WEBVTT 00:00:00.000 --> 00:00:03.140 You know that feeling. It's like 11 .0 p .m. 00:00:03.160 --> 00:00:05.120 You've got this amazing movie scene in your head. 00:00:05.160 --> 00:00:07.839 It's so vivid. You see the light, the mood, the 00:00:07.839 --> 00:00:11.980 actor. Everything. So you sit down, you open 00:00:11.980 --> 00:00:14.679 your video generator, you type in a prompt, you 00:00:14.679 --> 00:00:17.980 hit enter, you wait that agonizing minute, and 00:00:17.980 --> 00:00:20.519 then... That absolute garbage. It's just heartbreaking. 00:00:21.160 --> 00:00:23.679 The lighting is completely off. The character 00:00:23.679 --> 00:00:26.260 has morphed into a different person. Or, and 00:00:26.260 --> 00:00:28.280 this is my personal favorite, they suddenly have 00:00:28.280 --> 00:00:30.879 six fingers. So what do you do? You tweak a word. 00:00:30.940 --> 00:00:32.759 You pull the lever on the slot machine again. 00:00:32.920 --> 00:00:36.420 Exactly. You burn another 50 credits. It's what 00:00:36.420 --> 00:00:38.119 we've all come to know as the prompt lottery. 00:00:38.380 --> 00:00:41.020 It really is the universal frustration of this 00:00:41.020 --> 00:00:43.880 whole early AI era. You feel like you're gambling, 00:00:44.000 --> 00:00:47.439 not creating. Just crossing your fingers and 00:00:47.439 --> 00:00:49.380 hoping the machine spits out something usable. 00:00:49.659 --> 00:01:01.899 What if that whole... And it's such a fascinating 00:01:01.899 --> 00:01:04.680 read because the whole argument is that we've 00:01:04.680 --> 00:01:07.519 been doing it wrong. We treat AI like a slot 00:01:07.519 --> 00:01:09.780 machine hoping for a jackpot when we should be 00:01:09.780 --> 00:01:12.060 treating it like a film crew. It's a shift from 00:01:12.060 --> 00:01:15.560 just, you know, writing stories to actively managing 00:01:15.560 --> 00:01:19.219 visual continuity. So welcome to the deep dive. 00:01:19.400 --> 00:01:21.319 We're going to unpack this whole murder board 00:01:21.319 --> 00:01:24.019 method step by step. And I got to say, the name 00:01:24.019 --> 00:01:28.079 murder board, it's a little intense. It is, right. 00:01:28.260 --> 00:01:31.250 But the metaphor, once you get it. is actually 00:01:31.250 --> 00:01:33.629 spot on. Yeah. Think about any detective show 00:01:33.629 --> 00:01:35.590 you've ever seen. You've got that big cork board 00:01:35.590 --> 00:01:38.890 on the wall, photos, timelines, pieces of evidence. 00:01:39.030 --> 00:01:40.709 And the red string connecting everything. The 00:01:40.709 --> 00:01:43.090 classic whodunit map. Exactly. No, just apply 00:01:43.090 --> 00:01:45.989 that same logic to AI video. Max Anne's point 00:01:45.989 --> 00:01:48.670 is that in 2026, you can't just wing it with 00:01:48.670 --> 00:01:51.230 prompts. You need a visual control system. You 00:01:51.230 --> 00:01:53.709 have to track every single seed, every reference, 00:01:53.829 --> 00:01:56.189 every prompt. So the murder board is your project 00:01:56.189 --> 00:01:58.780 management layer. It's your continuity engine, 00:01:58.939 --> 00:02:01.819 because without it, the AI just hallucinates 00:02:01.819 --> 00:02:04.620 a new reality every single time you hit generate. 