The argument in one line.
A folder of markdown files on your desktop, pointed at Claude in code mode with an MCP connection to an image/video generator, functions as a complete AI film production office that retains full project context and generates shots in the background while you edit.
Read if. Skip if.
READ IF YOU ARE…
- You are already generating AI film footage and want a way to manage prompts, character references, scene boards, and asset tracking without switching between five tools.
- You have tried node-based AI pipelines and found them hard to customize or maintain.
- You use Claude regularly and want a concrete example of how MCP connections change what it can do in a creative production context.
- You want an honest cost and failure report from a real AI short film — specific numbers, specific bugs, specific workarounds.
SKIP IF…
- You are working on live-action or hybrid production — this workflow is 100% AI-generated footage.
- You want tool-agnostic advice — Martini is the sponsor and MCP target throughout.
The full version, fast.
Instead of subscribing to an all-in-one AI film platform, you can build a free production office from a folder of markdown files that Claude in code mode reads as its context. The folder holds story notes, style guides, character reference grids, scene boards, and a production tracker. An MCP connection lets Claude reach into an image/video generator and fire shots by chat while you continue editing in parallel. The post-mortem on Paperclip Heart — 53 shots, 160 generations — surfaces the real gotchas: Claude defaulting to Opus for menial tasks, context windows filling up mid-project, ring-light prompts producing robotic eyes, and character consistency breaking in multi-role shots. The handoff doc pattern and icon-shaped reference images are the two most transferable fixes.
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 →Where the time goes.
00:00 – 00:29
01 · The Future AI Film Workflow (No Nodes)
Cold open promise: the future workflow is not nodes.
00:29 – 00:57
02 · Paperclip Heart: The Case Study
Introduces the film this workflow built; recommends watching it first.
00:57 – 02:32
03 · Pre-Production: Claude + Deep Research
Story brainstorming with Claude (creative writing); fact-finding with Gemini Deep Research (20-page paper). Hybrid prompt/dialogue script format.
02:32 – 03:45
04 · Building the Production Office (One Folder)
Folder lives on desktop, parsed by Claude in code/cowork mode. QUICKSTART.md is the entry point.
03:45 – 05:28
05 · What It Does: Ref Grids, Scene Boards, Trackers
Story breakdown, style guide, character reference grids (Nano Banana Pro 2K), 2x2 scene boards, CDance 2.0 prompts, production brief, production tracker.
05:28 – 06:30
06 · MCPs: The Big Unlock
MCP = Model Context Protocol. Gives Claude hands into external platforms. Production office now asks: manual, MCP, or hybrid mode?
06:30 – 07:16
07 · Connecting Claude to Martini
Gear icon -> MCP server on -> paste URL -> always allow. Claude Code: terminal command. Also works on Gemini, ChatGPT.
07:16 – 08:36
08 · Directing Shots by Chat
Live demo: conversational shot request, Nano Banana Pro 2K, ~3.5-4 min per round-trip. Claude warned about In-N-Out trademark twice.
08:36 – 09:09
09 · Working in Parallel (Mind the Tokens)
Generate one scene while editing another. Downloading assets via MCP burns tokens — manual download + Claude organize is cheaper.
09:09 – 10:36
10 · Where It Breaks: Big-Brain Models on Dumb Jobs
Claude defaults to Opus for folder cleanup. Tracker goes stale. These are solvable but require active management.
10:36 – 11:35
11 · Context Windows & Handoff Docs
When context pushes red: ask for a handoff doc. New session reads the doc, starts at zero debt.
11:35 – 12:37
12 · Production Headaches: Ring Lights & Robot Eyes
Ring light in CDance 2.0 prompt = robotic eyes. Talk show host got content-flagged. Baby-with-glasses: 7 attempts.
12:37 – 13:42
13 · Character Consistency: Eli & Default Face
Eli (tech CEO) ran clean on omni reference. Default Face required icon-shaped reference image for multi-role shots — Claude suggested this fix.
13:42 – 14:24
14 · Editing with Full Project Context
Claude knows the whole film — mid-edit shot requests are fast because context is already loaded.
14:24 – 15:40
15 · Martini's Built-In Editor (XML Export Trick)
Basic NLE in Martini for rough-cutting during generation. XML export to DaVinci Resolve or Premiere.
