Engineering an Agentic Production Pipeline: Orchestrating Claude Code and Gemini Omni for Consistent Multimodal Video Generation

In the current landscape of generative AI, the barrier to creating a single high-quality AI video clip has plummeted. However, a significant technical challenge remains: temporal and stylistic consistency across a series. For creators attempting to emulate high-production motion graphics—specifically the "Kurzgesult-style" aesthetic characterized by clean geometry, specific color palettes, and structured educational narratives—the difficulty lies not in generation, but in orchestration.

The problem with standard prompting is that it lacks statefulness. Every new prompt is a "cold start," leading to drift in visual identity, typography, and narrative tone. To solve this, we have engineered an automated production pipeline using Claude Code, leveraging a system of reusable cloud skills to interface with Google Gemini Omni for media generation and 11 Labs for high-fidelity text-to-speech (TTS).

The Architecture: Reusable Cloud Skills vs. Prompt Engineering

The core innovation of this pipeline is the move away from monolithic prompts toward a modular "Cloud Skill" architecture. In this context, a skill is a structured set of instructions and resources that Claude Code loads dynamically when a task matches its definition.

This system bifurcates production logic into two distinct rule-sets:

1. Content Rules: Defines the pedagogical framework, including target audience, narration tone, story structure (e.g., introduction $\rightarrow$ curiosity gap $\rightarrow$ context $\rightarrow$ analogy $\rightarrow$ takeaway), and mandatory verification steps.
2. Visual Rules: Establishes the aesthetic DNA of the channel, defining the color palette, geometric primitives, typography, iconography, and motion principles.

By encoding these rules into a reusable skill, we ensure that while the subject matter changes with every video, the production standard remains immutable.

The Pipeline: A Multi-Stage Deterministic Workflow

The pipeline operates as a directed acyclic graph (DAG), where the output of each stage serves as the structured input for the next, subject to human-in-the-loop (HITL) validation at three critical checkpoints.

Stage 1: Research Synthesis and Fact Verification

The process begins with a single high-level topic. The system does not immediately generate a script; instead, it produces a Research Package. This package includes:

• A central research question.
• Potential narrative angles.
• Key claims extracted from primary sources.
• Identified areas of uncertainty and a list of facts requiring verification.

Checkpoint 1: The human editor reviews the Research Package, cross-referencing key claims against original sources to flag any hallucinations or overstatements before proceeding to the narrative phase.

Stage 2: Narrative Engineering (The Scripting Engine)

Once research is approved, Claude transforms the evidence into a structured narration script. This isn't merely a summary; it is an engineered narrative designed for engagement. The engine utilizes specific structural primitives:

• The Curiosity Gap: Opening with a hook that creates cognitive dissonance.
• Contextual Anchoring: Setting the scene.
• Analogy Injection: Translating complex biological or physical concepts into relatable mental models.

Checkpoint 2: The script undergoes an editorial review to ensure pedagogical clarity, factual alignment, and "flow" optimization.

Stage 3: Visual Identity and Style Sheet Generation

To maintain visual continuity, the system checks for an existing Project Visual Style Sheet. If this is a new production, the pipeline utilizes GPT Image to generate a foundational style board. This sheet acts as the single source of truth for:

• Color palettes (HEX/RGB).
• Typography and font weights.
• Main subjects and background textures.
• Motion principles (e.g., ease-in/out, physics constraints).

For subsequent videos, the system reads this existing sheet, updating only the subject-specific elements while preserving the global aesthetic parameters.

Stage 4: Audio-Visual Scene Planning (AVSP)

With a script and style sheet finalized, Claude generates an Audio-Visual Scene Plan. This is a highly structured metadata table where each row represents a discrete shot/scene containing:

• Narration Segment: The specific text to be spoken.
• Visual Concept: A description of the on-screen action.
• Shot Type & Movement Tags: (e.g., "Close-up," "Pan Left," "Zoom In").
• Reference Image Prompts: Instructions for the generative model.
• Duration: Precise timing to ensure sync with the audio.

Stage 5: Multimodal Generation and Orchestration

The final stage involves executing the generation via FileAI, which acts as an interface for Google Gemini Omni. The system generates individual scene assets, ensuring each one adheres to the prompts and reference images defined in the AVSP. Simultaneously, the approved narration is sent to 11 Labs (or similar high-fidelity TTS providers) to generate the voiceover.

The pipeline does not rely on zero-shot generation. We implement an iterative loop where generated assets are checked against the visual rules and style sheet. If a scene is "off-palette" or lacks the required geometric simplicity, it is flagged for regeneration using updated parameters.

Final Assembly: Manual vs. Automated

The output of this pipeline is a highly organized directory of matched assets (scenes, voiceover, captions) with predictable naming conventions. This allows for two assembly paths:

1. Manual Assembly: Importing assets into CapCut for fine-grained control over transitions and audio leveling.

2. Automated Assembly: Utilizing HyperFrames to programmatically ingest the exported assets and construct a finished timeline, matching captions to spoken timestamps automatically.

Conclusion: The Shift from Creator to Director

This architecture fundamentally changes the role of the content creator. We are moving away from the labor-intensive task of "rebuilding the process" for every video. Instead, our job is to design the standards once and then act as the final arbiter of quality—reviewing research, approving scripts, and refining generated media. The pipeline handles the repetitive, high-entropy work of production, leaving the human to focus on high-level creative judgment.