Google Stitch, Remotion, and Blender MCP signal a shift to command-line creative work
Nate B Jones, host of AI News & Strategy Daily, covers three recent AI creative tool releases and argues they represent a fundamental restructuring of how design, video, and 3D work gets done.
Summary
Nate B Jones examines three releases — Google Stitch's major update, the Remotion video framework, and Blender MCP — and argues they collectively signal that design is moving to the command line in the same way software development did. He contends that the traditional product-design-engineering triangle was always dysfunctional in practice, and that these tools don't eliminate designers but rather abstract away the operational work that was never the creative point. A key structural claim is that MCP is becoming the universal connector standard for AI tools, and that any product not yet available as an MCP server is falling behind. He also highlights Google's design Markdown file export from Stitch as an underreported feature that enables coding agents to read and build against a design system directly, eliminating the handoff layer that caused so many failures in traditional workflows.
Key Takeaways
FULL TRANSCRIPT
The Three Releases That Matter
Nate B Jones: The world is getting overloaded in AI output. But what makes it useful? When does the creative process assisted by AI become something we actually care about? I think three big releases dropped in the last week or so that we need to pay attention to as builders. This is not about taking jobs away from designers. It's actually about rethinking how we do design in the age of AI. I want to get into detail, talk about each release, talk about what it does for you, and talk about where we can go with it to build appropriately.
First, Google updated Stitch. It's a free tool where you describe an app in words and it draws the screens for you. It's not wireframes. It's actually a finished-looking UI with multiple pages at once, and you can talk about it out loud and it's going to design it in real time. Yes, you can actually do voice-to-design now. It's connected to coding tools so you don't have to ask yourself whether this is buildable — it's actually just going to read the design and build against it appropriately. And there isn't going to be any more exporting to Figma. That is why Figma stock tanked, because fundamentally when you have based your entire value proposition around this era idea that all of our job families are separate, you get into trouble in the AI era world where job roles are blurring and we're all moving toward the command line.
Second big release to pay attention to is Remotion. It's a video framework that's been growing quietly since January that has crossed over 150,000 installs as a Claude Code skill. All you have to do to use Remotion is describe a video — say, "Make a 30-second product demo of the product we've been building together. Please show these three features and show a call to action." And the agent will write the code that renders it into an actual MP4 video.
The third one is Blender MCP, and it takes a bit of a different approach. It's at about 17,000 GitHub stars. What Blender MCP does is it actually builds a 3D scene in front of you. So you can say, "I want to create a beach scene for my product and this is what I want it to look like," and Blender will build the three-dimensional version for you right in front of you. People who've never touched Blender are doing things like generating room walkthroughs or architectural assets for their general contractors. There are people who are using Blender for what it was originally designed for — video game graphics — but they're creating immersive walkthrough generative worlds. We're just at the beginning of figuring out what all this looks like, but I think we know enough now to start talking about the idea that design is shifting to the command line and to talk about the implications of that.
MCP as the Universal Connector
Nate B Jones: But first, let's talk through each of these pipelines and understand where we're going. The one commonality we have across all three is that they all depend on the universal standard for AI connectors: the MCP. We are seeing again and again that MCP is becoming the USB plug for AI, where you can just take a tool and say, "I want to make it available at the command line." It becomes an MCP server and you can just begin to work with it. That's how Remotion went from a little website that people might go to and do stuff on, to a fast-growing GitHub community. They made it an MCP and they made it available to people to use as a skill inside the terminal. That is becoming the growth hack of 2026. If you have a product, ask yourself: why isn't it an MCP? It should be an MCP. If it's not an MCP, you've got problems.
Why the Command Line Matters More Than the Tools
Nate B Jones: So why does the command line matter so much? I think this is a much bigger deal than the tool and the 2026 use case. Most of the time we like to talk about these products as if they're big flashy releases, and we miss some of the larger changes that are happening.
