The promise of “vibe coding” is seductive. You simply describe what you want, as you might in regular conversation, and then the AI agent magically makes it appear.
But if you’ve spent more than a few moments with any of the vibe coding tools, be it “beginner-friendly” options like Lovable to heavy-duty workhorses like Claude Code or Codex, the magic likely wore off really quickly.
The distance between prototyping an idea and building a production-ready tool you can use on a regular basis is still frustratingly vast. Having technical skills, as well as an understanding of how to write precise product feature descriptions, heavily impacts your likely level of frustration.
And when it comes to “vibe animating,” it’s so far really been about how to make ridiculous slop full of errors, hallucinations, and nonsensical garbage.
Then came Google’s updates to their animation tool Flow (released May 19, 2026).
And I’m genuinely excited. With some caveats.
The Leap Forward: Google Flow’s new integrated agent lets you use natural language to build custom, multi-step generative workflows (like an end-to-end animation studio) without needing to learn complex node-based logic.
The Proof: Within 90 minutes of “vibe-coding” and iterating, I built a (mostly) functional tool that scripts, generates assets, storyboards, and animates a scene.
The Roadblocks: Memory is currently limited (no persistent file storage yet), version control requires manual “remixing” and saving, and the AI will break unrelated code while trying to be helpful. And the usual bugs that can be tricky to track down.
The Verdict: It’s not magic (yet). Technical product knowledge still helps a ton to structure your requirements. But the potential is high for a unified, customizable AI production ecosystem.
If you aren’t familiar, Google Flow is a creative AI studio designed for filmmakers and media creators. Using the Google image and video models, you can create all sorts of assets for videos.
Historically, if you wanted to compile these assets into a video, you need additional tools to make it feel like more than slop. For multi-step generative workflows, you likely turned to ComfyUI, Runway, or your own custom-coded set up. Which means you’ve also probably been stymied by the steep learning curve of their node-based, flowchart-meets-logic tooling.
With its latest updates, Google Flow has added a powerful agentic layer that lets you bypass that learning curve entirely. Instead of wiring nodes together manually, you use natural language to vibe code custom tools. Flow’s integrated Gemini agent handles the logic, building bespoke workflows that can perform multiple actions simultaneously while maintaining visual consistency.
Prior to the update, I’ve already been a heavy user of Google Flow. This is just one of many articles and workshops where I share how I work with it.
Google also added an agentic option for interacting with the main image and animation tools, powered by Gemini.
The standard Flow animation interface now includes a button to use the Agent workflow.
Without the agent, I entered this prompt into the plain old animation prompt box.
I want to create a very short, emotional, nonverbal, 3D CGI Pixar-style short based on this exact series of clips: 1. A decorative vase sits on a table in a living room. There is nothing on the table, no flowers, no water. A tennis ball rolls under the table. 2. A bouncy, adorable Great Pyrenees dog runs straight at the camera in slow motion in the same room/setting as the vase. 3. The dog sits with the tennis ball in her mouth, looking very sad and guilty. Next to her is the broken vase. The dog should not be on the vase or on the broken pieces. Keep them safely separated, but clearly show both the dog and the broken vase.
And got this
Not bad. It’s not slop, per se, but also I personally wouldn’t publish it. It’s simple but has issues. The dog is running away from the table. The living room setting changes by the third shot.
Then I gave the same prompt to the agent, which created a multi-step process.
First it provided a storyboard with asset descriptions, including the dog, the living room, and the vase. As well as descriptions for each shot of the animation.
I approved it and it created a storyboard image.
After I told it to proceed, it asked for permission to generate the video, including how many credits it would use.
It made this video. One generation.
This test went generally well. I really like that I can click on the video and see the prompt that the Agent used. It helps me continue to understand how to prompt better.
But it’s not all sunshine and roses! Another test of a similar storyboard went very poorly. The agent insisted it had saved the storyboard. But due to how Flow is set up, there’s no mechanism or location for saving anything other than the images and videos it generates. It’s not tied to my Google Drive or my harddrive.
Yet the agent insisted the storyboard was in the chat panel (it wasn’t) and that it had been saved (it wasn’t).
So I need to test a whole lot more to see how reliable this is! And try out longer, more complicated story ideas.
I also suspect that as the newly released Tools mature, it will help solve these issues as well.
Now for the fun stuff. The tools. I started by simply playing with the tools that Google published with the update, like.
Shot Explorer: Explore different angles of the same scene
Transition Machine: Create seamless transitions between videos
Storyboard Studio: Write a script, create visual assets, and then storyboard
As I started to get a feel for what kinds of tools were available, I started brainstorming ideas.
Storyboard Studio was the tool where I kept thinking “What if I could...?” So I started there.
I clicked Remix (the option to save a new copy and begin modifying the tool). And then I made a big feature request.
Add a fourth stage after storyboard, which is Animate. Ideally, this would pre-populate the prompt from the script assets and storyboard with what needs to happen in the image or in the animation. I can edit it. Then there’s a button to submit individual images for animation, or I can submit the whole scene for animation. When it’s finished animating, the animation segment appears in that spot. I can edit the clip and regenerate if I want.
Guess what? It worked. And I was hooked.
I continued requesting features. These are some of my actual prompts, typos and all.
Auto-save: Is there a way to load or save a story so I can come back to it later and continue working on it? I was using this to play with a concept, updated the tool, and then came back to it, but I’ve lost everything I put into it. Do we have an option for how to save or load it so I can pick up where I left off?