15:40 – 17:07
16 · The Audio Cleanup Chain
Adobe Podcast Enhance -> Brainworx Cleansweep (free) -> multiband compressor (broadcast preset) -> Suno music -> Valhalla Supermassive reverb (free).
17:07 – 17:57
17 · Total Time & Cost: 53 Shots, 160 Generations
~3:1 generation-to-shot ratio. Abandoned: AI influencer storyline, CDance Jake Gyllenhaal outputs.
17:57 – 18:56
18 · What's Next: Gaussian Splats & Multiplayer
Martini's 'step into set' (Gaussian splat + camera reposition). Multiplayer canvas. MCP control of these features is still uncertain.
18:56 – 19:35
19 · Wrap-Up
Sponsor thanks (Martini), sign-off from Tim.
Lines worth screenshotting.
- A markdown folder on your desktop beats a node-based pipeline for AI filmmaking — it is customizable, portable, and Claude retains everything across sessions.
- The real value of an AI production office is retained project context, not generation speed — it makes mid-edit shot requests accurate without re-briefing.
- MCPs turn Claude from a text assistant into an agent with hands inside external platforms — the difference between directing by chat and clicking through a UI manually.
- Generating one scene while editing another is only possible if the agent has full project context loaded; without it you are back to babysitting every request.
- Claude defaults to Opus for menial tasks like folder cleanup unless you explicitly tell it to route those to Haiku — and Opus costs significantly more per token.
- The handoff doc pattern lets you reset a bloated context window without losing any project state: ask Claude to write a handoff doc, open a new session, point at it, continue.
- Ring lights as a term in a CDance 2.0 prompt reliably produces robotic eyes; removing the word fixed the issue immediately.
- Character consistency is easier with a tech-CEO archetype using omni reference than with a female AI companion who needs to appear in multiple configurations in the same shot.
- 160 generations for 53 usable shots is a roughly 3:1 ratio — a useful benchmark for budgeting AI film production time and credits.
- The homebrew advantage over managed AI studio platforms is direct visibility into what breaks and why — which generates ideas for what to automate next.
- An icon-shaped face reference image solves the character-appearing-in-two-roles-in-one-shot problem better than a photorealistic reference.
- XML export from a basic NLE into DaVinci Resolve already works — the bridge between AI generation and professional editorial exists today.
- A multiband compressor at broadcast preset is the single highest-leverage audio move for AI-generated dialogue — it fixes the dynamic range problem that makes synthetic voices sound thin.
- The Suno plus Valhalla Supermassive combination (both free tier available) can produce a convincing cinematic ambient score without licensed music costs.
Build the office, own the workflow.
WHAT TO LEARN
A folder of markdown files and a Claude MCP connection gives you a production office that retains full project context — and the three lessons that matter most are about token cost, context hygiene, and the one prompt word that breaks everything.
- A self-contained project folder that Claude reads as its context is more durable than any platform subscription — you own it, you customize it, and it travels with you across any tool that can parse markdown.
- MCP connections let an AI agent generate assets in the background while you work on something else; without them, every shot request is a context-switch that breaks your editing flow.
- When Claude defaults to a large model for menial tasks like file cleanup, you pay the full token rate for work that a smaller model could handle — explicitly routing low-stakes tasks to Haiku or Sonnet is a habit worth building.
- A context window handoff doc is the equivalent of a project save state: ask for it before the session fills, open a fresh session, point at the doc, and you resume with zero debt and no lost context.
- The ring-light bug is a specific example of a general principle: AI video models have undocumented visual triggers — when a character looks wrong, suspect the descriptive nouns in your prompt before anything else.
- Icon-shaped or simplified reference images outperform photorealistic references for characters who need to appear in multiple configurations in the same generated shot.
- A 3:1 generation-to-usable-shot ratio is a realistic planning benchmark for AI short film production — budget credits and time accordingly rather than assuming every generation is a keeper.
- The audio cleanup chain (detin + high/low pass + multiband compressor at broadcast preset) is transferable to any AI-generated voice project — it is the difference between synthetic-sounding and broadcast-ready.
Terms worth knowing.
