Google Shipped Six Creative Tools It Didn't Design. One of Them Hands You Separate Audio Stems.
Six new tools landed in Google Flow this week. Google built none of them. A sound designer, an interior designer and two filmmakers did, and the sound one solves the problem that has made generated audio useless for real post work.
Almost every audio generator released this year hands you one file. You describe a rainy street with distant traffic and a door closing, and you get back a single mixed recording with the rain, the traffic and the door welded together forever. If the door is too loud, your options are to regenerate and hope, or to give up.
Anyone who has cut sound for a living knows why that makes the tool a toy. Post-production is a layer game. You need the rain on its own track so you can duck it under dialogue. You need the door on its own so you can nudge it four frames later when the edit changes.
Mondo Sónico, which went live in Google Flow on September 23, generates ambience, foley and contextual effects and delivers them synchronized across separate, editable tracks, with individual stems you can export. That is a sentence from Google's own announcement, and it describes a different product category from everything else in this space.
Google did not design it. A creative director named Ricardo Villavicencio and a sound designer named Sebastián Carvallo did, because after years of working together they wanted better audio for their own projects.
The part that matters more than the tool
Six tools shipped that day, all of them made by working creatives rather than by Google's product teams. Google's announcement frames it as a collaboration with leaders across architecture, sound design and digital content, and every one of the six exists because a specific person was annoyed by a specific chore.
- CaptionCast, by digital storyteller Jay Pirabakaran, transcribes, styles and animates multilingual captions in a single pass, because subtitling by hand is a tax on every video anyone posts.
- ThumbnailForge, also Pirabakaran, turns one image plus a headline into cover art formatted for each platform, because manually rendering the same thumbnail at five aspect ratios is the last twenty minutes of every upload.
- Surface, by interior designer and technologist Vojtek Morsztyn, generates materials and maps them onto 3D walls, ceilings and floors in real time, because sourcing and test-rendering materials eats the early part of every interiors project.
- CollageMotion Pro, by filmmaker and science communicator Hashem Al-Ghaili, animates mixed-media collages from text, because documentary collage sequences are compelling and brutally slow to assemble.
- SwissFlow Studio, also Al-Ghaili, converts scripts into Swiss-style motion graphics without keyframing each move by hand.
Notice what those six have in common. Not one of them is "generate a video." They are all the boring middle of somebody's actual day: the captioning, the resizing, the material tests, the collage assembly, the stem separation. The unglamorous steps between having the idea and delivering the file.
That is the signal. For three years the pitch for generative tools has been make the thing. These six are about stop doing the chore around the thing. Those are different products and the second one is worth more per hour.
How Flow Tools actually work
Here is the mechanism in plain words, because the announcement assumes you already know.
Google Flow lets you build a custom tool by describing what you need in ordinary language. You are not writing a script or wiring nodes together. You say what goes in, what should come out, and what the thing should ask you for. Flow assembles a small interface around Google's generative models, with the inputs you named, and gives it a shareable address.
Every link in this article is one of those addresses. Click any of them and the tool opens, already built, someone else's. Google says you can duplicate and remix any of them to fit your own setup.
The closest familiar thing is a saved preset, except the preset can have its own form fields, its own multi-step sequence, and its own defaults baked in by someone who understood the problem. A preset remembers your settings. A tool remembers your process.
So when Villavicencio and Carvallo wanted separated audio layers, they did not file a feature request and wait two years. They described a tool that generates the layers separately and exports them individually, and then shared the address. The separation was a design decision made by the people who needed it, not a roadmap item prioritised by someone who does not cut sound.
That is the actual shift, and it is worth saying without decoration: the people who know what the tool should do are now the people who can make it.
Put this into practice
Start with somebody else's tool, not your own. Building from scratch first is how people bounce off this.
-
Open Mondo Sónico and describe a scene you have actually had to sound. Not "spooky forest." Something from a real job: "interior office, late evening, distant HVAC hum, one printer cycling, footsteps on carpet approaching from the left." Specificity is what separates a useful set of stems from mush.
-
Export the individual tracks, not the mix. This is the entire point of the tool. Pull them into your timeline and try what you cannot do with a single generated file: duck the ambience under a line, slide one effect two frames, mute a layer that fights the music.
-
Pick one chore from your own week and find the tool nearest it. If you post video, that is probably CaptionCast or ThumbnailForge. If you do interiors or sets, Surface. Run it against a live job, not a test case. Test cases lie.
-
Duplicate that tool and change one thing. This is the smallest real step into building. Do not start with a blank page. Take a working tool by someone who understood their own problem, then adjust the one input that does not match how you work. A field that asks for the wrong thing, a default aspect ratio, a style you would never use. You learn more from one edit to a working tool than from an hour on an empty one.
-
Only then describe your own. Open Flow, say what the chore is, what you hand it and what you need back. Aim narrow. "Generate videos" is not a tool. "Take my three product photos and output the same hero shot at the six sizes my client's brand guide specifies" is a tool.
Time cost: steps one and two, twenty minutes. Step four, another twenty. Step five is where it stops being a lunch break, and I would not attempt it until a duplicated tool has already shown you the shape of the thing.
Honest limitations
All of this lives inside Google's plan structure. Flow runs on a credit system tied to the Google AI subscription tiers, with a free allowance and larger monthly allowances on the paid plans. The announcement contains no pricing at all, and the numbers floating around third-party comparison sites are not something I would build a budget on. Check Google's own pricing page before you plan a project around this, because credit costs vary by what you generate and the figures move.
Nothing here runs locally and nothing here is yours. These tools are addresses on Google's servers. If a tool is changed, deprecated or priced differently, your process changes with it. That is a real risk for anyone building a repeatable client workflow on top of one, and it is the opposite of the tradeoff you make with a local setup you control.
I cannot tell you how good the audio separation actually sounds. Google's announcement shows animated previews, not audio you can audition, and the claim of clean synchronized stems is Google's own. Separated layers that are individually mediocre are less useful than one good mix. That is the first thing to test in step two, and it is the question the announcement does not answer.
"Built by creatives" is a framing, not a guarantee. Six people's chores are not your chores. Surface is built around an interior designer's material-sourcing problem, and if your work is character animation it will not help you. The value of the six is proof of concept plus a starting template. The value for your specific week comes from step four onward.
Describing a tool is a skill and it will take you a few tries. Anyone who has worked with these systems knows that "make it good" produces nothing and precise instructions produce results. Building a tool is the same problem one level up: you are specifying a process rather than an output, and vagueness compounds. Expect your first attempt to ask for the wrong things.
The question I keep coming back to
If the person who knows the problem can now build the tool in an afternoon, what happens to the software they used to wait on?
I do not think the answer is that professional tools disappear. Nobody is replacing a full edit suite with a described tool. What I think happens is narrower and stranger: the long tail of small utilities, the plugins and scripts and one-off batch processors that every studio accumulates, stops being something you buy or commission. It becomes something you describe on the Tuesday you get annoyed.
Which raises a question I have not resolved. Those accumulated utilities are part of how a studio's craft gets passed along. The script someone wrote in 2019 that everybody still uses carries a decision they made about how the work should be done. If every one of those becomes a disposable thing each person describes fresh, the craft does not vanish, but the record of it does.
Maybe that does not matter. Maybe shareable addresses are a better record than a folder of undocumented scripts, and Mondo Sónico being a link anyone can open is exactly the point.
If you make something with these, especially if you duplicate one of the six and bend it toward your own work, I would like to know what you changed and why. The edits are more interesting than the originals.
Medium metadata
Reading time: about 8 minutes