00:02:04.799 --> 00:02:07.420 It has no memory of what it did five minutes 00:02:07.420 --> 00:02:10.419 ago unless you force it to remember. OK, so we're 00:02:10.419 --> 00:02:13.759 shifting from being writers to being. What? Project 00:02:13.759 --> 00:02:16.780 managers. Visual directors. Precisely. And to 00:02:16.780 --> 00:02:20.280 do that, Anne lays out a very specific tech stack 00:02:20.280 --> 00:02:23.219 for 2026. It's a pipeline. All right. So let's 00:02:23.219 --> 00:02:24.939 map this out for everyone listening. First, we're 00:02:24.939 --> 00:02:27.039 going to cover the ingredients. Then the brain, 00:02:27.300 --> 00:02:29.860 which is the LLM. Then the visual base, which 00:02:29.860 --> 00:02:33.259 he calls Nano Banana Pro. A name I still cannot 00:02:33.259 --> 00:02:35.300 say with a straight face. It's sticky. You got 00:02:35.300 --> 00:02:36.939 to give them that. Then we move to the motion 00:02:36.939 --> 00:02:39.840 engine, Kling 2 .6. And finally, the part that 00:02:39.840 --> 00:02:42.319 I think is the biggest game changer. The secret 00:02:42.319 --> 00:02:45.860 weapon. Ah, shut deck. That's where the real 00:02:45.860 --> 00:02:48.560 magic happens, but we have to earn it. Fair enough. 00:02:48.860 --> 00:02:51.259 Let's start with a core idea. The guide says 00:02:51.259 --> 00:02:53.099 the prompt lottery is dead. We've heard that 00:02:53.099 --> 00:02:55.819 before. Why is this different? Because it changes 00:02:55.819 --> 00:02:58.280 the entire order of operations. See, most people 00:02:58.280 --> 00:03:00.659 think the workflow is write a prompt, then generate 00:03:00.659 --> 00:03:03.360 a video and says that's just a recipe for failure. 00:03:03.740 --> 00:03:06.580 The new workflow is a hierarchy. It's a sigh. 00:03:07.229 --> 00:03:10.250 Gather assets, then generate still frames, and 00:03:10.250 --> 00:03:12.490 only then do you even think about touching the 00:03:12.490 --> 00:03:15.409 video button. So you're never asking the video 00:03:15.409 --> 00:03:18.009 AI to actually compose the shot for you? Never. 00:03:18.069 --> 00:03:21.129 The video engine is just for movement. The composition, 00:03:21.469 --> 00:03:23.409 the artistry, that all happens way upstream. 00:03:23.569 --> 00:03:25.750 You've got your strategy layer, like a chat TPT 00:03:25.750 --> 00:03:28.229 5 .2, acting as your director of photography. 00:03:28.509 --> 00:03:31.449 Then Nano Banana Pro creates the high -res stills, 00:03:31.449 --> 00:03:35.659 and Kling 2 .6 just... It just animates the thing 00:03:35.659 --> 00:03:37.639 you already built. It's assembly line logic. 00:03:37.879 --> 00:03:40.419 Let's back up to that first step. Gather assets. 00:03:40.860 --> 00:03:43.400 The guide calls them base ingredients. And this 00:03:43.400 --> 00:03:45.560 part really surprised me. He says before you 00:03:45.560 --> 00:03:47.939 write a single word, you need boring images. 00:03:48.180 --> 00:03:50.340 Boring is the key word. You start with ass in 00:03:50.340 --> 00:03:52.810 A. the character reference. He uses that example 00:03:52.810 --> 00:03:54.949 of Captain Renfield. Right. And the advice is 00:03:54.949 --> 00:03:57.729 super specific. A clear, well -lit portrait facing 00:03:57.729 --> 00:03:59.909 forward with a totally neutral expression. No 00:03:59.909 --> 00:04:01.810 dramatic lighting, no weird angles, nothing. 00:04:02.050 --> 00:04:05.949 But wait, why neutral? If I want to make a dramatic 00:04:05.949 --> 00:04:08.509 movie, why am I starting with what is essentially 00:04:08.509 --> 00:04:13.250 a passport photo? It's because of how the AI 00:04:13.250 --> 00:04:17.050 maps geometry. If you feed it a reference where 00:04:17.050 --> 00:04:19.870 the character is, say, screaming or in heavy 00:04:19.870 --> 00:04:23.550 shadow, the AI bakes that emotion into the character's 00:04:23.550 --> 00:04:26.310 identity. Ah, so then every shot you generate, 00:04:26.389 --> 00:04:27.870 they look like they're screaming. You got it. 00:04:27.889 --> 00:04:31.149 You need a clean map of the face so the AI can 00:04:31.149 --> 00:04:33.709 apply emotions later without... distorting the 00:04:33.709 --> 00:04:35.889 actual bone structure underneath. Okay, that 00:04:35.889 --> 00:04:37.730 makes so much sense. It's like a texture map 00:04:37.730 --> 00:04:40.009 in a video game. You want the base layer to be 00:04:40.009 --> 00:04:42.170 totally flat so you can paint light and shadow 00:04:42.170 --> 00:04:44.670 onto it later. Exactly. And it's the same logic 00:04:44.670 --> 00:04:46.709 for asset B, the scene reference. If you want 00:04:46.709 --> 00:04:49.149 a pirate ship, find a clean image of a deck. 00:04:49.209 --> 00:04:51.269 It doesn't need to be some artistic masterpiece. 