What I am seeing is that the 2010s-era triangle of product, engineering, and design is starting to blur together. We've talked about that. But the thing we haven't talked about — and I've seen it discussed only rarely, because I think it's a nuanced topic that tends to run away into the job loss debate — is that the product-design-engineering triangle often did not work well in practice. I say that as someone who lived it. Of the product-design-engineering triangles I was a part of, the vast majority were not seamless. We had issues where engineering was the brake and said you couldn't build this. We had issues where design was slow and said, "I need some more time to push pixels." We had issues where product — namely me — liked to go fast and bias for action, and everyone else was worried about that.
The point is this: in all of those cases, we were gated by the ability to sequentially get pixels on the page and then later find out if it was buildable. The best circumstances I found, the best teamwork angles I found in that 2010s world, was when you could co-locate. The designer would sit with the engineer with product. We would all huddle around the screen, mess with stuff together, and see what worked. The beauty of that is that even though it was rare in the 2010s, this kind of tool with the command line is making that more possible for everybody today.
In the past, you had to depend on the right people being in the room and having some degree of role blurring, even though we didn't talk about it that way. The designer needed to understand a bit of how JavaScript worked. The engineer needed to understand a little bit about how you put polish on the front end. Product needed to be able to sit there and put the two together and look at the offer and the customer. Well, now at the command line, you can actually invoke something that is by definition buildable — because that's how this works. If it's not buildable, Claude Code can't touch it. And you can start to bring in multiple options for design iteration very quickly. That was always one of the things we struggled with: the idea that design takes time. Designing something in a polished way takes more time. If you want to get through a variety of iterations, design is something you can now do from the command line and quickly get through them.
Now, if you want extraordinary polish, you will still need to have a designer hand-finish this stuff. But so often right now, what the designer needs to do is just put together from a design library the next release. If you wonder how designers at Anthropic are doing this, how designers at OpenAI are doing this — they are in the code. They are working from the command line. They are working to code something up from the library of components they have and putting it together in a way that's very rapid so they don't slow the team down. That is command-line design thinking, and it applies to the video space as much as it applies to creative.
I don't see high-class designers leaving the room here. The "AI replaces designers" narrative is cheap and it's distracting. I see high-class designers — and I've talked to them — sitting at the command line. I see them sitting with engineering and product, pushing through ideas into code in real time. And instead of the whole table wondering whether the design is buildable, it's just right there. It's prototyped. It's exportable. It goes right to the codebase.
I'm really bullish on this, and it's not because it replaces designers. It's because it abstracts away the part of design that was never the point. The point was always the experience for the customer — the feeling, the flow, the moment when a user thinks, "This is what I needed." That judgment you can't automate away. It gets amplified, because now the person who has it can iterate at the speed of language instead of at the speed of Figma.
What shifted in the last few weeks is the beginning of that shift, and it changes the cost structure of creative work in the same way vibe coding changed the cost structure of software.
And if you're not a designer, this is even more relevant to you, because design has always been a scarce commodity. It's been hard to get design expertise. People often do without. Now you can get — maybe not excellent human-designer quality — but at least more designer quality than you could get without design in the room. Which is great.
Google Stitch: Voice-to-Design and the Markdown File
Nate B Jones: So let's walk through these pipelines. First, Stitch. Google originally launched Stitch at IO 2025 as a very quiet labs experiment. It was a text-to-UI generator running on Gemini, producing one screen at a time — useful but limited, and a lot of people thought so. It's the kind of thing you see in the demo and sort of forget about.
The March update is not that. Google redesigned the entire product around what they're calling vibe design — the design equivalent of vibe coding. Instead of opening a blank canvas and placing your components, all you do is describe your business objective, a user feeling, or a product concept in natural language. Stitch will then generate multiple high-fidelity UI directions simultaneously, and it will pull together up to five screens at once on an infinite canvas where you can drop in images, text, code snippets, and competitor screenshots as well.
