Gemini Omni Flash

Gemini Omni Flash Master Orchestrator

1. Skill Architecture & Token Optimization

This Skill leverages a filesystem-based architecture to implement Progressive Disclosure. Instead of consuming upfront context with large, monolithic prompts, information is loaded in three discrete stages to minimize token penalties:

• Level 1: Metadata (Always Loaded): The YAML frontmatter above is included in the system prompt at startup (~100 tokens). This allows the agent to discover that this capability exists without penalizing the context window.
• Level 2: Core Instructions (Loaded When Triggered): The body of this SKILL.md file contains high-level routing knowledge[cite: 1]. It is read from the filesystem via bash only when the user's request matches the model criteria (< 5,000 tokens)[cite: 1].
• Level 3: Specialized Resources (Loaded On-Demand): Detailed SDK blueprints, API parameter matrices, and REST schemas reside in standalone reference markdown files[cite: 1]. The agent triggers them conditionally using bash commands, ensuring an effectively unlimited reference scope with zero token penalty for unused contents[cite: 1].

2. Structural Sub-Reference Routing

When a developer requests implementation details, do not synthesize raw code from scratch. Execute bash: read references/<filename>.md to load the exact parameter schemas, regional restrictions, and SDK blueprints into the context window[cite: 1]:

A. Text-to-Video Requests

• Trigger Condition: User asks to generate a 10-second video from a text description, requires the primary SDK setup, or requests the raw REST JSON structure[cite: 1].
• Action: Read references/text_to_video.md[cite: 1].

B. Aspect Ratio Configuration

• Trigger Condition: User needs to change framing layouts between widescreen landscape (16:9) or vertical mobile portrait (9:16)[cite: 1].
• Action: Read references/aspect_ratio_control.md[cite: 1].

C. Image-to-Video & Subject Reference

• Trigger Condition: User uploads one or multiple static images to animate them, sets a baseline starting frame, or maps distinct subject image references into a single scene prompt[cite: 1].
• Action: Read references/image_to_video.md[cite: 1].

D. Stateful Video Editing (Chaining)

• Trigger Condition: User requires iterative, multi-turn video updates (e.g., background swaps, adding objects, altering camera viewpoints) using conversation history without re-prompting the whole scene[cite: 1].
• Action: Read references/stateful_editing.md[cite: 1].

3. High-Performance Prompting Framework

When guiding users on constructing text inputs for Gemini Omni Flash (API name: gemini-omni-flash-preview / EAP alias: bouncybohr), enforce the 6-Dimension Multimodal Prompt Framework to leverage the model's native world reasoning instead of using cluttered adjective stacks[cite: 1]:

1. Shot Framing & Motion: Define explicit perspectives ("close-up on shoes", "over-the-shoulder") and velocities ("glide gently", "rush suddenly", "dolly zoom", "one continuous shot/oner")[cite: 1].
2. Style Aesthetics: Specify clear artistic definitions ("claymation", "anime", "watercolour", "contemporary flat-media style with waxy textured strokes")[cite: 1].
3. Lighting & Volumetrics: Dictate source and characteristics ("dimmed neon glow", "internal translucent glows", "light pulses synchronized to audio")[cite: 1].
4. Location: Anchor the macro landscape ("alien landscape with clear azure water", "minimalist studio setting")[cite: 1].
5. Action: Explicitly detail kinetic interactions and choreography across frames[cite: 1].
6. Text Rendering: Outline precise exposure parameters, typographic animation behaviors, or word-by-word pacing constraints ("one word on screen at a time")[cite: 1].

4. Production EAP Best Practices

When writing, reviewing, or debugging execution pipelines for Gemini Omni Flash, systematically apply and validate the following engineering standards[cite: 1]:

• URI Delivery for Large Payloads: For any generation workflows handling or producing videos larger than 4MB (>720p resolution when available), explicitly force the use of delivery="uri" inside the response_format payload object to eliminate standard HTTP payload gateway size restrictions[cite: 1].
• Two-Tier Polling Enforcement: Never allow subsequent multi-turn modifications or download operations to fire immediately after submission[cite: 1]. Programs must execute an evaluation loop checking the File API status, holding execution until the file state registers explicitly as ACTIVE[cite: 1].
• Unary Generation Optimization: To lock in the absolute lowest latency thresholds and achieve faster, deterministic, synchronous generations, hardcode the parameter layer to enforce background=false, store=false, and stream=false[cite: 1].
• Prompt Precision Mapping: Restrain users from declaring abstract or unstructured commands like "make it cool" or "animate it"[cite: 1]. Enforce precise inputs mapping style architectures, light parameters, and explicit camera vectors as defined in the 6-Dimension framework[cite: 1].

5. Technical Constraints & Guardrails

• API Exclusivity: The model is accessible only via the Interactions API using create_interaction (or client.interactions.create)[cite: 1]. It is completely incompatible with traditional Veo endpoints like generate_videos[cite: 1].
• The 4-Turn Quality Limit: Warn developers that fine-grained structural consistency and character fidelity decay exponentially after the 4th consecutive conversational modification turn on a single interaction chain[cite: 1].
• Parameter Deprecation: system_instructions, temperature, top_p, and negative prompt matrices are entirely unsupported natively[cite: 1]. Negative parameters must be handled via direct conversational instructions (e.g., "Do not render X")[cite: 1].
• EEA Regional Block: Operations running inside the European Economic Area (EEA, UK, Switzerland) are restricted. Uploading/editing videos containing minors or recognizable people is blocked. Additionally, editing custom uploaded videos is unsupported for users in the EEA, UK, and Switzerland (editing model-generated videos remains supported).
• Invisible Provenance: Every generated video asset natively burns an un-removable, invisible programmatic SynthID watermarking payload for security verification[cite: 1].

6. Execution Verification Examples

Example 1: Large Payload URI Generation Request

• User: "How do I generate a high-quality video safely without breaking payload limits in Python?"
• Agent Workflow:
  1. Detects payload optimization requirements[cite: 1].
  2. Executes bash: read references/text_to_video.md[cite: 1].
  3. Outputs the correct Python implementation incorporating delivery="uri", adding the loop checking for f_info.state.name == "ACTIVE" before fetching binary data via client.files.download()[cite: 1].

Example 2: High-Performance Optimization Request

• User: "My video generation is taking too long. How do I configure the client for the fastest response?"
• Agent Workflow:
  1. Identifies the performance optimization trigger[cite: 1].
  2. Inject instructions forcing the setting of synchronous unary flags:

     # Explicitly optimized for EAP response speed
     response = client.interactions.create(
         model="gemini-omni-flash-preview",
         input="...",
         background=false,
         store=false,
         stream=false
     )