00:04:51.670 --> 00:04:54.310 It just needs to tell the AI, hey, here's the 00:04:54.310 --> 00:04:56.790 floor, here's the mast. We're anchoring the hallucination. 00:04:56.990 --> 00:04:58.709 That is the perfect way to put it. Language is 00:04:58.709 --> 00:05:00.829 just too slippery. If I type rugged captain, 00:05:01.149 --> 00:05:03.790 the AI has... What, a million definitions for 00:05:03.790 --> 00:05:07.329 that? But a specific JPEG of a face. That's hard 00:05:07.329 --> 00:05:09.610 data. You're putting a leash on the randomness 00:05:09.610 --> 00:05:11.810 before you even introduce the chaos of motion. 00:05:12.230 --> 00:05:15.730 Speaking of language being slippery, let's talk 00:05:15.730 --> 00:05:18.750 about the prompting itself. I am so guilty of 00:05:18.750 --> 00:05:21.050 the adjective soup approach. We all are. I'll 00:05:21.050 --> 00:05:23.970 just type cinematic, moody, dark, cool lighting, 00:05:24.110 --> 00:05:27.110 8K, masterpiece, and I'm just expecting Ridley 00:05:27.110 --> 00:05:28.910 Scott to pop out. And you get a blurry mess. 00:05:29.519 --> 00:05:32.720 Max Anne is brutal on this point. He says writing 00:05:32.720 --> 00:05:36.439 prompts by hand in 2026 is a huge mistake. Humans 00:05:36.439 --> 00:05:39.079 use adjectives. Machines need engineering specs. 00:05:39.339 --> 00:05:42.000 So we use the LMM as a prompt engine. Yes. You 00:05:42.000 --> 00:05:43.660 don't write the prompt yourself. You talk to 00:05:43.660 --> 00:05:45.980 Claude or ChatGPT and it writes the prompt for 00:05:45.980 --> 00:05:48.720 you. But you don't just ask it nicely. You give 00:05:48.720 --> 00:05:51.639 it this massive custom instruction block. I was 00:05:51.639 --> 00:05:53.540 looking at the template from the source. It is 00:05:53.540 --> 00:05:55.860 so rigorous. It has the specific character count 00:05:55.860 --> 00:05:59.100 between 2 ,200 and 3 ,000 characters. Why that 00:05:59.100 --> 00:06:01.339 range? Why not just write a novel? It all comes 00:06:01.339 --> 00:06:03.480 down to the attention mechanism in the image 00:06:03.480 --> 00:06:05.759 generator. Anne notes that if you're under 2 00:06:05.759 --> 00:06:08.279 ,200 characters, there's just not enough density 00:06:08.279 --> 00:06:11.199 to force a specific style. But if you go over 00:06:11.199 --> 00:06:15.480 3 ,000, the model gets overwhelmed. It suffers 00:06:15.480 --> 00:06:17.699 from what he calls loss of focus. It just starts 00:06:17.699 --> 00:06:22.589 ignoring things. Around 2 ,800 characters. That's 00:06:22.589 --> 00:06:24.529 the sweet spot where the AI is forced to pay 00:06:24.529 --> 00:06:26.949 attention to everything. That is incredibly specific. 00:06:27.129 --> 00:06:29.009 It's like finding the exact resonant frequency 00:06:29.009 --> 00:06:30.990 of the model. But the part of that template that 00:06:30.990 --> 00:06:32.930 really stood out to me was the grounding block. 00:06:33.170 --> 00:06:35.250 Yeah, this is mandatory. You paste this text 00:06:35.250 --> 00:06:37.089 at the end of every single prompt you generate. 00:06:37.310 --> 00:06:40.189 And it explicitly demands things like real materials, 00:06:40.529 --> 00:06:43.310 real lighting, real physics, and it bans stuff 00:06:43.310 --> 00:06:46.379 like fantasy glow. or illustrative techniques. 