Now, this was the killer app. And it turns out this is a nice parable for how difficult it is for even large companies to find product-market fit today. Google debuted this back at IO. Figma shares didn't really move. It took actually getting the product right — moving it to the command line, making it simple to use, expanding the blast radius of what you can do with design — for Figma shares to respond in the markets. Because now the Wall Street guys who are not designers see the obvious connection to Figma.
So here's what Stitch actually does. You can open a canvas and just talk to it, because there's voice mode. If that isn't specific enough, let me give you a prompt. Say: "Design a landing page for an AI writing assistant. The hero section has a headline and CTA. There's a features section with benefit cards. Please put in a pricing section — make it two tiers." And Stitch will just generate complete designs — not wireframes, by the way. High-fidelity screens with real typography, real color palettes, good spacing, good component hierarchy.
And it gets more interesting from there. The design agent doesn't just generate and forget. It actually holds the entire project context — all the screens, all the decisions, every piece of reference material you've dropped onto the canvas — and it reasons across all of it when you ask for changes or edits. So if you want three different navigation layouts, you can ask and it will see the original and know how to adjust. If you want to see the whole app in a dark palette, just say so. The agent will track the project's evolution the way a senior designer tracks a design system holistically, not just screen by screen.
The agent also lets you branch and compare design directions, which is one of the things I'm most excited about in the age of AI, because we were always gated in our optionality by the time it took to make designs. Not anymore. You can now version-control, do creative exploration, pursue five different approaches simultaneously, evaluate them side by side, and merge the pieces you like. And you get instant prototyping that converts any static design into a clickable flow. Just hit play and you can walk the user journey. Stitch even auto-generates the logical next screen based on user interactions — generative UI. It predicts what comes after a button press by reasoning about the product structure and the way you've talked about it.
And this is what really affected Figma. It's free. It gives you 350 generations a month for nothing. Zero cents. That's what you get when you're Google and all of AI is assigned to you.
But the feature that's most interesting — and that most coverage missed — is the design Markdown file. We all know Markdown files are critical in AI. What the design Markdown export from Stitch gives you is a full agent-readable file that captures a project's design system as you've evolved it in Stitch. So you not only export the code, but you get a durable record of colors, typography, spacing rules, and component patterns. And you can extract them from any URL you want. You can say, "Have a look at this URL. Please pull the design Markdown," and you can get a sense of what that site you admire actually looks like from a designer's perspective.
People who aren't thinking carefully will use this to just copy a site they like. People who are thinking carefully will use it as inspiration and push it farther, because they have a sense of where they want to go. Clarity of intent is a differentiator in the age of AI.
And Google isn't locking this into their ecosystem. As I said at the top, this is MCP-readable, which means you can grab it from Claude Code, you can grab it from ChatGPT, and yes, you can grab it from Google's own tooling. Your coding agent of choice can read your design system while it builds. Google also shipped official Claude Code skills — which is itself a sign of how dominant Claude is, when Google has to ship Claude Code skills with their own launch. But they shipped installable playbooks that teach the agent how to work with Stitch outputs.
The pipeline works right now. You can design to a business objective today. Stitch will generate the UI. The coding agent will read the design Markdown file and just start building. There's no Figma export. There's no handoff document. There's no "the developer got the design document wrong" — which I've heard a lot. It simplifies the number of handoffs in the system, which is one of the big trends we're seeing in AI. You start to delete these handoffs as agents get more capable. That's what Google did here.
Now, here's the rough part — and I want to be honest. The design quality is not yet at a point where most senior designers would sign off on it and say, "This is great, I want to show it as-is to a client." But it's better than a prototype, and this is the dumbest and worst it will ever be. So I would look at it as: A, is it a way to get an MVP into market if I can't work with a designer? B, is it a way for a designer to work faster and be able to polish more effectively because we've tested a bunch of different directions? And C, can we make it this good without pre-built components? Can we start to bring our own design system into play and bring extra polish to the process?
