<aside> 💡
Watch the full video here: https://youtu.be/4TXaAnittHs
</aside>
<aside> 💡
Download the Claude Skills here
</aside>
If you watched the video, you already saw what these skills do. This is the post for the people who want to actually use them. No course. No upsell. The skills to download are above, and the prompts are below — take them.
I'm writing it because the version of me who started CTRL: Hunters would have killed for this doc. I’ve burned literally thousands of credits dialing in these skills so you don't need to. They’re not perfect but I feel we’re getting close.
Two updated Claude skills and the sample prompts I'm running every day:
banana-pro-director-2.0 — every still. Character faces, outfits, scene plates, reference sheets, detail shots.cinema-worldbuilder-2.0 — every Seedance video. Cinematography grammar locked, audio rules locked, character markers locked.The skills got smarter in three specific places, and these are the ones that actually moved the needle on output quality.
Volumetric depth, every frame. The thing that makes AI footage read as AI is flat air. Real cinematography has atmosphere — light cutting through haze, fog catching practicals, particulate in the foreground softening edges in the deep background. The skill now bakes volumetric language into every prompt by default. Haze density, particulate in air, light shafts, atmospheric falloff between subject and background. You don't have to remember to write it. It's there.
Grey backgrounds for character builds. This one's a small change with big consequences. When you build a character on pure white seamless, the model has nothing to anchor skin against — everything gets that plastic, over-lit, AI-portrait look (most of the time). Mid-grey seamless gives the model a value to read skin tones against, lets the shadows do real work, and the character comes out with actual dimensionality instead of looking like a render. I switched every new character build to mid-grey by default and the plastic problem mostly went away.
Cameras described by behavior, not brand names. The old prompts in the pipeline ended with "ARRI Alexa 35, Panavision Ultra Vintage 75mm anamorphic at T2.3, Kodak Vision3 250D pushed 800 ASA." It worked, but the model isn't actually rendering an Alexa — it's pattern-matching on the vibe of that gear stack. The updated skill cuts the brand names entirely and writes the behavior instead: "wide-latitude cinema capture, vintage 75mm 2x anamorphic character at a wide aperture, oval bokeh, color-negative film rendition with fine 35mm grain." Same look, none of the noise. Same logic on lighting — direction, quality, temperature in plain physical terms. No fixtures, no model numbers, no jargon the model has to translate.
A few things that aren't on the marketing page for any AI tool, and I'm going to say them out loud because if you don't know them going in you're going to waste credits.