00:06:46.639 --> 00:06:49.579 No fantasy glow. I feel like default AI art is 00:06:49.579 --> 00:06:52.040 90 % fantasy glow. Everything looks like it's 00:06:52.040 --> 00:06:54.399 been smeared with Vaseline. Why is that? It does, 00:06:54.420 --> 00:06:56.199 and there's a technical reason. These models 00:06:56.199 --> 00:06:58.879 are trained on millions of images. So when they 00:06:58.879 --> 00:07:00.860 get confused or when they try to optimize an 00:07:00.860 --> 00:07:03.420 image, they just regress to the mean, they smooth 00:07:03.420 --> 00:07:05.699 everything out, they average the data. So smoothness 00:07:05.699 --> 00:07:08.800 is actually the AI failing to be specific. In 00:07:08.800 --> 00:07:12.600 a way, yeah. Realism is messy. realism has grain 00:07:12.600 --> 00:07:17.620 dust noise sharp edges by banning stylized rendering 00:07:17.620 --> 00:07:20.439 and demanding gravity you're forcing the model 00:07:20.439 --> 00:07:23.579 to stop averaging and start making specific gritty 00:07:23.579 --> 00:07:26.339 choices you're fighting that plastic look you're 00:07:26.339 --> 00:07:28.970 telling the cinematographer Don't use the beauty 00:07:28.970 --> 00:07:31.730 filter. I want to see the pores. Exactly. You 00:07:31.730 --> 00:07:34.129 want the imperfections. That's what our brains 00:07:34.129 --> 00:07:36.730 register as real. Okay. So we have our neutral 00:07:36.730 --> 00:07:38.769 assets. We have this massive technical prompt 00:07:38.769 --> 00:07:41.850 from our LLM. Now we go to the visual -based 00:07:41.850 --> 00:07:44.870 Nano Banana Pro to make the image. And here, 00:07:44.990 --> 00:07:47.829 the guide suggests a workflow hack he calls the 00:07:47.829 --> 00:07:50.709 two -by -two grid. The contact sheet strategy. 00:07:50.870 --> 00:07:53.189 Love this. Instead of making one image, you generate 00:07:53.189 --> 00:07:56.769 four at once. Is that just to save time? It's 00:07:56.769 --> 00:07:58.990 about consistency and coverage. Think about a 00:07:58.990 --> 00:08:00.970 real film shoot. You don't just set up one camera, 00:08:01.029 --> 00:08:03.269 take one shot and move on, right? You get coverage. 00:08:03.529 --> 00:08:06.529 Wide shot, medium, close up, a reverse angle. 00:08:06.670 --> 00:08:08.810 Exactly. But usually in AI, if I generate four 00:08:08.810 --> 00:08:10.410 images, they look like four totally different 00:08:10.410 --> 00:08:12.949 movies. That's where Nano Banana Pro is different. 00:08:13.389 --> 00:08:16.470 If you run a two by two grid in a single generation, 00:08:16.870 --> 00:08:20.029 all four images usually share the same seed noise. 00:08:20.149 --> 00:08:22.370 That means the lighting, the color palette, the 00:08:22.370 --> 00:08:25.990 texture. It all stays consistent across all four 00:08:25.990 --> 00:08:28.509 angles. So we get a wide shot, an over -the -shoulder, 00:08:28.649 --> 00:08:30.490 and a close -up that actually look like they 00:08:30.490 --> 00:08:32.250 were filmed on the same day with the same camera. 00:08:32.370 --> 00:08:34.570 That's the idea. You get an instant scene kit, 00:08:34.730 --> 00:08:36.690 you just pick the best one, crop it, and you're 00:08:36.690 --> 00:08:38.929 good to go. It saves you from burning hundreds 00:08:38.929 --> 00:08:41.529 of credits on 50 disconnected attempts. That's 00:08:41.529 --> 00:08:43.990 huge for continuity. Now, there's one more hack 00:08:43.990 --> 00:08:45.830 in this section that I just have to mention it 00:08:45.830 --> 00:08:47.929 because it's so simple, but the reasoning blew 00:08:47.929 --> 00:08:51.250 my mind. The aspect ratio. Ah, the cinematic 00:08:51.250 --> 00:08:54.629 hack. The guide basically screams, use 21 .9 00:08:54.629 --> 00:08:57.389 ultra widescreen. And he says, to strictly avoid 00:08:57.389 --> 00:09:01.090 16 .9, which is your standard TV shape, why does 00:09:01.090 --> 00:09:02.990 the shape of the rectangle change the quality 00:09:02.990 --> 00:09:05.470 of what's inside it? That sounds like magic. 