By Novae Caelum · Originally published
And he only balked a few times
And it only took around 500 hours of my own expertise and another bazillion hours of app and workflow building, testing, reworking, and building again to get there.
mic drop
Clickbaity headline aside (and, truly I HATE the “AI just did X thing!” headlines, because no, AI executed what a HUMAN set up in their own judgement, with time, experience, and taste), Claude really did make a short film for me this weekend, in my app. Which again, I’ve been building for over a year. It’s judgement, process, and taste all the way down.
The culmination of all of that is: I gave Fable a short story I wrote a few years ago about a headless telepath just trying to find work. He ripped it into the app and ran through the ingestion process, ran through the character and location and assets references, ran through the screenplay parsing—double checked the screenplay parsing to make sure all was well, and then proceeded to roll on into the video rolling sequences.
We had a few false starts. I was still working out the particulars of how to MCP my app—we’d do things like:
Me: Okay dude, go make the thing, woot!
Fable: Oh yeah, forgot to say, we didn’t actually wire up this part yet, my bad! Want me to do that for you?
Me: WHY YES DUDE GO DO THAT.
We had to work through things like the app accidentally casting everyone as iterations of the same East Asian woman, because that woman was prominent in the style sheet, where I had to tell Fable to go diversify the cast. And he did and could, ripping through the prompts, making those decisions.
I had to veto his first decision for the Surgical Bot character, too, because it was basically Lt. Commander Data. We had to reroll that one.
And, we carefully ramped into the full video sequence for each scene, until he got his legs and then he happily queued up as many scenes as he could in parallel (because he needs to run concurrent things one at a time, each dependent on what comes before). But scenes not sharing the same location/time are fine to rip through simultaneously.
He knew to do that on his own. He also found several errors and rerolled them on his own, with a hard gate of no more than three rerolls on a video before pinging a human to see what’s up.
After Fable had run through several scenes, I handed that off to Opus 5 and he did just fine, too.
When everything was all rolled, around 6 hours of Opus chugging working through it later, we’d made around 10 minutes of raw footage, which I then cut down to 5 in sharp, social media pacing.
It’s a little hard to gague my initial time in planning with Opus and Fable, and QA-ing their decisions, because I was working through some bugs and fixes along the way. But, I think it’ll take ~1/2 hr to an hour total time—unless you’re really digging into the screenplay, which you might want to do (at very least have the AI make sure it all ads up to the beats in your book before you start laying down serious cash with videos). But, I gave it a glance, and chugged on.
My own cutting at the end took around an hour and a half. It did over-generate. But quite frankly, I’d rather it over-generate than under-generate. I was an editor working with wonderful, surprising, and on-point coverage already laid out for me. I could pick and choose some of the shots and moments to include, discarding the rest, and I think it’s better for it.
I spent another hour with some fixes - mainly because I threw it a doozy of a story to work with (said headless telepath, which fought us along the way and showed me just how well Fable and Opus can QA and reroll with judgment calls) and some of the audio went off because of it. But, I basically rerolled 2-3 scenes, grabbed a snippet of the voice in another and rolled a few extra lines in Seed Audio in Fal to make the voice consistent. A character with a head (who isn’t also nonbinary, heh) wouldn’t have that problem. Mostly everything else stayed consistent.
I made a single soundtrack song in Suno and just reused it on the seem, fading it in/out of itself. I edited in Capcut, which is my go-to when I’m making anything social-facing, vs Davinci when I’m rolling my main horizontal show. Capcut will lay down the trendy-looking captions. Plus, it’s just fast and no-nonsense.
Lastly, I’m waiting on the Topaz Starlight 2.6 (a model I’ve never used before, so, fingers crossed) to upscale the 480p we rolled in to 1080p and we’ll be ready to go. At five minutes of film, this is gonna take a while. (I’ll circle back.)
My process isn’t revolutionary. I mean, it IS revolutionary. Because it’s hella new, and hacking at what’s possible all the way. But, most AI filmmakers are doing similar things.
But, I needed a way to speed this all up, because I’m a one person empire with an amazing assistant who needs to speedrun through IP like whoa so I can rule the world. As one does. So, we’re solving these problems.
This film cost me ~ $110, all said and done. Which is more than I’d hoped, but, in the grand scheme of things, quite ridiculously cheap. And it only, in the end, cost around 3 hours of my own time. FOR A FIVE MINUTE FILM.
