Big Idea

The argument in one line.

Three targeted updates to a Claude prompting skill stack — volumetric depth baked by default, mid-gray character backgrounds, and behavior-based camera language — are what separate AI footage that reads cinematic from AI footage that reads generated.

Who This Is For

Read if. Skip if.

READ IF YOU ARE…

You are already running an AI video pipeline (Nano Banana Pro, Seedance, Higgsfield) and hitting a realism ceiling you cannot explain.
You use Claude to write prompts for AI image or video generation and want a structured, repeatable skill-based system.
You are building character-consistent AI content — music videos, short films, branded footage — and want to know what actually moves the quality needle.
You have seen the plastic composite look in your own work and want to understand the root cause.

SKIP IF…

You have not started with AI video tools yet — this assumes working knowledge of the pipeline and tools named.
You are looking for a tool-agnostic comparison; JOEY has already chosen his stack and teaches only within it.

TL;DR

The full version, fast.

Two Claude skills drive JOEY's entire AI cinema pipeline, and this video is the full walkthrough of their biggest update. Three changes moved the quality needle most: volumetric depth baked into every prompt by default, switching character reference builds from white seamless to mid-gray to eliminate the plastic skin-tone problem, and replacing camera brand names with behavioral descriptions the model can actually execute. He demos the complete workflow then closes with the argument that AI reliably gets you to 70% and the gap to 90 is irreducibly human editorial work.

Free for members

Chat with this breakdown — free.

Sign in and you get 23 free chat messages on us — ask for the hook, quote a framework, find the exact transcript moment, generate a markdown action plan. Bring your own key when you want unlimited.

Create a free account →

Chapters

Where the time goes.

00:00 – 00:22

01 · Intro

Hook and 10K subscriber milestone acknowledgment

00:23 – 02:06

02 · Zara/Mira interrupt

AI co-hosts tease episode content; explicit not-sponsored disclosure

02:07 – 03:36

03 · MV showcase

Clip reel of OUR TURN music video built with updated skills

03:37 – 05:11

04 · Placing AI characters in your space

How JOEY composites AI characters into his own filmed footage using Claude + Topaz

05:12 – 06:12

05 · AI vs. real cinema (depth)

Core thesis: volumetric depth is the single biggest realism unlock; atmospheric spray demo

06:13 – 07:33

06 · Using the skills — character build

Gray background vs. white; live rebuild of character sheet in banana-pro-director-2.0

07:34 – 08:38

07 · Soul Cinema vs. NBPro vs. GPT-2

Tool selection guide: cost vs. accuracy tradeoffs for character stills

08:39 – 10:54

08 · The outfit and outfit replacement

Outfit build and outfit-swap workflow baked into the updated skill

10:55 – 11:19

09 · 6-panel character sheet

Running the sheet prompt to get full-body and detail shots in one output

11:20 – 12:02

10 · Building the world plate

Tokyo Canyon scene: haze, rain texture, no rain streaks to avoid static-frame artifacts in video

12:03 – 13:07

11 · Building the scenes (Seedance 2.0)

Dropping characters + plates into Claude; prompt delivery format with image reference tags

13:08 – 13:59

12 · Why ref sheets matter (or don't)

When a reference starting image helps Seedance vs. when to skip it

14:00 – 16:03

13 · Most important takeaway

Image reference tagging (@image1-9) baked into skill; why it eliminates the biggest generation friction point

16:04 – 16:24

14 · Take it from 70 to 90

AI gets you to 70%; the last stretch requires human editorial judgment

16:25 – 16:56

15 · The edit bay

Color grade, J-cuts, sound design, pacing — things no model hands you in a code block

16:57 – 19:03

16 · Honest closing thoughts

AI as a brush not the artist; keep the 10% gap; gratitude for 10K growth

Atomic Insights

Lines worth screenshotting.