00:09:05.690 --> 00:09:07.350 It's all about the training data. The neural 00:09:07.350 --> 00:09:09.450 network is just a giant association machine. 00:09:09.590 --> 00:09:12.029 So think about it. What kind of images in the 00:09:12.029 --> 00:09:16.570 world are cropped to 16 .9? TV shows, the news, 00:09:16.970 --> 00:09:20.009 YouTube videos. Right. And what do those generally 00:09:20.009 --> 00:09:22.769 look like? Kind of flat lighting, digital cameras, 00:09:22.909 --> 00:09:26.190 broadcast quality. Okay. Now... What kind of 00:09:26.190 --> 00:09:29.490 images are cropped to 21 .9? Big budget Hollywood 00:09:29.490 --> 00:09:32.690 movies. Exactly. So when you force the aspect 00:09:32.690 --> 00:09:35.870 ratio to 21 .9, you're subconsciously triggering 00:09:35.870 --> 00:09:39.470 the movie magic pathways in the AI's brain. It 00:09:39.470 --> 00:09:42.610 associates that wide rectangle with better color 00:09:42.610 --> 00:09:44.710 grading, more dramatic lighting, higher production 00:09:44.710 --> 00:09:47.710 value. That is wild. So just by changing the 00:09:47.710 --> 00:09:49.990 crop, I'm tricking the AI into thinking we're 00:09:49.990 --> 00:09:51.710 making a blockbuster. You're hacking the data 00:09:51.710 --> 00:09:54.190 set. You ask for 16 .9, you get a soap opera. 00:09:54.330 --> 00:09:56.929 You ask for... 21 .9, you get Dune. It's a probability 00:09:56.929 --> 00:10:00.149 game. I love that. And then there's a quick mention 00:10:00.149 --> 00:10:02.629 of an optional polish step using something like 00:10:02.629 --> 00:10:05.730 Topaz Gigapixel to add fabric fibers, dust, stone, 00:10:05.830 --> 00:10:08.029 just really leaning into that texture we talked 00:10:08.029 --> 00:10:10.090 about. It's all about removing that digital sheen. 00:10:10.210 --> 00:10:12.690 You want the viewer to subconsciously feel the 00:10:12.690 --> 00:10:14.610 grit. All right. So at this point in the murder 00:10:14.610 --> 00:10:18.450 board method, we have a beautiful, gritty 21 00:10:18.450 --> 00:10:21.269 .9 still image. It looks like a movie still, 00:10:21.409 --> 00:10:23.840 but it's frozen. Now we have to make it move. 00:10:24.000 --> 00:10:25.440 And this is where it all falls apart for most 00:10:25.440 --> 00:10:27.299 people. This is where it falls apart for me. 00:10:27.340 --> 00:10:29.220 The shimmering, the weird morphing, the floating. 00:10:29.440 --> 00:10:33.840 Enter Kling 2 .6, the motion engine. Right. So 00:10:33.840 --> 00:10:37.059 Kling is the tool of choice here because it respects 00:10:37.059 --> 00:10:41.100 that 21 .9 aspect ratio we fought so hard for. 00:10:41.259 --> 00:10:44.620 But the real secret isn't the software. It's 00:10:44.620 --> 00:10:46.639 the video prompt. Which is totally different 00:10:46.639 --> 00:10:48.279 from the image prompt. Completely different. 00:10:48.360 --> 00:10:50.279 In the image prompt, you describe the scene. 00:10:50.720 --> 00:10:52.740 In the video prompt, you describe the camera. 00:10:52.919 --> 00:10:54.980 And there's one instruction in the murder board 00:10:54.980 --> 00:10:57.759 guide that seems completely counterintuitive. 