- MCP (Model Context Protocol)
- An open standard that lets an AI agent reach into external applications and platforms to control them programmatically — turning a chat model into an agent with real-world hands.
- Omni reference
- A character consistency feature in some AI image generators that lets you upload a reference image and have the model reproduce that specific character across new generated shots.
- Nano Banana Pro 2K
- A high-resolution AI image generation model available inside Martini, used for creating character reference grids and scene boards at 2K resolution.
- CDance 2.0
- An AI video generation model used for text-to-video shot generation within the Martini platform. Known quirk: ring light in prompts causes robotic eyes.
- Gaussian splat
- A 3D scene representation technique that converts a 2D image into a navigable 3D space, letting you reposition a virtual camera within the scene.
- Handoff doc
- A summary document generated by Claude at the end of a session capturing all current project state, context, and next steps — used to restart a new session from zero context debt.
- Context window
- The maximum amount of text (measured in tokens) that a language model can hold in active memory within a single session — when this fills up, older context is dropped.
Things they pointed at.
Lines you could clip.
00:00
“Today, we're taking a look at what I consider to be the future workflow of AI filmmaking. And no, it is not nodes.”
09:30
“I would catch it when I sent it off to do dumb things — clean up the folder — and it was using Big Brain Opus. I'm like, no. That's costing a lot of money. Why don't you send the stupid one to go clean it up?”
13:35
“Just don't use ring lights. Don't call out ring lights in your prompt.”
03:11
“What all of this is is a set of instructions for Claude to read. It's really meant for Claude — for it to become your production office.”
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
Today, we're taking a look at what I consider to be the future workflow of AI filmmaking. And no, it is not nodes. Breathe a sigh of relief, node haters.
Rather, this is something that you'll be seeing quite a bit of in, well, different iterations over the next few months. Like all game changers, there will be good versions of it and there will be bad versions of it. So today at baseline, I'm gonna go over how you can build a customized one for yourself plus some tips and tricks I learned along the way.
Kicking off, a few days ago, I released Paperclip Heart, a short kind of Black Mirror esque film about the AI apocalypse that no one sees coming. Yes, your AI girlfriend is going to enslave humanity. If you haven't caught the film yet, I I mean, I do recommend checking it out first.
The link is down below. Mostly because we're gonna talk about it in the context of this workflow and, you know, I mean, I'm pretty proud of this one. So do feel free to hop over and give it a watch or, you know, if you want to watch it after.
Although, spoilers ahead. And to that, before we actually head into production, I did just briefly wanna talk about preproduction and kind of the why of it all. As someone in the comments mentioned, no, this was not a weekend project.
It it was definitely a longer stew for sure. Although, I do have to say production itself did go fairly quickly, but the idea itself has been noodling around for quite a bit. Well, I mean, to be honest, uh, since OpenAI suggested that they were going to release a spicy mode for chat GPT voice.
They did ultimately pull back on that particular side quest, but I would say that combined with the doomer Galvanic Sable scenario that was presented in if anyone builds it, everyone dies. Like, both of those kinda got the old creative noggin humming.
What followed was a lot of brainstorming and research, kind of poking at it in the background over, like, the last couple of months. I utilized Claude, which in my opinion is the more creative writer for the story breaks and arcs and Gemini's deep research for the fact finding and logic. In fact, deep research actually came up with a, like, a 20 page paper on the technological, economic, and societal ramifications, which actually becomes the basis for most of the news report sections in the film.
Now I do wanna note that the script itself wasn't written in, you know, traditional script format. There really wasn't a need to do it that way. Uh, instead, it's kind of done in this, like, hybrid, like, prompt and dialogue version that actually, well, Claude and I came up with.
That said, I mean, for a production document, for my use, this worked totally fine. So after our script was locked, it was time to move into To that, I once again revisited the AI film pipeline from Dragon Blue, which was the last AI film that I did.
So I'm not gonna go too deep into the woods on this. I did cover this in our last production breakdown. But, you know, essentially, this is a folder that lives on my desktop that can plug into Claude or ChatGPT or basically anything that can parse an MD file, which is markdown file or kind of like the instructions for one of these LLM systems.
If you're new to all of this, don't stress it. It actually is very easy. But I will say that I do recommend, uh, using Claude in either co work or code mode.