Volumetric depth — haze, particulate, atmospheric light falloff — is the single thing that separates AI footage that reads cinematic from AI footage that reads generated, and it costs nothing extra to add.
White seamless backgrounds for AI character builds are actively working against you: pure white blows edge detail and bounces light onto the face, creating the plastic look.
AI tools reliably get you to 70% quality; the gap from 70 to 90 is color grade, pacing, J-cuts, and audio — editorial work no model hands you in a code block.
Seedance accepts 9 reference images; prompts that do not tag them by position (@image1, @image2...) leave the model guessing upload order and produce inconsistent outputs.
Nano Banana Pro handles most character stills accurately and cheaply; Soul Cinema unlocks extreme styling that NBPro consistently softens.
Replace camera and lighting brand names in prompts with behavioral descriptions: 'camera operator breathing' beats 'ARRI Alexa handheld' because the model executes behavior, not brands.
A pile of beautiful AI generations on a timeline is not a film; assembly, structure, and cutting four seconds of a good generation are where craft separates the output.
The moment AI imagery becomes indistinguishable from reality it stops being entertainment — deliberately leaving a 10% human gap is both an artistic and ethical choice.
Screenshots sent to Claude iterate faster than full-resolution images for prompt generation; the quality loss is irrelevant at the prompting step.
Six-panel character sheets eliminate the need to manually composite reference views and give Seedance consistent anchors for every scene.

Takeaway

What actually kills the plastic AI look.

THE WORKFLOW

The gap between AI footage that reads cinematic and AI footage that reads generated comes down to three specific, free-to-implement changes — not a better model or more credits.

Add volumetric depth to every AI image prompt by default: haze, particulate, atmospheric light falloff. This is what makes footage feel photographed rather than rendered, and it costs nothing extra.
Build AI character reference sheets against mid-gray backgrounds, not white seamless. White blows edge detail and bounces light into the face; gray gives the model a neutral anchor for skin tones and eliminates the plastic composite look at the source.
Replace camera and lighting brand names in prompts with behavioral descriptions. 'Camera operator breathing' and 'Dutch angle' give the model executable instructions; brand names give it jargon to guess through.
Tag reference images by position (@image1 through @image9) when working with Seedance. The tool accepts up to nine references; untagged prompts leave reference ordering ambiguous and produce inconsistent outputs.
Choose AI generation tools based on styling ambition: Nano Banana Pro handles most character work accurately and cheaply; Soul Cinema unlocks extreme silhouettes and translucent fabrics that NBPro consistently softens.
AI reliably gets you to 70% quality output. The gap from 70 to 90 — color grade, pacing, J-cuts, sound design, structural editing — is earned in the edit bay and cannot be delegated to a model.
Use screenshots rather than full-resolution images when sending references to Claude for prompt generation. The quality difference is irrelevant at the prompting step, and screenshots process faster.
Six-panel character sheets (full body plus detail shots in one output) give every downstream tool consistent anchors without manual Photoshop assembly.
When building scenes with multiple characters, give Claude names for internal tracking but instruct it to describe characters by appearance in the actual prompt, not by name, so the generation model has something it can use.
Deliberately leave a gap between what AI produces and what a human editor shapes from it — the edit bay is still where the work becomes something worth watching.

Glossary

Terms worth knowing.

Volumetric depth: The visual quality created by haze, particulate matter, and atmospheric light falloff — the sense that light travels through air rather than just hitting surfaces. What makes cinema footage feel physically present rather than rendered.
Nano Banana Pro (NBPro): An AI image generation model optimized for realistic character stills. Accurate and cost-efficient for most character work; less capable with extreme styling like very baggy silhouettes or translucent fabrics.
Soul Cinema: An AI image generation model that handles extreme fashion styling better than NBPro, at higher cost per generation.
Seedance: An AI video generation tool that accepts up to 9 reference images and a text prompt to produce short video clips. Used to generate character-consistent scenes from reference sheets.
Topaz Video AI: A video upscaling tool used after Seedance generation to bring 720p outputs to final delivery resolution cleanly.
Character reference sheet: A multi-panel image showing an AI character from multiple angles with detail shots of distinctive features. Used as a reference input for Seedance to maintain character consistency across scenes.
Outfit swap: A Claude skill prompt that transfers an outfit built on a generic model onto the character body, so stylized clothing from Soul Cinema can be applied without rebuilding from scratch.
World plate: A background scene image generated separately from the characters and used as the stage into which characters are composited.
Camera operator mode: A behavioral camera description baked into the updated skill that introduces natural handheld movement — breathing, micro-shake — rather than locked-off or gimbal-smooth motion.

