Claude Code picks up a camera

What it does This toolkit wraps Claude Code in a full video production workflow. It gives the agent structured skills for Remotion compositing, Playwright screen recording, FFmpeg processing, and open-source AI models for voice (Qwen3-TTS), images (FLUX.2), and music (ACE-Step). You drive the process through slash commands like /video and /setup, using templates for sprint reviews, product demos, or custom explainer videos.

The interesting bit The real machinery is a project management system that treats video production like a multi-session software project. It tracks scenes, assets, and audio across separate Claude Code sessions, automatically reconciles your original plan against what actually exists on disk, and generates a per-project CLAUDE.md so you can resume editing days later without losing context.

Key highlights

• Built-in brand profiles (brand.json, voice.json, assets) that auto-apply colors, fonts, and voice settings to every project
• A lifecycle tracker that moves videos through phases: planning → assets → review → audio → editing → rendering → complete
• Custom scene transitions (glitch, RGB split, light leak, clock wipe) alongside standard Remotion effects
• Experimental migration script to port the Claude Code skills to OpenAI’s Codex runtime
• Heavy reliance on self-hosted open-source models and cloud GPU (Modal or RunPod) rather than locked-in APIs

Caveats

• The Codex bridge is explicitly marked experimental, and modifying skills requires restarting Claude Code to pick up changes
• The author admits the original use case—sprint review videos for Digital Samba—is fairly specific, and “autonomous video creation is a lofty ideal for such a subjective field”
• Cloud GPU setup is required for the open-source AI tooling, though Modal’s free tier and Cloudflare R2’s 10GB tier keep initial costs near zero

Verdict Worth exploring if you already live in Claude Code and need to produce scripted demos, walkthroughs, or sprint reviews. Those who prefer direct, hands-on control over timelines and keyframes will likely find the agent-mediated workflow indirect.