I stand by it being possible to make a 60 minute film for right around $1000 USD, give or take a wobble. It would probably take Claude or Codex or Grok or whoever running a solid week, so do add that in, too. Maybe a tad more than $1000. If you’re running with cheaper models like Minimax H3 in the gaps, a tad less. Or, even the whole way. Running H3 in 2K will cost a tad more on the front end but you won’t have to upscale on the back end, and good upscaling can be pricey.
I had Fable cook up a rough pricing estimate using several models with this system - TAKE THIS WITH A GRAIN OF SALT. Every project will have its own scope. Model pricing can and will change week to week sometimes, though for the most part, they don’t tend to get more expensive (at least, within the same model) over time, just more expensive/better models coming out.
If you’re rolling at 480p you WILL need generative Topaz upscaling. If you’re rolling 720p, you can get away with some of the local upscaling out there, though it’s not as sharp.
This doesn’t count the cost of GPT Image 2, as well, which will run into high double digits at least across a large project. But! It’s one of the big drivers of why this is all possible (though I’m gonna be testing Seedream 5 Pro in the same role shortly, too).
It’s not that people are at a lack of story these days, but people are at a lack of uniqueness. Of personality. Of niches Hollywood won’t touch and representation it’s punting out again.
Production quality is important, but so are feels. And I think (as I’ve punted on watching SO MANY shows and movies this year as same old vanilla flavors) that a scrappy author with a back catalog and a scrappy strategy has much, so very much to gain in this space.
Because it me. I am doing this, so is Cassie, so are a lot of you, or upcoming!
My dream is to have an ongoing queue of book projects lined up to be turned into film projects—the agent pulls from my command center on schedule, fires up the project, and pings me only when it needs decisions and when it’s done. I edit, which is my very favorite part, and then make glorious, glorious shows and movies in all directions.
THAT IS THE DREAM. THE DREAM IS BECOMING REALITY.
Anyhow! This version of the app with MCP + agent skills will go to our paid folx later this week, I think (I very much plan!). And one of my goals with this project is for it to always be in motion—new pipelines will be built shortly after new models and workflows come out; old ones will retire as they phase out.
You can use the app on its own, but, I think the true power really lies in the MCP driving it now, with the app as the process and interface. The agent is now the decision layer first line, the troubleshooter, and the pusher-alonger. And that was a whole problem I’d needed to solve, because you do need eyes on things in between the bigger decisions, and the agent can be those eyes.
I’m releasing to public beta next month, so my cutoff is end of August for sub folx to have ongoing access to this at just the Substack cost. You all are my early testers, and you helped me figure out some of the problems that still needed to be solved. And now, they are. (Some of them, anyway. Enough of them!)
Starting Friday, too, I’ll be taking over our 10AM PST Friday calls for a few weeks to dive into an AI filmmaking mastermind. This is gonna be as comprehensive as I can make it, arming you to have all the know-how and skills levers to know how to make things on your own, build your own systems, or how to run the app + agent and make the raw output into something polished and solid.
This, like anything with AI, is going to be something that can be done automatically, but to rise above the flood, and that flood is definitely coming (judging by the numbers of AI verticals coming out of China right now, which are huge) you’ll do much better knowing why something is working or not, and how to fix it. Or, even how to roll a sex scene on your own that all the regular methods refuse, etc etc.
It’s just a different medium of storytelling. If you’ve written stories, made decisions on what to put on your covers, and driven your own social media images and videos—you can do this. Full stop.
And the final is ready! Woot! (after 2 1/2 hours of upscaling, lolsob, so yeah that’s gonna be a time factor too—but at least you can walk away from it and set Opus or Sol to check in on it every half hour or so)
You all are gonna get to see alll the weird ass stories I wrote the last decade or so with all my experiments while testing everything. I decided to be smart and release a short story collection with all of these in a few weeks because if I’m gonna be dropping cash to make these shorts, I am by god gonna promote shit! 🙌 I’ll post this on my IG tomorrow.
Is it perfect? Nope. Could I spend another few hours to fix it? Yep. Am I going to? Nope. That’s the point. Speed and volume. Make people feel. Roll with it. MOST of this is first-roll. The app works on vibe-pushing-through-it.
It’s been a wild and bumpy and crazy ride to get this far—I’ve dropped at least $4K on this app project between now and last year when I started (and when both I and the tech weren’t ready for it yet).
But we’re here. We really are here. Give me a few days, then you all can go play with it. <3
(And make sure to join us Friday to start the masterclass!! I’ll be going deep into all the nuts and bolts of how and why, all the craft I’ve learned, all the latest methods, all of it. It’s gonna be a blast!)
-Novae