00:10:58.000 --> 00:11:01.480 The shaky cam. Yeah. Camera is handheld and visibly 00:11:01.480 --> 00:11:05.440 shaky. It lists terms like persistent micro jitter, 00:11:05.620 --> 00:11:09.100 imperfect motion blur, breathing. It sounds like 00:11:09.100 --> 00:11:11.820 you're asking for a bad cameraman. Yeah. I usually 00:11:11.820 --> 00:11:14.100 want those smooth gliding drone shots. Why are 00:11:14.100 --> 00:11:16.460 we asking for shake? Because smoothness is the 00:11:16.460 --> 00:11:20.159 enemy of realism in AI video. Explain that. Well, 00:11:20.220 --> 00:11:22.620 when an AI generates perfectly smooth motion, 00:11:22.759 --> 00:11:25.019 it often looks floaty. The physics just don't 00:11:25.019 --> 00:11:26.460 feel quite right. It falls straight into the 00:11:26.460 --> 00:11:29.360 uncanny valley. But when you add micro jitter 00:11:29.360 --> 00:11:32.340 and handheld shake, that chaotic motion works 00:11:32.340 --> 00:11:35.059 like visual camouflage. It hides the crimes. 00:11:35.220 --> 00:11:38.200 It hides the crimes, exactly. If a background 00:11:38.200 --> 00:11:41.240 element warps a little bit or a shadow glitches, 00:11:41.340 --> 00:11:44.080 the viewer's brain just forgives it because the 00:11:44.080 --> 00:11:46.419 whole frame is shaking. It masks the artifacts. 00:11:47.440 --> 00:11:49.559 It's like a magician using misdirection to hide 00:11:49.559 --> 00:11:52.299 a cut. And it adds emotional texture. A handheld 00:11:52.299 --> 00:11:55.039 camera feels like a documentary. It feels like 00:11:55.039 --> 00:11:58.039 you're there. A perfect glide feels like a computer 00:11:58.039 --> 00:12:00.299 simulation. I have to admit, this is a bit of 00:12:00.299 --> 00:12:03.039 a vulnerable moment for me. I still wrestle with 00:12:03.039 --> 00:12:05.679 prompt tricked all the time. I'll have a character, 00:12:05.759 --> 00:12:07.679 and as soon as they turn their head, they're 00:12:07.679 --> 00:12:11.360 a different person. This shaky cam trick, it 00:12:11.360 --> 00:12:13.710 feels like a cheat code. I wish I knew a year 00:12:13.710 --> 00:12:15.750 ago. It pretty much is. But we have to be honest 00:12:15.750 --> 00:12:17.690 about the limits. The guide calls it the Hulk 00:12:17.690 --> 00:12:19.409 out problem. The Hulk out. You've seen this. 00:12:19.549 --> 00:12:22.629 The character smiles and suddenly their jaw unhinges 00:12:22.629 --> 00:12:25.009 or their neck muscles bulge out like a bodybuilder. 00:12:25.029 --> 00:12:27.110 Or they grow a second row of teeth. Yeah, it's 00:12:27.110 --> 00:12:30.309 horrifying. So cling 2 .6 is good, but it's not 00:12:30.309 --> 00:12:33.110 magic. The work around here is just volume. You 00:12:33.110 --> 00:12:35.389 generate three to five variations of the same 00:12:35.389 --> 00:12:37.919 motion. Trim the ends, right? Always trim the 00:12:37.919 --> 00:12:40.360 ends. The first second is usually the AI figuring 00:12:40.360 --> 00:12:42.679 out the physics. And the last second is where 00:12:42.679 --> 00:12:44.600 it runs out of steam and everything starts to 00:12:44.600 --> 00:12:47.320 melt. The gold is in the middle. So perfection 00:12:47.320 --> 00:12:50.980 looks fake. Chaos looks real. That's the lesson 00:12:50.980 --> 00:12:53.460 here. That's the mantra. We have one more big 00:12:53.460 --> 00:12:56.720 segment to get to. The secret weapon. And honestly, 00:12:56.879 --> 00:12:59.379 this is the part that made me feel like I needed 00:12:59.379 --> 00:13:01.519 to go back to film school to even understand 00:13:01.519 --> 00:13:03.919 it. It's the biggest level up in the entire guide. 