Resources

Things they pointed at.

01:38toolbanana-pro-director-2.0 (Claude skill) ↗

01:38toolcinema-worldbuilder-2.0 (Claude skill) ↗

07:34productNano Banana Pro

07:34productSoul Cinema

12:03productSeedance 2.0

15:00toolTopaz Video AI

02:07productHiggsfield UI

02:07productSuno

Quotables

Lines you could clip.

05:12

“Real cinema has never been about what's sharp. It's about what's not sharp.”

Punchy inversion, no setup needed, works as a standalone principle→ TikTok hook↗ Tweet quote

16:16

“A pile of beautiful generations sitting on a timeline is not a film.”

Tight, self-contained truth — lands without context→ IG reel cold open↗ Tweet quote

17:44

“It's just a brush. It's the most powerful brush we have in our tool bag right now if we can use it correctly, but it's a brush nonetheless.”

Clean closing thesis; quotable as newsletter pull-quote→ newsletter pull-quote↗ Tweet quote

04:30

“AI is only as good as the human behind it.”

Succinct, repeatable, high-agreement hook→ TikTok hook↗ Tweet quote

The Script

Word for word.

Read-along

Don't just watch it. Burn it in.

See every word as it's spoken — crank it to 2× and still catch all of it. The same dual-channel trick behind Amazon's Kindle + Audible.

metaphoranalogystory

00:01Oh my god. I'm so sorry. I

00:06think I just made the biggest update to the skills that are already helping thousands of you generate better AI video content while spending less credits. And we broke 10,000 subscribers in the span of seven days.

00:21What the 10,000.

00:24I was about. Yeah. Yeah.

00:25He's gonna share all about how to use these new skills to prompt us how we even stepped into the studio. What? Like, didn't leave the door open?

00:32Okay. Let him cook. Anyways,

00:34so earlier, I started filming this whole breakdown video on my latest video about the new updated skills, but it wasn't enough. I'm new to this, and I really wanna give more value than just a repeat of the last video that you guys absolutely blew up for no reason, and I appreciate every single one of you. So today, I'm gonna go a lot deeper That's what she said.

00:56Uh, if you're okay with that. Hi. Hi.

00:58I'm back as a fan favorite. Or did you want Mira in here? Hello.

01:02I just wanna say briefly before we get any further, I know Joey's not gonna keep most of you. I mean, you're looking at this eighteen minute video going, yeah. I don't have time for that.

01:10We won't keep you too too long, but good AI videos take some learning and we're trying to do the hard parts for you. It's still gonna be hard. Two big things that changed in the skills that are giving us way more realistic skin and cinema are going to be shared.

01:23So please, please stay here. Okay? Bye for now.

01:26Oh, that's good. If you wanna figure it all out for yourself, I've got the skills and a brief breakdown linked below, but I'm hoping you'll stay for the full workflow and the biggest update to the skills that I think no one's talking about in the AI video space. But before I say anything, this is where I'm gonna say it's sponsored by no one.

01:45It's sponsored by no one. This video is not sponsored. I had so many people thinking the last video was sponsored.

01:50Guys, if I make a sponsored video, I'll be real and straight with you. Yes. I've had so many offers.

01:54Right now, I'm doing all of this for the love of the game. Truly learning, sharing, growing together. I'll only do one if it's truly good enough and valuable for all of us, not just a quick cash grab.

02:05So let's dive in.

03:09I don't know why we decided to film all this at 4AM. They're delirious.