Uh, basically, that isn't chat. The reason for this is that once again, we want this project living on our desktop or somewhere on our hard drive. Uh, just not in a folder called stuff.
That is a deep cut callback to longtime viewers. Basically, yeah, you want it living somewhere on your computer so it knows where your stuff is. What all of this is is a set of instructions for Claude to read.
I mean, you're not supposed to read this. You can if you want to, but, I mean, it's really meant for Claude, um, for it to become . All you have to do is open up a new session, point it at your folder, and ask it to review the quick start.
And, well, I mean, you're pretty much off and running. Now what are you running? Well, again, I mean, it is pretty much your AI film production office.
Um, so it'll do a story breakdown of anything that you throw into it, breaking things down into scenes. Um, you can go through a style guide with it where it'll lock in the style, uh, character reference grids for all of your characters that can be generated in Nano Banana Pro two k or really whatever you want.
Um, we'll go over that later on. Um, it'll do every all of your scenes as a two by two grid, um, in Nano Banana Pro or once again image two if you want to or really again anything else.
Um, this particular one has CDance two point o prompts that I personally like to use. Again, you can swap them out, but the default there is the prompt format that I I like. And then it also has a production brief and a full production tracker as well.
So, um, yeah. I mean, it is again, it's kind of your home office. Now is all of this useful?
Useful? In my opinion, yes. Very much so.
Considering that the project itself knows what your film is about, knows what it looks like, knows who's in it. So it really does speed up the creation process a pretty considerable amount.
The other big part of it is the asset tracking because I well, I mean, frankly, I'm I'm always a mess with that. So it's nice to have something that's keeping track of it. Is it perfect?
No. It's not. We'll go over that stuff later on, but I I mean, I'll point out that this is like the eighth iteration that I've done on this mostly because, again, things keep changing.
Uh, I would do wanna note that I'll have all of this available for you in this week's newsletter. I'll just, you know, have the whole folder. You can download it.
So do make sure you're subscribed below. But the big change that's happened since the last time, uh, I built this was the introduction of MCPs.
Now, MCPs are a big unlock. So big in fact that, uh, actually with this new skill, uh, the first time you load it up, uh, I actually had Claude ask, like, how do you wanna run this? Do you wanna run this with MCPs manually or some sort of hybrid format?
Briefly, if you're not familiar with what an MCP is beyond, uh, being the ultimate bad guy in Tron, the good Tron, the old old Tron, uh, an MCP is a model context protocol. At at, you know, very baseline, it allows for your agent to, you know, have hands that reach into external programs and platforms to control them. To that, Paperheart was generated via MCP over on martini.film who, yes, were kind enough to sponsor today's video.
Now, one thing I do wanna say is and just be very upfront with you all, m c p's are popping up on a lot of platforms. So I don't wanna front and say like this is exclusive to Martini. Uh, that said, again, I I happen to like Martini, there are a lot of other helpful features here as well as we'll go over later on in the video.
So hooking up, in this case, a clawed MCP is kind of stupidly simple. Um, you just come up to your little, uh, gear icon here. Um, make sure that MCP server here is turned on.
Um, and then there are some, like, directions that are here. It's all fairly easy in terms of directions. I do recommend using clawed desktop for this.
But basically, you know, you just add a connector in desktop and then point it at this URL and, you know, pretty much you're you you hit up, you know, always allow and off you go. Uh, in Claude code, there is a terminal command that you throw in, uh, very simple. And then, you know, you essentially open up Claude code, uh, hit forward slash m c p, and you're off and running.
Uh, there's also it's also available on Gemini and on ChatGPT as well, though I haven't actually hooked either of those two up. Up.
Once you've done that, I mean, you're pretty much just directing your production office conversationally. Um, so I can say something like, um, create a two by one, uh, image of Eli eating a hamburger in an In N Out.
Uh, let's use Nano Banana two. No.
Nano Banana Pro in two k. And there you go. Three images of our tech billionaire enjoying a hamburger at In N Out.
Um, it's funny. I did ask for, uh, three additional ones. It looks like the there's pushback on the GPT image two version, uh, but they they did come in via Nano Banana.