You should not view this the way Google probably markets it — as a magic designer in a box. View it more as a magic junior designer in a box that does a faster, better job at prototyping, but isn't going to be perfect, especially if you're not clear on intent. And by the way, that is valuable, especially if it's free. The phase we're talking about — design exploration, MVP development — used to cost thousands of dollars just to get through before you got to the polishing. Now it's free. This is what we mean when we say the cost of software is falling to zero.
Remotion: Video as Code
Nate B Jones: Number two — now we're going from design to video. Video is becoming code.
In January 2025, Remotion — which is a React framework that treats video as code — launched an agent skill for Claude Code. Within 48 hours, the demo video they made had over 6 million views. Eight weeks later, the skill sits at 150,000 installs or more on skills.sh, which is an open directory for AI agent skills. It's the number one skill not made by a company named Vercel, Anthropic, or Microsoft.
What it does is really simple. You install the Remotion skill into Claude Code with a single terminal command. Then you describe the video in plain English, and Claude writes the React components — the code components that define every single frame of the video. It defines the text animations, the motion graphics, the data visualizations, the captions, the transitions. Remotion renders all of those components into a video file — MP4. The whole loop runs locally on your machine for free, other than the Claude Code subscription.
I want to be really clear here, because there's a ton of misunderstanding. This is not the same thing as AI-generated video. Tools like Sora and Runway generate pixels from prompts and can be impressive but also inconsistent. They're not easily editable. They can be expensive to iterate on. Remotion generates code that renders video. Every element is a React component that you can modify, version-control, and parameterize. If you change one variable and re-render 100 localized versions, that's fine — free, other than your Claude Code subscription. If you update the data source and every chart in the video updates automatically, again, it's fine. This is programmable video, which is in some ways more powerful than generative video because it's easier to edit and adjust. That distinction matters enormously for anyone producing video content at scale.
Remotion collapses the video production last mile. The person who knows what the video should communicate just says what it should say directly. The agent handles the timeline. And because the output is code, not a locked export, anyone on the team can modify it without re-entering the editing tool.
One creator, Sabrina.dev, documented her complete pipeline. She was able to generate a promotional video from one prompt, having Claude browse the web to take real screenshots of GitHub repositories for inclusion in the video, then added her headshot and background music from follow-up prompts. The whole workflow was the command line. No Premiere, no After Effects, no timeline editor. She just got it done.
I want you to look at the falling cost of compute and the value of design here. This is freeing our imaginations. That's the through line across these first two releases — through Stitch and through Remotion. We can imagine and therefore we can create cheaply. Before, if you wanted to make a polished clip on a MacBook Pro, it was heavy CPU usage, it took a while, it wasn't easy to edit, and sometimes it would take an hour or two hours to render. Complex animations could break down fast.
Now, I want to be honest: if you're doing extremely complex animations in Claude Code with Remotion, you also have issues. If you have multiple overlapping elements, fancy transitions, or very intricate timing, you may not get a perfect render. The sweet spot right now seems to be clean motion graphics — text animations, data visualizations, product demos, terminal recordings. Honestly, that's a ton of product space to operate in, and it's going to do very well. Think of this as shorthand for the kind of content that drives so much of a digital business: product demos, feature announcements, data stories, social clips. All of those can be done from the command line in Remotion, with no incremental cost, because Remotion is giving Claude Code the skill it needs to write code that generates that video. The gating factor on your video is the quality of your input.
Blender MCP: 3D Scenes from Natural Language
Nate B Jones: Number three — also a creative tool released in just the last few weeks — Blender MCP. Three-dimensional gets a chat window.
Blender is, think of it as a nuclear reactor for creative tools. You get professional-grade 3D modeling, animation, rendering, and simulation all in one thing. It's used in feature films, game studios, and architectural visualization. It is by universal consensus one of the most complex pieces of software a human can attempt to learn. The interface has roughly 1,500 operators — I'm not kidding — a Python API that exposes almost every internal function, and a learning curve that is measured for some people in years.