03:15So that's just a brief clip of the latest video, and you can watch it here later. Unfortunately, we got slapped with a restriction from a YouTube bot that I couldn't dispute.

03:27So it's not getting pushed through the algorithm, which sucks because I kinda spent a lot of time on it, um, but that's besides the point. Before I get any further, you're also probably thinking back to earlier with Zara. Was she she's not.

03:39Is she she is AI. All of them are AI. My very real girlfriend has a real job.

03:44I have a real job, but so I've got AI to help me. Before I get into the skills, I'll briefly explain how I'm doing that bit. And, honestly, this feels like a separate video, but I'm going to do it fast.

03:57And if you want a full, like, nine minute version, let me know in the comments. So real quickly, I start out with the frame of me here. I'm giving enough room for the character to fill the scene.

04:06I go ahead and grab the character with their outfit sheet, spin up a script, and send it through Claude. Claude takes the reference of me and the character so it knows what to build off of, the camera angles, all of it. I tell it to keep the frame locked.

04:19I run the script through the skills, paste the prompt, generate the scene, run it through topaz to upscale, and I place it on top of my existing footage in the timeline, mask it off if needed, and there you have it. It takes a bit of thinking and timing, and this is why I think AI will never be just as simple as a prompt output.

04:39AI is only as good as the human behind it. So the skills, which you're here for. Sorry if I kept some of the waiting tool.

04:45You saw in the last video, you know, the two skills that do the heavy lifting. You probably already downloaded them. I already know a ton of you have already been using them.

04:53I love seeing it, but those got big updates. Massive, actually. There's one change I want to start with because I genuinely think that no one really is talking about it yet.

05:03And once you see it, it's like you can't unsee it. And then we're also gonna dig into why plastic happens. Plastic, no good.

05:09We're gonna be talking about volumetric depth. Real cinema has never been about what's sharp. It's about what's not sharp.

05:16Think about the films we all rip off of lovingly. Blade Runner, Deacons doesn't light a room.

05:22He lights the air in the room. The haze, the way a face emerges out of Merck. They use a lot of times, use stuff like this.

05:33They use obviously commercial grade stuff, but atmospheric spray. I just have it because I'm also pretty filming. That allows the light to grab onto the air in space.

05:44It's grabbing on to the haze to get a lot of depth. So that's depth doing the emotional work. Even in my other videos, you can tell and it kind of just looks like the character is placed on the scene.

05:55It's tough. It's it's figuring out. We're figuring it out as we go.

05:58I we're think getting close. So that's the single biggest jump toward photograph not generated. And that's what I've seen, and it costs you nothing extra to use because it's now baked into both of our skills.

06:11That's the why. Now let's actually build something. First, before you think about a scene, you wanna lock character.

06:16Here's the first big change from the last time because I had this wrong too. I used to build characters on white seamless backgrounds. It seemed like that's what everyone was doing.

06:23It seemed right. It made sense. Clean, neutral, nothing to distract from white was working against me a lot of the time.

06:31And I'll tell you why. It's because the pure white blows out the edges of the face. It kicks light back up into the jaw.

06:37You get a lot of shininess. You get a lot of light. Or even in my older skills, I would just say diffuse lighting, soft lighting.

06:43You would still just get too much light My eyes. That's bouncing onto the character. And what that does is when you place that character into a scene, it's going to fight as hard as it can to look appropriate, and you're going to end up getting pasted on plastic looking faces.

07:00So what I've found after many generations is a solid gray background. Some of you are gonna say, bro, we already knew this. Solid gray background works wonders.

07:10So let's build a character here. I've already got a character that I started building, but I built it on a white background. So we're gonna bring this into Claude.

07:18We're gonna say, let's rebuild this character reference sheet with our new updated skills.

07:25It's gonna take that. It's gonna go through, read the banana pro director skill.

07:30It's gonna ask you if you wanna do this in banana pro. If you want to build your character in two or Soul Cinema, those are other options that it'll let you do and help you prompt.

