Make Cool Things With AI That People Buy: Three Ways to Use Gemini 4 Argon

Most people are going to use Argon like every other AI: ask a question, get an answer, move on.

After using Gemini 4 Argon nonstop for a week, I think that misses the opportunity. Three use cases stood out: marketing that brings people into your business, personalized physical products, and focused software that replaces a frustrating legacy workflow.

I built examples of all three. Some results were impressive immediately. Others needed real creative direction. That distinction matters, because the model can do a lot of the work, but you still need to know what is worth making.

Google sponsored my testing and gave me early access. These are my builds, my experiences, and my opinions, including the parts I think still need work.

Table of Contents

• Where Argon Shines, and Where the Tooling Needs Work
• 1. Vibe Marketing: Build Something People Want to Explore
• Teach Argon Your Writing Voice Instead of Asking It to Be “Less AI”
• 2. Vibe Manufacturing: Turn Customization Into a Physical Product
• 3. Boring SaaS Is Still an Opportunity
• Six Techniques That Improved My Argon Results
• The Plan I Used for Building
• Marry the Niche, Date the Product

Where Argon Shines, and Where the Tooling Needs Work

The three strengths I kept coming back to were:

• Marketing with AI: interactive websites, useful experiences, and writing that sounds like you.
• Physical AI: customizable product designs that can become files for 3D printing.
• Complex engineering: focused applications with the details needed to support a real business workflow.

My experience with the model was stronger than my experience with Antigravity, the environment I used to work with it. Antigravity is not my favorite harness yet. I missed parts of other development environments, and I think Google has catching up to do.

That is separate from the model itself. Argon can also be used in environments such as Cursor and OpenCode. You do not have to love one interface to find the underlying model useful.

The bigger point is that Google has a frontier-level model worth taking seriously. The question is where to point it.

1. Vibe Marketing: Build Something People Want to Explore

In 2021, my team built an illustrated website for our holding company. Different buildings represented different businesses, including our design agency, LCA, and our marketing company, Boring Marketing.

It was a huge project. By my accounting, it took about 40,000 hours and cost roughly $200,000. The business has generated more than $40 million in revenue, and I believe the website has played an important role in that. People remember it. They bring it up.

But it was also difficult to update. Every change took time and energy, so we often left it alone.

I asked Argon to make a better version, then gave it a more specific creative direction: take inspiration from Mondrian.

The stronger version became a small interactive world. It had movement, cars, sounds, a world passport, and little games, including shooting a basketball. Clicking around became part of the experience rather than just a way to reach another page.

That is what I mean by vibe marketing. Why default to an ordinary website when you can build something people actually want to spend time with?

Interactivity should have a purpose. Here, the purpose was to encourage people to discover the different companies. The passport and games supported that exploration.

The First Attempt Was Pretty Brutal

I do not want to pretend the beautiful version appeared from a vague, one-line request.

My initial instruction was essentially to make a better version of our existing site. Argon understood the broad concept, but the result was a sparse little 3D scene. It had the energy of an early metaverse demo, not a premium company website.

Other iterations had too many buildings. Another felt too much like Roblox. Those were not engineering failures. They were creative-direction problems.

Argon needs good references to produce good creative work. That was one of my biggest lessons from the week.

You still have to make judgments:

• Is the scene too crowded?
• Does the visual style feel premium or toy-like?
• Are the interactions helping people discover the business?
• Does the result have a point of view?

I actually like that the human still has a job here. Outsource every creative decision and you are likely to get something that looks like everything else.

The practical upside is not just a cheaper initial build. It is the ability to keep improving the site through conversation rather than treating every update as another substantial development project.

An Interactive Experience Can Come Before the Business

Another example came from Justine Moore at a16z. She shared an interactive museum of Apple computers built with Argon. In her account, she did not supply images or data. The system gathered archival computer photos, wrote an image-cleaning tool, collected decades of specifications, and built a 3D browsing experience.

The interface changed to match the era of the computer being explored. That is a much more memorable idea than a basic catalog.

This is the kind of project with the potential to spread on X, TikTok, or Instagram. It gives people something interesting to interact with and share.

It also suggests a different order of operations for starting a business:

1. Build a compelling experience for a specific interest group.
2. See whether it attracts that audience.
3. Learn what those people care about.
4. Explore relevant services, memorabilia, or physical products.

An audience of Apple enthusiasts is not a business by itself. But it is a much more useful starting point than a generic product with no clear customer.

Teach Argon Your Writing Voice Instead of Asking It to Be “Less AI”

Writing was the other big marketing surprise. In my testing, Argon was probably the strongest frontier model I had used for writing at that point.

But I did not just ask it to write a newsletter and hope for the best. I asked it to study how I already communicate and turn that understanding into a reusable skill.

The request was straightforward: research my posts on X, my newsletters, and how I speak, then internalize those patterns. After that, I would give it a topic and test the result.

