Automating Product Demo Videos with Claude Code and GPT-6 Astra

Orchestrating AI Video Production

The author uses a specialized workflow to produce product demo videos by decoupling creative direction from execution. By running GPT-6 Astra inside Claude Code, the agent acts as a director that plans camera movements and shot sequences, while Claude Code manages the local environment to capture and edit the footage. This setup avoids the limitations of chat-based video generators by allowing the agent to interact directly with a live application in a browser.

Implementation and Tooling

To avoid high API costs, the author uses a CLI proxy to route Claude Code requests through an existing subscription. The process relies on specific browser-automation tools and a custom-built skill for music synchronization.

• Environment Setup: Use CLI Proxy API to connect Claude Code to an existing coding agent subscription. Set the BASE_URL to the local proxy server and provide the API key to initialize GPT-6 Astra within the Claude Code interface.
• Planning and References: Provide the agent with the URL of the running application and reference videos (e.g., Anthropic or Linear launch videos). The agent renders a visual plan in the browser with a comment box for iterative refinement before locking the direction.
• Modular Recording: Record shots individually at full-screen resolution. This allows for immediate review and re-generation of specific clips if camera movement or zoom levels deviate from the plan, preventing the need to re-render the entire sequence.
• Music Synchronization: Create a custom skill that analyzes the video's cut frequency and product style. The agent uses this data to fetch tracks from Mixkit that align with the beat, ensuring the audio pacing matches the visual transitions.

Workflow Refinement

The author emphasizes that the quality of AI-generated video depends on the agent's ability to perform iterative checks. By running built-in validation scripts after combining clips, the agent identifies alignment issues or timing errors before the final render. This modular approach ensures that the agent remains in control of the tools, rather than relying on a black-box video generation model.