final Workflow

Generate cheap drafts first, review, then produce the final with the quality model. This saves 60-70% on failed iterations.

Veo Path

1. Generate 2-3 variants with veo-3.1-fast-generate-preview at 720p
2. Review candidates (silence test at 1x, slow-motion scan at 0.5x)
3. Regenerate winner prompt with veo-3.1-generate-preview at 1080p

Sora Path (Preferred -- Automated)

1. Drafts: sora_direct.py production --stage drafts -- generates 2 clips with sora-2 at 720p
2. Review: sora_direct.py review-drafts -- Gemini scores each draft on 5 dimensions (prompt fidelity, temporal stability, surface realism, lighting coherence, text preservation)
3. Finalize: sora_direct.py finalize-from-review -- selects winner, launches sora-2-pro final at 1080p

All three steps require --run-dir with an absolute path. Relative paths break the review subprocess.

Camera Movement Reliability

Camera choice directly affects generation success rate. Summary:

Movement	Success Rate	Recommendation
Static	94-97%	Default for hero shots, UI demos
Zoom	81-87%	Good for reveals and emphasis
Pan	73-85%	Acceptable for environment shots
Tilt	67-81%	Use descriptive phrasing, not "tilt"
Tracking	58-68%	B-roll only, generate 3 variants
Crane	44-52%	Expect retakes
Combined	~29%	Never -- split into separate shots

For movements below 70%, always generate 3 variants. Full percentages and prompt phrasing in references/provider-details.md.

Key Constraints

Veo

• Always pass the negative prompt block (see references)
• All paths must be absolute
• extend_video_clip and generate_video_with_style require Veo 3.1
• Videos stay on Google servers for 2 days -- download promptly
• Named physical light sources required in every prompt (never "well-lit")

Sora

• OPENAI_API_KEY must be set
• --run-dir must be absolute for production workflows
• No real people, public figures, copyrighted characters
• Only non-human character uploads
• 720p drafts auto-map to compatible sizes when final target is 1080p
• I2V auto-resizes source images to match --size

Cost Reference

Provider	Model	Resolution	Duration	Cost
Veo	3.1 Standard	1080p	8s	~$3.20
Veo	3.1 Fast	720p	8s	~$1.20
Sora	sora-2	720p	4s	~$0.85
Sora	sora-2-pro	1080p	4s	~$3.50
Sora	sora-2-pro	1080p	8s	~$5.60

Typical scene (draft + final): $4-6 total.

Prompt Structure

Follow this template for both providers:

[Camera+Lens]: [Subject with physical detail] [Action with force verbs],
in [Setting with atmosphere], lit by [Named physical light source].
Style: [Texture micro-details, film grain]. Audio: [Ambient/SFX].

Rules:

1. Named physical light source -- always specify concrete light ("cool fluorescent overhead strips", not "well-lit")
2. Avoid dynamic physics -- no falling objects, fluids, collisions
3. Force-based verbs -- "grips handle and pulls" not "opens"
4. Micro-imperfections -- append "slightly grainy film-like quality, fine surface texture"
5. Exact descriptors -- copy-paste character/scene descriptions across shots, never paraphrase

Configuration

• Veo MCP server: veo in ~/.claude/.mcp.json
• Veo auth: GOOGLE_API_KEY or GEMINI_API_KEY
• Veo output: ~/Videos/veo-generated/
• Sora script: /Users/fausto_home/.claude/skills/sora/scripts/sora_direct.py
• Sora auth: OPENAI_API_KEY
• Run with: uv run --with requests python <script> <command> [args]

Related Skills

• video-production — multi-shot orchestration, chaining patterns, QA workflow
• image-generation — hero image generation for style anchors
• ffmpeg-production — post-processing, encoding, platform export