Use existing assets: In the scene image creation stuff, I want to be able to load in a previously existing character or images (for objects and locations and props too). I often have an existing reference that I need to load in.
Explore new angles and manage continuity: I also want the shot explorer option within the editing of storyboards. So I can create new angles of the image that are things like zoom in/out, pan left, right, top down, low angle, etc And then we also have a continuity issue. If you look at the images generated here, the tower changes, the placement changes, etc. So at times, one of the earlier storyboards is going to become the reference for later storyboards. How can we best handle that?
After about 15 feature requests, I was getting more bugs than successful runs.
I stepped back, made a list of the features I wanted, decided on an order to implement, and started a fresh remix of the original tool. In other words, I put on my product manager hat.
In version 2, I focused on updating the Assets stage first. The main features focused on being able to load a character reference from existing files, editing the generated images via a chat box, and saving and restoring progress.
In version 3, I focused on the Storyboard stage. The features included inheriting generation features from the Assets stage, like loading reference files, as well as creating continuity references. So when the first image included a particular setting and character, the second image would also include the same.
In version 4, I again added the Animation stage, making sure I could see the entire prompt that was being submitted for generation, and make edits. I also added a download all feature at a target resolution, rather than having to click on each item alone.
By version 5, I was focusing on bug fixing and making sure the full pipeline worked end to end. The performance is still quite fussy, so I get an error of some sort about 1 out of 4 times. It doesn’t always follow the style info (as you can see in the screenshots). It doesn’t always automatically populate information when I move to the next stage. So interesting as a concept, but not yet usable without additional work.
Yet all of this took 90 minutes.
Then I turned into a super crazy nerd.
While in editing mode in Flow, you can actually view the underlying code for the tool you’re building. So I asked Claude (Anthropic’s model) to review the code via Claude Code with Playwright to access my Chrome browser. The Claude Chrome plug-in worked equally well for this.
Claude provided feedback and prompts I could provide to the agent in Flow.
Now fully obsessed, I tried creating a few other tools to really see where the limits are.
First a video editor. I edit in DaVinci Resolve and via code with HyperFrames or Remotion. I’ve long wanted to be able to edit within Google Flow. They do provide tools, but it’s like cutting your filet mignon with a hammer.
I need a tool where I can select animated clips, edit their length (Select parts of it), add text to parts of the video for timed portions, and then export the whole thing as a single file.
I added tools to crop the videos, transition between videos, and download the full viewable.
The code is using ffmpeg to render the videos. It’s so-so on quality so I’m not abandoning my current flow anytime soon. But really cool to see the possibilities.
The ideal tool would allow me to have the context of the show, the world, the characters, the episodes, the voice. All the things I need to reference to produce episodes. I currently manage this level of product management through custom code with Claude.
How close are we to achieving an entire one-stop ecosystem with this Google Flow update? Still far.
I asked Claude for a simple product requirement document based on my preferences for creating videos.
Then I fed the agent a hefty list of features and let it go.
The resulting app was great in theory, but had a fatal flaw.
Memory is limited.
Just like the storyboard issue I ran into with the agent, it can only save what’s hard-coded into the code. Even though the agent built structures that would save content IF there was a way to do that.
The “Add a New Episode” button was effectively useless. There’s no persistent memory storage except for the assets you’ve generated (and even that takes some work to get it to save and reference reliably). So until there’s some sort of persistent memory option, like hooking up to Google Drive, you would need to build in an option to upload a json file with the show context. And you’d have to upload that file every time.
As with all new AI tools, there are caveats:
Bugs will happen. Flow is generally pretty good at auto-fixing errors, but they definitely come up. If it doesn’t offer an automatic fix, you can usually paste the error into the chat and it will resolve it.
You need to actively manage version control. Flow isn’t currently set up to auto-save your tool iterations perfectly. You have to do it manually by going in, remixing your own tool, naming it, and keeping track of the version history yourself. You can roll back within the chat, but I’ve had mixed success with that actually working.
It doesn’t always code where you tell it to code. A classic curse of vibe coding is the AI “helpfully” fixing other places in the code that you didn’t ask it to touch. It’s the equivalent of going to the doctor for a sprained ankle, and them testing you for strep throat. (Again: save versions frequently!)
Technical and product knowledge still matters. I love that you can use natural language, but like all vibe coding, the more detailed and exact you are, the better your output. I can move quickly because I know what’s possible in software and I know how to structure requirements.
The best way to understand what Flow can and can’t do is to build something with it yourself. Reading about it only gets you so far. I’ve run into the memory walls, the version control headaches, and the moments where it does something genuinely surprising.
So jump in! And remember:
Be the Product Manager: The AI is your developer, not a mind reader. The clearer and more precise your feature requirements, the better the tool you’ll get.
Prototype, Then Rebuild: Don’t expect your first chat thread to yield the final product. Experiment to find out what’s possible, make a list of exact features, and then start a fresh tool to build the clean version.
Save Relentlessly: Manually version control your tools by remixing and renaming them frequently.
Embrace the Hacks: Don’t be afraid to use other tools (like Claude Code) to audit and refine the code your Flow agent generates.
Is it there yet? Nope. Is it a great leap forward? Yep. The limitations I hit today feel like solvable problems. That’s not something I say about most tools.
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.