It's funny too because, again, production office here was like, hey. By the way, In N Out is a real trademark, so you might run into trouble with that. Uh, and then, you know, it ended up generating them out.
And then at the end, it was like, you know, by the way, again, you know, that's that's a real trademark, so you might wanna rethink that. Um, we'll talk about this part of it a little bit later on. Now did all of that happen quickly?
I mean, I don't know. Somewhere in the neighborhood of, like, three and a half to four minutes, um, including, you know, shot off to my Claude agent over to Martini generation and back.
Um, would it be faster doing it manually? Perhaps. But, you know, the idea here is that, uh, you can be working on simultaneous things as at once.
So this whole system does have visual context of what's happening in your project as well. Uh, and you can ask it like, hey, can you download that image or that video or anything from your canvas into my folder, um, and then organize it there. Although, that can be a bit of a token burner.
So, uh, honestly, mean, I think that it just I get save your tokens, just manually download it yourself, pop it in the folder, and then have Claude go through and organize it. As I said before, this can be a pretty massive time saver if you're using it to, like, generate one scene while you're working on another. Speaking of which, again, is this a perfect system?
No. It is not. Um, it's actually you know, there's actually still a lot of labor in here as always, and the system can get derailed from time to time.
One example is when you're sending out your agents to do menial tasks. Uh, if you're not familiar with the, like, this the Claude ecosystem, uh, Opus is kind of the big brain model. Fable just came out, like, like, a few hours ago, so that actually hasn't migrated its way into code just quite yet.
But, um, Opus is like the big brain model that, uh, you know, ends up costing more tokens. Uh, from there, there's Sonnet and Haiku. Oftentimes, like, I would catch it from time to time when I sent it off to do dumb things like, uh, you know, clean up the folder or whatever that it was using Big Brain Opus.
Um, and I'm like, no. That's costing a lot of money. Why why don't you send the stupid one to go clean it up?
Other thing is that it would quite often forget to update the tracker. Like, I would pop in just to see where we were with the project, find it kind of, like, wildly out of date, message Claude and be like, what come on, man. What's going on here?
And, like, to to Claude's credit, it would always be like, sorry, boss. On it. All of this is essentially a homebrew version of what we're gonna be seeing across a lot of the platforms or most of the platforms soon enough.
Uh, but what I do like about kind of this homebrew version of it is well, two things. One, you really, you know, get a chance to play with it and see how the thing is working. And two, you could really customize it to the way that you want to work.
Now one thing I do wanna note to folks that don't use Claude code very often is this little circle thing down here. If you pull that up, you can see the context window here. So within each session, um, you're gonna have an allotted amount of context.
When that starts pushing into the red, uh, basically, what you're gonna wanna do is is just ask Claude for a handoff doc. It'll populate that in the folder as well. And then all you have to do is open up a new session.
Uh, you're back to zero on your contacts and just say, like, hey. Go read the last, like, hand off doc, and it'll have all of the contacts of everything you talked about. Again, the bigger platforms are going to have sort of, like, behind the scenes systems that are going to take care of all of this.
But, you know, obviously, you know, you lose out a little bit on the manual control. This is the part that I I figured out. I like, hey.
Are you spawning sub agents for the dumb work, by the way? I was like, straight answer, one so far. I was like, caught you, Claude.
But ultimately, I mean, I do think that going this route is the most rewarding, mostly because, again, you you're sort of behind the scenes and, uh, you just kind of end up with getting more and more ideas about other things that you can do. Now in terms of, like, overall production headaches, I mean, kinda typical in all honesty.
I'd say most shots because we were working in text to video, uh, for the most part came out fine. There were a few, like, you know, I mean, the the baby with the glasses. This one, uh, was what?
Seven times I wanna say. And then there was, like, some exploratory stuff like Default Face, that was her name, um, when, you know, she's saying I have an idea.
I was running just a number of those just for, um, you know, uh, just for exploratory purposes. Um, I will say, yes, we did run across a few content flagged issues, but that was more on the C Dance Interestingly, the talk show host, that that was actually giving us a pretty significant amount of trouble.
Um, Yeah. A lot of content flags there. I'm not sure exactly what was triggering it.