Blender MCP simplifies all that. All you do is type a natural language description like, "Create a beach scene with palm trees and sunset lighting," and you watch the 3D environment assemble itself in real time. By the power of MCP — because the function of MCP is to enable an LLM to creatively manipulate tool sets — Claude manipulates the tool sets of Blender for you without that learning curve. The beach scene assembles itself. Objects appear, materials get applied, lighting adjusts, and all of it is controlled and edited by AI writing and executing against Blender's Python API through a socket-based bridge.
The repo is obviously popular — over 17,000 GitHub stars, 1,500-plus forks as people build off of it. It integrates with Polyhaven for free high-quality assets like HDRIs, textures, and 3D models, and with Sketchfab if you want an even wider library. You can also connect it to Hyper 3D for text-to-3D model generation, meaning you can describe a character or object, have it generated, and then drop it into your scene without leaving the conversation. I can describe a character like an open-claw crab and drop it into the beach scene without leaving the conversation.
The community response has predictably been massive. Users have described it as a game-changer, and I think they're underselling it. When you have deterministic software that is as complex as Blender, MCP plus the command line is a massive simplifier. Which is why I like to say over and over again: don't be afraid of the terminal. It's actually simpler than most of your other options, because you can now describe in natural language where the camera should be, what the setting should be, what you want to appear in a three-dimensional scene. And instead of spending months or years learning Blender, you just have it made for you in a few seconds.
If you think that's not relevant to you, then your job probably isn't architecture previsualization, gaming prototyping, product mockups, or any workflow where the bottleneck is getting from concept to proof of concept very fast. There are a lot of cases where we have three-dimensional visualization needs and it has been difficult to get to that level of fidelity quickly. Again, we see a collapse in the cost of creativity here. It's not that you're not going to be creative. It's that you're more free to be creative because it's faster to get from idea to voice to the thing being there. It's as close as the breath now.
Context Loss, Creative Primitives, and Scheduled Pipelines
Nate B Jones: There's a larger thing I've been thinking about. When we were working through the classic product triangle in the 2010s, so much of what we wrestled with was context loss. So much of what made those dynamics work when they worked was all of us together at one computer talking, with zero context loss. You have zero context loss in all three of these tools, because it's so easy to go from describing what you want to getting something back — something visual.
And I want to call out that these are primitives. You can think of these visuals as components in your workflow. One of the big skills of 2026 is taking these individual components and smartly integrating them into a larger workflow.
Let me give you a specific example — hot off the presses. On March 20th, Noah's Way announced cloud-based scheduled tasks for Claude Code. You set a repo, you schedule it, and you prompt. Claude then runs on a cadence in its cloud infrastructure, and your laptop can be closed and your terminal can be off, and the work still gets done.
The example posted was developer-facing: you can schedule a daily job that reviews all the PRs shipped since yesterday and updates the docs. That's useful. But think about it in the context of what we've been talking about today — creative pipelines.
You can schedule a weekly job that pulls all of the shipping you've done for your product this week and generates a Remotion video for customers that summarizes the key points with motion graphics, renders it in the right aspect ratios, and queues it up ready to upload — so you can have an attractive launch video for customers every single week. You can schedule a daily job that checks your analytics dashboard and generates a data visualization video showing yesterday's metrics and sends it to the team Slack channel. You can schedule a job that monitors your product's change log and generates updated UI screenshots via Stitch and rebuilds the marketing site's feature gallery every time a new version drops.
You think that one's not worth it? I've had so many conversations with marketing and documentation teams as product, just to explain: this thing launched, you've got to remember to update the site, we need to update the docs, is everything up to date? These kinds of things are now trivial if you hook up your creative primitives to your scheduling primitives.
