Prompt 2 edit

The Backblaze Generative Media Hackathon brief asks for real production workflows on top of B2 and Genblaze — not just another prompt-to-video demo. We started there, then went looking for what was actually missing from every "media library" idea we sketched: a folder full of clips with tags on them isn't a library, it's a junk drawer with better labels.

The unlock came from reading Memories.ai's research, specifically their paper on prompt-driven agentic video editing. Their thesis is that understanding long-form footage — knowing what's actually in it, not just what it's called — is a separate, harder problem from generating or storing it. That pattern is built for real camera footage. Nobody had pointed it at the pile of AI-generated clips a pipeline like Genblaze produces in one evening of testing. Once you've generated forty product-video variants, you don't remember which prompt made "the one from the side angle" — you just remember what it looked like. Prompt 2 edit exists because that's a search problem, not a memory problem for a human to solve.

Real embeddings once a library outgrows what fits in one search-time prompt — same index records, just add a vector field and swap the ranking step The full recut feature — instead of returning matching clips, auto-compose them into a finished cut from one prompt, closing the loop from raw generated pile to finished deliverable Continuity memory for anyone generating a series — each new run checks what's already been established (a character's look, a brand's tone) instead of starting from a blank slate Cross-variant consistency checks — before shipping six localized renders of the same ad, confirm the same product/logo/person actually appears consistently across all six