So we ended up kind of going with more of a, well, frankly, Jimmy Fallon, uh, type look. And, uh, yeah, that that that went through just fine. Another issue that I kept running into is with our, like, TikTok influencer who kept coming up with, like, these crazy robotic eyes.
Uh, I did figure out at some point or another, it was, uh, in the prompt, the ring light, and, uh, that was causing the issue there. Now for our recurring characters, and really there are only two in here. There is, of course, Eli, the, you know, uh, tech CEO and, uh, default face.
Um, and, you know, Eli pretty much ran flawlessly throughout, you know, any generation. Uh, just basically popped him in, uh, said use omni reference and, uh, whatever he was doing and, yeah, he came out he came out fine. Default face, uh, she was a bit of a pain.
We were running into a lot of issues in particular when, you know, she's supposed to appear on the phone as being also the person holding the phone or her appearing as two characters in actually, this whole part was cut, but this was supposed to be POV shots from the AR glasses. Interestingly, one of the solves there, and actually Claude came up with this one, was to use kind of more of an icon, uh, shape for her face, uh, as the reference image, and that did seem to work a little bit better.
And then with some prompt tweaks, uh, we actually got her to, you know, become full screen. Actually, you can see the full screen version of her here in another, uh, generation with ring lights, uh, where we got this as well.
So overall, just don't use ring lights. Don't don't call out ring lights in your prompt. When we hit the edit, I mean, that's probably where the project context really comes into play.
Again, because it knows everything, like, the idea of creating a YouTube thumbnail for this guy, um, you know, I could just send Claude off to do that because it it knows who this guy is, who the scene is, uh, and can generate up a pretty good, like, YouTube style thumbnail. There was a lot of this kind of stuff, like these stills that I kind of discover that I wanted to lay in during the edit that I I don't necessarily I'm not saying that it would have been hard to do or anything.
It was just much easier to simply go into Claude and say, like, hey. I need this shot of Eli doing this and then just get back to the edit, um, and then, you know, come back in five, six minutes and then pick up the shot. Speaking of discovering things and editorial, one thing that I do like about Martini, by the way, is, like, they have like this I mean, is, I will admit, a fairly basic, uh, nonlinear editor down here, um, that you can use.
The way that I ended up using this was as I was generating, when I would find things that I liked, uh, I would just quickly pop them onto the timeline. So, um, again, wasn't doing a lot of editing per se here, but that is something that, you know, you you can sort of play around with doing a rough cut here. It's a little bit you know, uh, obviously, it's got all of sort of your basic tools that you need in here, but, I mean, obviously, it's gonna lack a lot of the finesse of a full blown editor.
Um, but what I did find was very handy about this is that you could actually export to an XML. So that is probably something that I'll do for another project, um, you know, as I can build it literally here, um, you know, all of my clips and then just XML it and have it send it right over to my editor.
So that could actually be pretty handy. Actually, I could probably even get, like, Claude to do that, couldn't I? That's what I mean about, like, kinda building it yourself is that little ideas like that are going to occur to you.
Uh, honestly, you know, from here, it was basically more or less just, you know, a standard edit. Um, some transitions between the shots, uh, and just kind of figuring out, uh, you know, what to cut and what to keep in terms of the storyline. Other than that, audio, which I did futz around with a little bit more on this particular project, uh, track six here is our audio.
It is kind of all nested together, I should say, Uh, what I did here was I just exported all of the dialogue and then took it over to Adobe Podcast Enhance, uh, ran it through that just to kinda get rid of that sort of AI, uh, tininess, I guess.
Um, and then from there, it was run through this is this Brainworks, um, audio cleanup, uh, freebie that I found, uh, on Plugin Alliance.
I don't if it's still free, but you can go check it out. You know, it's it's just basically a high and low band pass, uh, filter. And then before that, there is a multiband compressor on here that, uh, I just set to the, uh, preset of broadcast.
The music was generated up in Suno, just kind of an ethereal glassy kind of ambient sound. Uh, and then I ended up using another free plugin called Valhalla Supermassive.
Uh, this thing is actually sort of a lot of fun. I'm not gonna go into a whole thing on, uh, super massive, but, uh, basically, it's like giant, uh, swelling reverbs and delays. Uh, very handy when you have like these ambient type tracks.