00:13:04.279 --> 00:13:06.139 We'll get into the metadata magic of Shotdeck 00:13:06.139 --> 00:13:10.740 right after this. Midroll sponsor read. Okay, 00:13:10.759 --> 00:13:12.960 we are back. We are deep in the murder board 00:13:12.960 --> 00:13:15.200 method, and we've reached the final piece of 00:13:15.200 --> 00:13:18.710 this puzzle. The secret weapon. Shotdeck. Now, 00:13:18.750 --> 00:13:21.090 I know ShotDeck, it's a tool for, you know, real 00:13:21.090 --> 00:13:23.210 filmmakers, directors, DPs to find reference 00:13:23.210 --> 00:13:26.049 images. It's this huge library of high -res frames 00:13:26.049 --> 00:13:28.929 from actual movies. How does that fit into an 00:13:28.929 --> 00:13:31.049 AI workflow? This is all about moving beyond 00:13:31.049 --> 00:13:33.730 the word cinematic. The guide argues that cinematic 00:13:33.730 --> 00:13:37.210 is a lazy word. It means nothing to an AI. It's 00:13:37.210 --> 00:13:40.029 too vague. Cinematic could be the Avengers or 00:13:40.029 --> 00:13:41.809 it could be the Godfather. Two very different 00:13:41.809 --> 00:13:44.490 things. Exactly. So the workflow is this. You 00:13:44.490 --> 00:13:47.220 go to ShotDeck. You find a frame from a real 00:13:47.220 --> 00:13:49.879 movie that has the exact look you want. Maybe 00:13:49.879 --> 00:13:52.940 it's Andor for that gritty, industrial, desaturated 00:13:52.940 --> 00:13:56.000 look. Or maybe it's 2001, a space odyssey for 00:13:56.000 --> 00:13:58.559 that sterile, precise, bright white feeling. 00:13:58.620 --> 00:14:01.330 Okay. So I find my shot. And you don't just look 00:14:01.330 --> 00:14:03.929 at the picture, you look at the metadata. ShotDeck 00:14:03.929 --> 00:14:06.230 lists the specific lens that was used, the camera 00:14:06.230 --> 00:14:08.870 body, the f -stop, the lighting diagram, even 00:14:08.870 --> 00:14:11.070 the film stock. We're talking literal hardware 00:14:11.070 --> 00:14:13.909 specs. Extremely specific hardware specs. You 00:14:13.909 --> 00:14:16.250 screenshot that data, you feed it to your LLM, 00:14:16.269 --> 00:14:18.549 and you say, extract these specs, and then you 00:14:18.549 --> 00:14:21.090 paste that exact technical data into your image 00:14:21.090 --> 00:14:23.090 prompt. Whoa, wait. So instead of saying cool 00:14:23.090 --> 00:14:26.009 lighting, I'm telling the AI what exactly? You 00:14:26.009 --> 00:14:37.460 are telling the AI, cook as... Does the AI, does 00:14:37.460 --> 00:14:39.750 it actually know what that means? And that's 00:14:39.750 --> 00:14:42.690 the moment of wonder. Yes, it does. Because the 00:14:42.690 --> 00:14:45.090 AI was trained on the entire internet. It was 00:14:45.090 --> 00:14:47.990 trained on photography forums and tech reviews 00:14:47.990 --> 00:14:51.269 and film databases where all these images were 00:14:51.269 --> 00:14:54.230 tagged with that exact data. It knows how light 00:14:54.230 --> 00:14:57.509 physically bends through a 50mm Cooke lens versus 00:14:57.509 --> 00:15:00.700 how it bends through a 14mm fisheye. That is 00:15:00.700 --> 00:15:02.720 incredible. So it's not just applying a look 00:15:02.720 --> 00:15:05.679 or a filter. It's simulating the physics of the 00:15:05.679 --> 00:15:08.500 glass. It stops guessing. It stops hallucinating 00:15:08.500 --> 00:15:11.139 a style and starts emulating a specific engineering 00:15:11.139 --> 00:15:14.419 signature. The guide says when you do this, the 00:15:14.419 --> 00:15:17.019 lighting gains real depth. Shadows hold detail. 