That's what I want you to start thinking in terms of — these are Lego building blocks for workflows that you care about. What gates it now is you. What do you care about? What are you passionate about? How does your workflow work well when you can have the power of these tools at the command line? Because the entire pipeline from idea to artifact — from idea to creative — is now describable in English, executed on by agents, and scheduled to repeat autonomously.
And I keep emphasizing: this is not expensive. You can do the Remotion skill for the cost of Claude Code. Google is giving you free iterations on Stitch. It's cheap now.
The Floor Dropped, But the Ceiling Didn't Move
Nate B Jones: I've been watching creative tools get disrupted for 20 years. Photoshop by Canva. After Effects by simpler motion tools. Sketch by Figma. Figma by AI. The pattern is always the same. A new tool lowers the floor of who can produce good-enough work, incumbents start to panic, and then reality sets in. The floor dropped, but the ceiling didn't move. More of us got in the door as creatives, but excellence still matters.
The difference now is how deep the floor dropped. When it's basically free to generate designs, the floor dropped into the basement — anybody can now generate designs. And the distance between good enough and extraordinarily excellent is now something we have to talk about as a different skill for designers, because before, so much of design was: can you execute the thing? Now so much of design is: can I polish the thing? Can I look at the thing and say, "This is specifically what's not good enough, and this is the edit I want to make"? Can I do it quickly and consistently and apply that skill — and potentially even teach that skill so others can learn it? Can I make the thing I know to be right or wrong, my sniff test, something I can act on very fast as I look through all of these designs, so that I am not the bottleneck for AI and the design creativity we can now have?
The people who win here are not the people who individually master these tools. That's sort of like being the person who masters the 1,500 commands of Blender and then MCP comes along and disrupts you. It's more like you need to be the person who can articulate precisely enough how you want something to be expressed visually, so it conveys the experience to the customer that actually matters. You're going to have to explain it, then edit it, then possibly polish it at the end.
Stitch is better than prototyping, but probably not polished all the way through. Remotion's clips can take a little while to run and don't do super sophisticated animations. Blender MCP does a pretty good job with basic scenes, but if you're doing very complex organic modeling, it's not going to work well. These are all present limitations as of March 2026. I expect models to continue to get better. But the thing to remember throughout is that we humans have an extraordinary ability to care for detail. It is going to be up to us to care for detail here so that we are actually able to express our taste and get what we want done out of our new superpowers.
And they are superpowers. For so many of us — including me, I am not a designer — this feels like I have Superman's cape. I can do so much.
Design Is Following Development to the Command Line
Nate B Jones: Ultimately, what this last several weeks of creative developments has freed us from is so much of the time we spent in creative tools that wasn't creative. It was operational. It was moving layers. It was adjusting keyframes. It was exporting formats. It was rebuilding the same layout for the third time because the stakeholder changed one word in the headline. The actual creative decisions were few and far between. That density of creative decisions can now go up. It's like: this is the decision I want — now go do it.
One of the things I think is really interesting is that design is following development here. Developers knew for a long time that the command line was the right interface to get execution done at scale. Creative work has arrived at the same insight. We are now at a point where you can generate, iterate, convert, and do all of the scheduling and executing work for design from the command line. And it doesn't matter if you were a developer before. As long as you have the ability to describe it, you can do it.
This is the most widespread, everybody-can-get-in change that I've seen in AI in a while, because design has always been gated — gated by our ability to imagine and believe that we can actually move pixels around and create beautiful things. We've seen some of that start to shift over the last few years as Figma has made some of these components easier to adjust. But this just blew the doors off. We are now in a place where anybody can do reasonably good prototyping design. And I think that frees us all to be so much more creative and focus on the value of decisioning.
And yes, I'll say it again: high-quality designers are still needed to offer their great judgment and help us take these tools — these new paintbrushes in the palette box — and actually use them to create incredible designs for customers, for ourselves, for our families. Have fun. Use your new powers wisely.