07:41GPT two gonna give you very realistic faces. It's gonna cost more. I found that Nano Banana Pro works just as fine for all my character renders.

07:49I did end up using soul cinema for three of the null characters, the the boy group. Cool guys. Nana banana pro versus nano banana two.

07:57Two is good. It's just not as accurate as I'd like it to be in banana pro. It's gonna ask you a pre prompt checklist.

08:04I'm gonna hit run it. I don't know how to spell. Yes.

08:07But it's gonna give you the prompt. Gonna give it to you in a code format that you can just copy and paste. So we're gonna grab her.

08:14Use that as a reference image. I'm gonna paste this in. Just gonna do one batch and generate.

08:21Let's also talk about our outfit. Here, it came out really nice, very realistic, gray background, soft lighting on the face. You can tell already there's no gloss.

08:33There's no shine on the face at all. This is gonna help hugely going forward. So I've already got an outfit that I built.

08:40Built it with Nana Banana Pro. I'm gonna give you real quick brief thing on building outfits. If you want to build an outfit that is relatively simple, makes sense, not a whole lot of styling effort goes into it, I would use NanoVanana Pro.

08:55It does very well, and you can literally build it onto your character already. If you want to build a outfit that is a little bit more stylish, maybe the rain jacket is kind of cool and translucent and Nano Banana Pro isn't cutting it. If you want baggy parachute pants, trying to force Nano Banana Pro to do that, it will only give you slightly baggy pants.

09:15If you put that into Soul Cinema, it is going to give you the baggiest parachute pants you've ever seen in your life, and it might be perfect. I've built into the skill for the character sheet a outfit swap. So if you did happen to use Soul Cinema, just building an outfit on a generic model that isn't quite your character, doesn't look like it, I've built something to help us put that outfit onto the character.

09:39It's a super easy prompt. We're gonna grab both of the reference images. We're gonna take screenshots of both.

09:45I use screenshots instead of full images because it's a lot easier and faster to send it to Claude. So we're going to make an outfit swap.

09:54We're gonna drop the two into Claude for reference, and it's going to give us this very simple it's gonna tell you which image should be first. The image should always be the outfit first and then the character second.

10:08I don't have reference tags in all of the prompts, but just for the sake of outfit changes, it's kind of important just to do it here. Boom. Looks great.

10:17Exactly how we want. Now, if there are specific markings or things on your character that are important.

10:25For example, any like tattoos or piercings or markings or anything of that sort that is on your original character. Make sure that some of those are in the description or in the prompt. I don't have that built into this skill because it's this is a very generic outfit replacement swap.

10:39Just add some some extra text like make sure that the tattoos and piercings remain the same from at image two. There's that. I wanted to change her hairstyle just for the sake of this video.

10:50So we went ahead and built her a bun, messy bun. And now we just need to drop that image into Claude and say, run the six panel sheet. So we get a six panel character sheet that shows us the what she looks like full body, some of the detail shots.

11:05She has nail colors or necklaces, jewelry, rings, things like that.

11:11It's going to give us a really nice character sheet. Already here, we don't have to go and photoshop or do whatever, make make them all in one prompt.

11:19While I've been doing that, I've been trying to build this scene, um, so that we can have a scene to put them in. I went ahead and used a scene reference from our last video, but I'm just bringing it in here.

11:31I'm gonna build the f 40 here on the background of this Tokyo Canyon. Got the haze. We've got the rain.

11:38I'm not using rain streaks in my image because sometimes if I'm using an image that has rain streaks, sometimes in seed dance generations, you'll still see the the rain streaks kind of still not really moving. So what I did was I did no rain streaks, but I want rain texture on the ground, raindrops on the roof, things like that.

11:59These are the small things that we don't really think about, but we should. So now I'm building the scenes. Basically, I'm just gonna do a quick scene for you guys just to show the skills at work.

12:08We're gonna take Zara here with an umbrella and a cheesy t shirt that says mama didn't raise no AI slap. We're gonna use this new character.

