20-30 mins and 10-20 dollars. It's fully automated evals with Gemini 3 Flash as a judge, but I verified manually a lot of them and it grades the outputs reliably.
Really impressive work guys! It seems like YC has funded a few companies attacking this but I think you all might have the best approach so far. Behind the scenes is the agent just editing using text/annotated timelines? I feel like the move is probably text for roughcut/narrative, then a vlm for digesting the initial roughcut, then adding broll and fixing timing issues. Feel free to steal my FCP xml generator. https://github.com/barefootford/buttercut
happy that you liked our approach! also, i think it's a better idea to just give agent these tools and let it figure out its course of actions than giving it a specific workflow to work on - it seems like the world keeps reminding us the bitter lesson [http://www.incompleteideas.net/IncIdeas/BitterLesson.html] more frequently these days
Theoretically I agree, but practically without guidance agents aren't really able to edit video ATM. Without hand holding Claude will just call ffmpeg and look at a few frames.
Sorta wish they would ship the last watches before announcing more that are months away. I’m pretty sure I preordered 9 months ago and still have nothing to show for it.
I switched from structured outputs on OpenAI apis to unstructured on Claude (haiku 4.5) and haven't had any issues (yet). But guarantees are always nice.
I've wanted to build a video editing agent this year but I kept putting it off because I thought it would take too long. Having simple Claude Skills as markdowns + a little bit of Ruby made this a lot easier than I expected.
Behind the scenes we transcribe audio, analyze frames, and then build a combined/combed transcript of all the footage. Claude can then build rough cuts by just laying out clips with the correct timestamps. After a yaml rough cut is built, a Ruby library generates basic XML for Final Cut or Premiere.
The design taste on almost all these humanoid bots seems off. If I'm going to have some AI-powered alien in my house I want it to be and appear harmless, helpful, maybe even cute? These look creepy and dystopian.
I look forward to trying this out. Any benchmarks or demos on how long it actually takes to restore? I ended up cooking my own boring S3 backup because previously litestream took 20 minutes to restore something like 1000 rows. It felt extremely unoptimized. How long does restoration take today?
Hey I’ve been trying to get Claude code to generate Final Cut xml myself. Mostly I just have a Claude.md with the FCP xml reference and some guidelines. What does this do differently?
Well, the format is incredibly complex. You have to generate the XML correct in every last detail or FCP will crash when you import it. I learned the hard way you need a robust validator system and not just tests alone. Before XML goes out the door you need a last line of defense to catch problems. The go version of cutlass is pretty good at this now and python is catching up.