It created a skill package with several distinct pieces:

• A core skill file: the main voice rules, sentence rhythm, mental models, and structural patterns.
• A voice and lexicon file: characteristic vocabulary, recurring phrases, and terms I use.
• Samples and templates: examples of my writing and the formats I tend to return to.

This was closer to the work I would expect from a thoughtful head of content than a generic “write in my tone” prompt.

Test the Skill on a Topic You Know Well

I asked it to write about marketing engineers. The result used ideas I actually care about: micro products instead of PDF lead magnets, old-world versus new-world comparisons, and landing pages short enough to fit into roughly 280 characters.

It even used my usual newsletter sign-off.

If a writer I had hired delivered that draft, I would have been excited. My subjective score was about 9.2 out of 10.

The important part is that it recognized more than surface-level vocabulary. It picked up the ideas, contrasts, and structures behind the writing.

You do not need a large public archive to try this. If your writing is not online, upload examples you already have and ask the model to study those. The source material is what gives it something specific to learn.

I would also use subagents for this research. Different agents can study different sources or aspects of the voice, rather than making one agent handle every direction at once.

AI does the plumbing. Humans bring the taste. A reusable voice skill gets you closer to a good draft, but you are still responsible for deciding whether it says something worth publishing.

2. Vibe Manufacturing: Turn Customization Into a Physical Product

Physical AI is the opportunity I think too few people are exploring.

The basic idea is simple: use AI to help design a physical product, export the design as an STL file, and move it into a 3D printing workflow. The interesting business is not merely printing objects. It is giving a particular audience something personal enough to care about.

I asked Argon to build a Vibe Manufacturing Studio: an embeddable product customizer for a Shopify store.

The prototype included a live 3D product preview alongside estimates such as material weight, cost, print time, and suggested retail price. It also included an STL export.

Imagine a customer choosing a size or color, seeing the result, and ordering that version directly from a store. There are millions of Shopify stores. A useful customization layer for even a narrow category is an interesting place to build.

Start With a Niche That Already Cares

For a snowboard-oriented store, you could explore models of recognizable mountains or personalized figures. A customer might want something that represents a favorite place or includes their own likeness.

The point is attachment. People often care more about an object when they helped shape it, especially when it is a gift or a visible signal of an interest.

One example in my studio was an espresso portafilter station. The demand hypothesis came from the espresso community: if someone has spent $2,500 on a machine, a $49 accessory customized to their setup is a plausible purchase.

That is a hypothesis to test, not proof of sales. But it is a much more grounded starting point than asking AI to invent a random object.

The niche gives you the insight. Argon helps turn that insight into something you can show people.

The Made-to-Order Workflow

Here is how I would approach the store:

1. Build the product customizer. Start with a narrow product and a manageable set of options.
2. Embed it in Shopify. Let customers choose supported variations and see a live 3D preview.
3. Take payment through the store. Keep the purchase connected to the chosen configuration.
4. Check the generated print file. Validate it against the size limits and configurations you have tested.
5. Print and ship. Use your own equipment or a production partner.

The file check matters. A preview on a screen is not the same thing as a successfully manufactured object. I would not skip testing the supported configurations.

If you own a Bambu Lab printer, you can use the exported STL as part of your printing workflow. If you do not want printers filling your house or office, services such as Slant 3D offer a print-on-demand route, including API-based workflows.

You do not have to own the factory. You do have to understand demand, choose the right product, and check that the production workflow works.

You Can Sell the Service Instead of Running the Store

There is another business here: build these customizers and product programs for existing brands.

The pitch is not “I can make you a 3D widget.” It is that you can help the brand create objects its customers want on their desks or in their homes, keeping the brand present in their everyday lives.

You could explore a margin-sharing arrangement to make the offer easier to accept. I mentioned potential margins of 40%, 50%, or 70%, depending on pricing. Those are illustrative possibilities, not guaranteed economics.

Start with a prototype, put it in front of the audience, and ask what they wish they could customize. That feedback is more useful than endlessly polishing a product nobody has asked for.

3. Boring SaaS Is Still an Opportunity

I keep hearing that SaaS is dead because AI can build software.

I think that argument confuses people who enjoy making software with people who just want a problem solved. Most customers do not want to become software developers. They are still happy to pay for something that makes their work easier.

The opportunity is not necessarily to rebuild a giant platform. It is to find a recurring, boring problem and solve it for a narrower customer.

Find Complaints Before You Pick Features

I would start with sources such as G2, Capterra, Reddit, and Idea Browser, the startup research product I built.

Look for repeated complaints about software people already use:

• A workflow that remains unnecessarily complicated.
• A legacy product that has not kept up with its customers.
• A broad tool that fits a particular niche poorly.
• A recurring task still handled through awkward spreadsheets.