12:17I've told Claude the names of these characters just so for when we're talking to Claude, it helps us keep on track of who's who, but it's also going to describe in the prompt who the characters are and not using the names, which is important. We then grab screenshots of each one. We've got the plate with the car.

12:33I built this plate with her putting on lip gloss in the car. It's gonna make sense later. We're gonna drop all those into Claude.

12:39We're gonna start building our scene. The first scene here is going to be a banter inside the car of between Zara and May.

12:48May is going be applying some lip gloss. Zara is going to grab her jacket and say, that's a lot of jacket because it is a big jacket. And then she's gonna say, Lynn, where's yours?

12:57So just a little bit of dialogue just to show the tool at use. Gonna show how it doesn't look plasticky, how we've got really nice camera features. It doesn't look like they're just pasted into the scene.

13:07I did make this reference image of her inside the card just so we have a little bit of reference to go off of. But when we're doing all of our seed dance generations, it's not going to specifically use that as the starting image.

13:20That's important. It's important to have for reference for seed dance to take and know what it looks like inside or outside, but you don't have to use it as a starting reference. I've just found sometimes that helps, sometimes not necessary.

13:33But when you're wanting specific people in specific locations, it's helpful for c dance to know. So we're gonna drop that in there.

13:39We're gonna tell it what want. It's gonna give us this prompt important information when it's giving you the prompt delivery.

13:46This is the biggest thing that I didn't have in my last set of skills that I was fighting against multiple times. It is something that people are already aware of. There was a bit of, does it really work?

13:58Does it not? I found building it into my skill to make it way easier to prompt saves so much time.

14:05So it's kind of important the way that we do this. Once you upload those reference sheets into Clot and you tell it the scene that you wanna make, it's going to use those reference sheets. It's going to spit out the prompt, but it is going to give you the reference sheets in order of that which you upload to c dance.

14:21One, two, three, four, all the way up to nine. You get nine reference images to upload. This is important because in the prompt, it's going to tag each one as at image one, image two, image three, and so on.

14:33It's also gonna tell you the title of the scene, and it's going to give you a runtime. We're gonna paste this into seed dance. We're gonna set to seven twenty p.

14:41I use seven twenty because I use Topaz to upscale really clean. I can do upscale video too if you guys want. There's a few buttons that I press that are pretty important.

14:50Another thing that makes these videos really real is the camera movement. We had some good stuff in the old skills. In these skills, I updated some of the modes to have camera operator.

15:02A camera operator, if we're not using a gimbal or a Steadicam, something that keeps the camera really still, the person holding the camera is breathing. Right? Or he's about to die.

15:11So there is going to be some operator movement. There are Dutch angles that give some really cool scenes. These are the things that I've baked into the skills that just help give some more cinematic feel and vibe.

15:24There we go. I have two that are rendering out and I'm gonna show you what those look like right now.

15:34That's a lot of jacket. And where's yours?

15:40Come on. You you couldn't have done that in the car. I'm freezing.

15:46Are you rushing me? Okay. Sorry.

15:48Hurry up, please. So that is what those look like. Not the best.

15:52I mean, I'm just trying to build a quick little video here for you. Those came from, like, four different generations, I think. And those came from editing those generations, which brings me to the part that is kind of important.

16:02You do all this. You build the characters. You lock the outfits.

16:06You build the scenes. You get the real depth. You get your prompts clean, your references are in order, you generate everything, upscale it, drop it onto the timeline, and you're at 70%.

16:15A pile of beautiful generations sitting on a timeline is not a film, and the last 20, part that takes you from 70 to 90, happens in the edit bay and it's all the stuff that AI can't do. It's the color grade.

16:27It's what every shot feels like it came off the same roll of film. It's the j cuts and the audio. It's the sound effects.

16:34It's the sound design. It's pacing. It's structure.

16:37It's cutting out four seconds of a really good generation because those four seconds just don't tell the story. These are all the things that are you. It's your craft that no model is going to hand you in a code block.

16:49Anyone can get to 70. That's the whole point of these tools. They can get you to 70 pretty fast.

