# PrismFlow

- **Event:** [fal x Sequoia 72-Hour Video Hackathon](https://cerebralvalley.ai/e/fal-x-sequoia-virtual-hackathon)
- **When:** Jul 17 at 12:00 PM – Jul 19 at 8:00 PM (PDT)
- **Where:** Online
- **Team:** [Anup Ghatage](https://cerebralvalley.ai/u/ghatage)
- **GitHub:** https://github.com/Ghatage/PrismFlow
- **Demo video:** https://www.youtube.com/watch?v=oFMFrujFdBQ
- **Gallery:** https://cerebralvalley.ai/e/fal-x-sequoia-virtual-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/fal-x-sequoia-virtual-hackathon/hackathon/gallery/18

PrismFlow is the "Cursor for AI Generated Video".
It gives AI agents real context and real editing power, but keeps the human in a review loop where every change is a visible, reversible, priced proposal.

AI can now generate video clips, but editing with AI is broken in two ways.
First, indie creators and solo media makers drown in toil — trimming, nudging, reordering clips, tightening pacing, hunting for "that one shot" by scrubbing through footage.
Second, the existing "AI editor" answer is a chatbot bolted next to a timeline is not enough, you need the agent to do more of the busy work for you and give you choice on what to do next.

1. Agents propose diffs, they don't mutate. You tell an editing agent "tighten the pacing in act two"; it works the timeline and produces a changeset — ghost clips rendered semi-transparently on the timeline, which you preview, accept, reject, or rebase per change, exactly like reviewing a code diff. The agent's tool calls are visible in a rail, like Cursor's.
2. The project is indexed like a codebase. Every clip carries provenance (prompt, model, seed, parent asset) plus derived context — a vision-model description of each shot and audio transcripts, embedded locally. That powers semantic search ("the shot where the fox jumps") and gives agents real context about your project, not generic prompts.
3. Clips are cached evaluations of prompts, not files. Because provenance is kept, any clip can be re-rolled, prompt-edited, or regenerated with the same prompt on a different model — with semantic search over the entire FAL model catalog to find a better model, side-by-side variant comparison, and the winner dropped back into the timeline.

Around that core: story-first creation (script beats generate scenes; the screenplay and timeline stay bound), character and style locking (cast a character once, every generation is forced through the locked references so it stays coherent across shots), and cost transparency (per-model pricing shown before you generate, half-cost draft tiers for iteration, and a live project spend dashboard).

---

Markdown version of https://cerebralvalley.ai/e/fal-x-sequoia-virtual-hackathon/hackathon/gallery/18. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt