Directors Cut

Inspiration

Video creators waste hours on thumbnails, text overlays, and voiceover. Most AI tools are one-size-fits-all. We wanted to build an editor where AI actively collaborates with you—adapting to your style instead of forcing generic presets.

What it does

Directors Cut is a video editor with AI agents for thumbnails, text overlays, and voiceover. Each agent proposes a strategy (e.g., bold vs. minimal thumbnails, different voice tones). You give a thumbs up or down, and the agents learn what works for you—favoring successful strategies over time. A Learning Panel shows per-agent performance so you can see how they’re improving.

How we built it

We built it with Next.js 16, React 19, Remotion for playback, and Zustand + IndexedDB for state and persistence. Each agent has its own API route that selects strategies via an exploration–exploitation loop. Gemini powers image generation, text styling, frame description, and script generation; ElevenLabs handles TTS. The frontend includes a timeline, chat panel for actions and feedback, and a right panel for clips, thumbnails, and learning stats.

Challenges we ran into

ElevenLabs free tier limits and rate blocks made live demos difficult. Managing overlapping text overlays when users said "No" and regenerated was tricky—we had to track the latest overlay IDs and remove them before adding replacements. Syncing audio overlays with timeline playback and handling IndexedDB blobs for voiceover storage required careful state design.

Accomplishments that we're proud of

Getting a working learning loop where agents reliably prefer better strategies. Building three self-improving agents (thumbnail, text, voiceover) in one unified flow. Delivering a usable editor that demonstrates the vision in under a day at the hackathon.

What we learned

Self-improvement can be implemented without retraining: a simple exploration–exploitation loop with thumbs up/down works well. Keeping agents focused on single tasks (rather than one monolithic “run everything” agent) made the learning loop clearer and more reliable. The same pattern—feedback → strategy preference → better outputs—applies beyond video editing.

What's next for Directors Cut

Adding more features for video editing exploring multi-user preference learning so agents can generalize across creators while still adapting to each user.