16:54The 90 is still earned. Why not a 100? I don't think we should get to a 100.

16:59There should always be a gap with AI. A little scene where you can tell that it's human made. The moment AI worlds become so indistinguishable from reality, we've stopped making entertainment and started making something that I don't really wanna be a part of.

17:14Doing deep fakes, fake news, we've seen those before. That those kind of things, that hurts people instead of moving them emotionally. It might move them emotionally to anger.

17:24And that's that's the line for me. What I'm doing here, what I hope you're doing is building worlds.

17:31I hope that you're building them for entertainment, for the love of it. A story you wanna tell that you don't have a 100 person crew to tell it. That's not taking anything away from real movies, real artists, real entertainment.

17:44It's just a brush. It's the most powerful brush we have in our tool bag right now if we can use it correctly, but it's a brush nonetheless. So I leave you with the skills to go paint.

17:54Go do something. Go make something. But keep the scene.

17:56Keep the realism in there. Get to 90 and be proud of the 10 you left on the table. And that's the workflow.

18:02That's the update. That's me. I've got a lot to dive into still, and those videos will come a lot faster than this one.

18:08I just wanted to give you these skills quickly and in great detail as much as I could as soon as you mean the possible. Again, I don't understand how we went from 40 literally, like, ten days ago to 10,000 of you wanting to watch this content.

18:23If you think about how much 10,000 is online, 10,000 might not seem like much. To me, it's a lot.

18:29But to some, it's it doesn't seem like much. If you picture 10,000 people, you can't fit in this entire radius.

18:36So that's kinda crazy, and I'm really not normal about it. Thank you guys seriously from the bottom of my heart. I'm excited to keep giving you guys content, keep helping, keep doing the thing.

18:45For the love of the game, we're here and we're not going anywhere. Peace, love, and AI.

The Hook

The bait, then the rug-pull.

The apology lands before the setup does. JOEY opens mid-thought — breathless, slightly sheepish — as if he has been awake too long working on something that got bigger than expected. Then comes the actual promise: a skills update that thousands of people are already using, and three changes inside it that no one in the AI video space is talking about yet.

Frameworks

Named ideas worth stealing.

16:04concept

The 70/90 Rule

AI tools reliably get any creator to 70% quality output. The gap from 70 to 90 — color grade, pacing, audio design, structural editing — is earned in the edit bay and cannot be automated.

Steal forAny conversation about AI-assisted creative work and where human craft still lives

05:12concept

Volumetric Depth Baking

Add haze, particulate, atmospheric falloff, and light shafts to every AI image/video prompt by default. Cinematographers light the air in a room, not just the subjects.

Steal forAI image and video prompting, cinematography education

06:13concept

Gray Background Character Builds

Build AI character reference sheets against mid-gray backgrounds instead of white seamless. White blows face edges and bounces light, creating the plastic composite look.

Steal forAI character consistency workflows

15:00concept

Behavior-Based Camera Language

Remove brand names (ARRI, Panavision, Aputure) from AI prompts and replace with behavioral descriptions. Models execute behavior, not brand jargon.

Steal forAI video prompting, any prompt engineering context

14:00model

Image Reference Tagging (1-9)

Seedance accepts up to 9 reference images. The skill outputs prompts with explicit @image1 through @image9 tags matched to upload order. Eliminates the most common point of generation failure.

Steal forAny multi-reference AI video generation workflow

CTA Breakdown

How they asked for the click.

VERBAL ASK

01:38link

“Download updated skills here: tinyurl.com/claude-skills-2-0”

Mentioned early and repeatedly, description link only — no hard sell, no course pitch. The repeated not-sponsored disclosure functions as trust-building that makes the CTA more effective by contrast.

MENTIONED ON CAMERA

01:38toolbanana-pro-director-2.0 (Claude skill) ↗

01:38toolcinema-worldbuilder-2.0 (Claude skill) ↗

FROM THE DESCRIPTION

PRIMARY CTAWhere the creator wants you to go next.