I am 100% not an audio expert. I will say this though, adding in a multiband compressor and, um, just like a little bit of a clean sweep or like a low high band pass filter will get you a pretty long way though.
So, uh, definitely, uh, try to incorporate those at least into your next project. Now in terms of overall time and cost, I think it was about 53 shots in total, give or take, with some title cards, uh, for the final output.
We ran about, uh, a 160 generations. Although, I will say that I was being pretty liberal with my generations.
I probably could have done it in less. Additionally, there were a couple of, uh, like, abandoned concepts including, like, AI influencers, uh, that was, uh, we kinda set that part up in the, um, they get their own social network beat, but, yeah, I just kinda dropped it.
There are also a number of, like, straight up, like, CDNs. I can't use this, uh, outputs like, you know, Jake Gyllenhaal here, but CDNs really loves Jake Gyllenhaal. The other one that kinda made me laugh was a tech reviewer in the style of MKBHD, and, I mean, it kinda gave us, uh, well, I mean, essentially, that is like Tebu Marquez Brownlee.
So, yeah, obviously, I did not use this one. So that kinda wraps up the production walk through. I will say that, you know, again, going back to Martini, there are some, like, really cool features in here that kind of have me thinking, like, what else we can do via MCPs.
They have, this step into set feature which basically takes any image and turns it into a Gaussian splat where you can, you know, sort of, like, move around within your image space.
The red outline is your character. So you could do some stuff with that too. We went over this in another video but this allows you to, yeah, change your camera angle to whatever you want.
And then when you find something that you like, essentially you you capture it and then it'll rerun on banana things. To be fair, I'm I'm not sure how, you know, a clawed MCP would handle this. Probably not well.
But, you know, the the idea is there. They also have a multiplayer feature in which, you know, two users can be, you know, in the same canvas.
Again, I don't think that MCPs are quite there yet, but it is an interesting idea to think that it won't be long before they are. So yeah. I mean, overall, a lot of really interesting new ideas and workflows are going to be coming our way.
Thanks to MCPs. Uh, again, I thank Martini for sponsoring today's video and essentially the movie. Um, Yeah.
So definitely, they are very filmmaker forward in terms of their approach. So high recommend martini.film. Other than that, I mean, I guess I'm gonna go back to seeing what other massive things have dropped over the last day or so.
And I'll be back with a, I guess, like news report coming up pretty soon. It feels like I'm behind on the news. Feels like I'm always behind on the news.
As always, I thank you for watching. My name is Tim.
The Hook
The bait, then the rug-pull.
The title is the argument: stop subscribing, start building. Tim opens by explicitly rejecting the node-graph paradigm that dominates AI video discourse, then spends 19 minutes proving that a folder of markdown files and a Claude MCP connection can outperform anything you could rent.
Named ideas worth stealing.
02:32model
AI Film Production Office
- Story breakdown
- Style guide
- Character reference grids
- Scene boards (2x2 grid)
- Video/seedance prompts
- Production brief
- Production tracker
A self-contained folder of markdown files on your desktop that Claude in code mode reads as its operating context for an entire film project.
Steal forAny multi-session creative project where continuity of context matters
10:36concept
Handoff Doc Pattern
When the Claude session context window approaches red, ask Claude to write a handoff doc. Open a new session, point at it, and you restart at zero context debt with no information loss.
Steal forAny long-running Claude Code session that spans multiple work sessions
15:40list
The Audio Cleanup Chain
- Adobe Podcast Enhance (AI detin)
- Brainworx Cleansweep high/low pass (free)
- Multiband compressor at broadcast preset
- Suno for ambient music generation
- Valhalla Supermassive for reverb/delay (free)
A five-step chain for making AI-generated voice dialogue sound broadcast-ready.
Steal forAny AI film, AI voiceover, or synthetic dialogue project
How they asked for the click.
VERBAL ASK
18:56newsletter
“I'll have all of this available for you in this week's newsletter. I'll just have the whole folder. You can download it.”
Mentioned twice — once mid-video and once in wrap-up. Soft ask, no urgency pressure. The free template is a strong incentive.
MENTIONED ON CAMERA
FROM THE DESCRIPTION
Visual structure at a glance.
