Last month I burned through a chunk of credits on a “simple” product clip. Client said make it feel real. I opened the generator first, because that’s the fun part. Got eight almost-good takes. None of them survived review. The note was basically: cool, but what’s the point.
That one stung because the footage wasn’t bad. I just never locked what the clip had to prove.
So I flipped the order. Paper first. Model last.
Here’s what I actually do now, and it’s messier than a framework slide.
I write the brief like I’m texting a tired editor. Who’s this for, in a real moment. What is the one job. What would make me believe it if I was scrolling half-asleep. Three things we will not do. Length and ratio so nobody “surprises” me in the cut.
Then I storyboard badly on purpose. Sticky notes, or a dumb table in Notion. Hook. Mess. The fix happening. The proof you can see. Done. If I can’t draw the proof beat without jargon, I don’t generate yet. That rule alone saved me from another vanity render last Tuesday (nice lighting, zero point, client ghosted the cut for two days).
Only after that do I touch the model. One beat at a time. If proof fails, I regen that beat. I don’t throw the whole timeline in the trash like I used to.
Failed shots used to feel random. They’re usually me skipping the ugly paper step.
If you’re about to open the generator with a vibe and a prayer, close it. Scribble the proof beat first. Your future self (and your credit balance) will be less annoyed.
Related:
Free client brief worksheet: https://scienceartist.gumroad.com/l/free-client-brief-worksheet — if you want the full AI video brief + storyboard kit after: https://scienceartist.gumroad.com/l/ai_video_creator_brief_storyboard_kit
Top comments (2)
"One beat at a time" is the part I had to learn the expensive way too. Regen cost isn't linear with how much of the timeline you throw away - it's closer to quadratic once you count re-approving takes you'd already accepted the first time.
My version of your ugly paper step: before generating, I write down what has to stay identical across takes - the character's face, wardrobe, lighting direction. Those consistency constraints are the real budget killer, because a regen that drifts on any of them is worthless no matter how good the motion looks. Locking that list before the first render took my regen rate from "most takes" to "a beat here and there".
Also stealing your brief format, especially "three things we will not do". Half my failed clips weren't model failures, they were scope that never got written down - the client and I had different clips in our heads and the generator found out first.
Curious how you handle the proof beat when it involves text or UI in frame. That's where paper storyboards and generated footage disagree most in my experience: the scribble promises legible text, the model promises nothing.
For text or UI, I treat the proof beat as a compositing plan rather than a generation request. The storyboard can specify the camera/hand/device moment that establishes context, then reserve an owned UI capture or designed overlay for the exact readable state. That keeps the claim testable without asking the model to be a typesetter, and makes the consistency list easier to protect across regenerations.