Then build a focused product around that problem. Go narrower than the incumbent where it helps you serve the customer better.

Give Away the First Tool

The usual plan is to build something, charge $50 a month, and hope to acquire thousands of customers.

I would seriously consider a different approach: make the first useful tool free.

Not a PDF about the problem. A tool that actually solves a piece of it.

That tool puts you in front of likely buyers and gives them a reason to trust you. Your second, third, or fourth product can be paid. Over time, you can build a small product studio around one ideal customer profile, or ICP.

The free product is a customer acquisition strategy, not the entire business model. Once you have earned trust and learned more about the customer, you can offer something else they value.

The Test Build: A Commercial Cleaning Quote App

I tested this with a very specific workflow: turning a commercial cleaning company's site visit notes into a reviewed price proposal.

The goal was a focused replacement for a quoting spreadsheet. Creating a quote needed to be the main experience, not one feature buried inside a sprawling dashboard.

My build request included these ingredients:

1. The customer and job: a commercial cleaning company preparing a quote after a site visit.
2. Supporting evidence: customer complaints, source links, and an anonymized workflow where available.
3. Project inspection: examine the existing project and explain the implementation plan briefly.
4. Implementation constraints: reuse the existing technology stack where possible.
5. Action boundaries: do not deploy the app or send proposals.
6. An honest handoff: list what remains unfinished.

If you do not have the supporting research yet, you can ask the agent to find it. But the request should still make clear which customer problem it is investigating.

The result included interactive pricing and margin calculations. It was not the most beautiful interface, but the engineering and workflow details were the impressive part.

That matched my experience elsewhere: Argon often got the functional structure right early, while the design needed references and more iteration.

I was encouraged by how the app worked in my testing, including the deployment-related work I tried in Antigravity. That is not a blanket guarantee that every generated application is secure or production-ready. It is why the prompt asks for unfinished work explicitly, and why testing remains part of the process.

There are thousands of narrow, paid software workflows worth examining this way. If you can offer a genuinely useful free alternative, you have a reason to start a conversation with the right customer.

Six Techniques That Improved My Argon Results

These are not magic tricks. They are ways to give the system a clearer job and build a better feedback loop. The biggest improvements for me came from planning, delegation, and goal-directed iteration.

1. Use /plan Before Changing Files

Agree on what to build before the agent starts implementing it.

I should have done this with the holding company website. A better conversation about the visual direction, layout, and interactions would likely have improved the first version and reduced the amount of wandering afterward.

2. Use Subagents for Independent Work

Delegate jobs that can run separately, such as researching different sources or checking different aspects of a project.

For the writing skill, separate agents could study posts, newsletters, and spoken material in parallel. The idea is ordinary delegation: do not make one worker switch among every task when the work can be split.

The subagent command was described as /subagents or /sub-agents in my walkthrough. Use the form supported by your environment.

3. Use /goal With a Clear Reference

One of the turning points for the website came when I used /goal, gave it a reference image, and pushed it toward a stronger result.

This used more tokens, but the output improved noticeably. For an asset with meaningful business value, spending more on iteration can be an easy decision.

The reference is important. “Make it amazing” is much more useful when the model can see what amazing means to you.

4. Use /browser to Test the User Journey

Have the agent open the application and exercise the actual experience.

Reading code and using a product are different activities. Browser testing helps bring the work back to what a customer actually does, rather than stopping at an implementation that merely looks complete.

5. Use /learn to Preserve Corrections

Turn recurring feedback into reusable project rules or skills.

If you keep correcting the same thing, the answer is not to repeat yourself forever. Build a system that carries that correction into the next task.

6. Try High Effort After Learning on Medium

I started on medium effort and later switched to high. High became my default because I generally preferred the output.

I would still start on medium while learning the workflow. Then compare the results and decide where the extra effort is worth it.

The Plan I Used for Building

During my test, I used Google AI Ultra at a quoted price of $100 per month. The access I described included the Antigravity app, its command-line interface, and an allowance described as up to one million output tokens.

I contrasted that with the $19.99 Google AI Pro plan, which I described as better suited to smaller text-generation and web-search tasks and as lacking the same development access.

Those were the plan details presented during my testing, not a permanent pricing guide. Check the current prices, limits, and included tools before buying. For the kind of building I was doing, Ultra was the option I used.

Marry the Niche, Date the Product

The hard part is increasingly not getting the model to produce something. It is finding the right idea, gathering good references, and choosing a customer you genuinely want to serve.

An interactive website can attract that customer. A personalized physical product can deepen their attachment. A focused free tool can earn their trust.

Those are different products, but they can all serve the same niche.

That is why I keep coming back to this rule: marry the niche, date the product. You can change the thing you make. You should care enough about the people you serve to keep learning what they need.

The model gives you more room to experiment. Your taste, judgment, and understanding of the customer determine whether those experiments become things people actually buy.