00:15:17.240 --> 00:15:19.700 The image stops looking like AI art and starts 00:15:19.700 --> 00:15:21.960 looking like actual photography. This feels like 00:15:21.960 --> 00:15:24.179 a fundamental shift in what the creator is even 00:15:24.179 --> 00:15:26.850 doing. If I'm just copying lens data from Blade 00:15:26.850 --> 00:15:28.750 Runner, am I really a prompter anymore? That 00:15:28.750 --> 00:15:30.710 is the big question. The guide suggests you're 00:15:30.710 --> 00:15:32.769 not a writer finding cool adjectives. You're 00:15:32.769 --> 00:15:35.129 a technical director managing a list of equipment 00:15:35.129 --> 00:15:37.990 specs. You're building a virtual camera rig inside 00:15:37.990 --> 00:15:40.350 the machine. It's less once upon a time and more. 00:15:40.919 --> 00:15:44.000 Set aperture to f1 .4. And that's a hard pivot 00:15:44.000 --> 00:15:45.779 for a lot of creative types who just want to 00:15:45.779 --> 00:15:48.799 tell a story. But in 2026, that technical precision 00:15:48.799 --> 00:15:51.600 is the difference between amateur and professional 00:15:51.600 --> 00:15:54.200 output. So let's pull this all together. We have 00:15:54.200 --> 00:15:56.000 the murder board to track our assets. We have 00:15:56.000 --> 00:15:58.440 the LLM creating these engineered prompts. We've 00:15:58.440 --> 00:16:01.240 got the 21 .9 aspect ratio hacking the training 00:16:01.240 --> 00:16:04.299 data. And we have ShotDeck providing the actual 00:16:04.299 --> 00:16:07.960 lens physics. It's a complete end -to -end ecosystem. 00:16:08.440 --> 00:16:10.519 If I'm a listener and I'm just sort of... dipping 00:16:10.519 --> 00:16:13.809 my toes into this. What is the big idea? Is it 00:16:13.809 --> 00:16:15.870 that I need to buy all these new tools? No, not 00:16:15.870 --> 00:16:18.389 at all. The big takeaway is that creativity is 00:16:18.389 --> 00:16:20.610 no longer the bottleneck. The bottleneck is discipline. 00:16:20.950 --> 00:16:23.029 Discipline. The murder board method is fundamentally 00:16:23.029 --> 00:16:25.809 about documentation and constraints. It's about 00:16:25.809 --> 00:16:28.970 resisting that urge to just hit generate and 00:16:28.970 --> 00:16:30.710 see what happens. It's about doing the prep work. 00:16:30.789 --> 00:16:32.889 It's all the boring stuff. The neutral face assets. 00:16:33.090 --> 00:16:35.590 Yeah. The 2 ,800 character limits. Right. And 00:16:35.590 --> 00:16:38.590 using handheld motion to mask the AI's flaws. 00:16:38.669 --> 00:16:41.730 It's a rigorous system. Max Ann argues that in 00:16:41.730 --> 00:16:45.049 2026, The real AI director doesn't write stories. 00:16:45.210 --> 00:16:47.990 They manage visual continuity. That is a powerful 00:16:47.990 --> 00:16:51.210 thought. Because anyone can generate one cool 00:16:51.210 --> 00:16:55.129 image. But can you generate 50 cool images that 00:16:55.129 --> 00:16:57.250 all look like they exist in the same universe? 00:16:57.590 --> 00:16:59.590 That's the real challenge. And that's the difference 00:16:59.590 --> 00:17:01.750 between a slot machine and a film crew. Well, 00:17:01.850 --> 00:17:04.170 I am definitely going to try this two -by -two 00:17:04.170 --> 00:17:06.769 grid approach. Even if I'm not making a movie, 00:17:06.869 --> 00:17:09.130 just seeing the difference in composition in 00:17:09.130 --> 00:17:12.380 one run. Seems incredibly valuable. I think you 00:17:12.380 --> 00:17:14.440 should. It forces you to think about framing 00:17:14.440 --> 00:17:17.839 wide, medium, close instead of just content. 00:17:18.039 --> 00:17:19.819 And I challenge everyone listening to try it 00:17:19.819 --> 00:17:22.480 too. Next time you open your tool of choice, 00:17:22.660 --> 00:17:25.960 don't just generate one square. Ask for a contact 00:17:25.960 --> 00:17:28.500 sheet. See if you can get that continuity. And 00:17:28.500 --> 00:17:30.500 maybe build your own murder board. Get the red 00:17:30.500 --> 00:17:32.700 string out. I think I'm going to need a bigger 00:17:32.700 --> 00:17:35.259 cork board. We all are. That's it for this deep 00:17:35.259 --> 00:17:37.450 dive into the murder board method. Thanks for 00:17:37.450 --> 00:17:39.890 listening and we will catch you in the next one. 00:17:39.950 --> 00:17:40.589 Keep creating.