What makes a good chatgpt app

The App Store gold rush taught us one thing: the presence of a new platform does not automatically make you a better product person. Most apps were noise and very few were signal. The same dynamic is now playing out with AI.

ChatGPT Apps - the new plugin-like ecosystem that lets you embed your product directly into a ChatGPT conversation - opened up a genuinely interesting design space. OpenAI's own guidance on the topic is well-written and worth reading. But as a team that has built over +150 software products across 12 years, we want to add our own perspective to the conversation.

Because the technical question ("how do I build one of these?") is not the hard part. The hard part is the design question: what should you actually build, and why?

The wrong starting point

When most teams think about building a ChatGPT app, they start with: "We already have a product. Let's bring it into ChatGPT."

That's a reasonable instinct but it's also usually the wrong frame.

The problem is that "porting" a product into a new context rarely works. It didn't work when mobile apps first arrived and every company tried to squish their desktop experience into a 4-inch screen. It doesn't work here either.

Inside a ChatGPT conversation, your product isn't the destination. It's a capability the model can call on at the right moment. Users aren't opening your app. They're mid-conversation about something, and your app might or might not be useful to that particular moment.

The apps that work inside conversational AI are the ones that offer a small number of very well-defined, specific things. Not a full product. Not a navigation hierarchy. Not an onboarding flowThink of it as a toolkit. What can your product do that the model can't do on its own? OpenAI distils this into three categories, and we think the framing is genuinely useful:

Know - does your app give the model access to information it wouldn't otherwise have? Live inventory, account data, proprietary datasets, real-time prices, user-specific history. If your product sits on a valuable, current, permissioned data source, that's your superpower.

Do - does your app take real actions on behalf of the user? Booking things, updating records, sending notifications, triggering workflows. This is where conversational AI starts to feel genuinely agent-like rather than just conversational.

Show - does your app present information in a structured, visual way that plain text can't? Comparison tables, timelines, shortlists, charts. Not because it looks nice, but because some information only becomes useful when you can see the structure.

If your app doesn't move the needle on at least one of these three, you're probably building something that the base model can already do.That's a demo, not a product.

The real question

Here's the thing we keep coming back to at MarsBased: none of this is purely a technical question.

We've been watching the AI ecosystem evolve from close range. We use Claude and Cursor heavily in our own development workflows. We've built agent-based features for clients. We've seen what good looks like, and we've seen what happens when teams treat "we added AI" as the finish line.

The teams that build useful AI-powered products are asking a different set of questions. Not "how do we integrate with this API?" but "what does the user actually need at this moment in the conversation?" Not "what features can we port?" but "what would genuinely be useful that they couldn't have before?"

That distinction matters because it's where craft comes in. Good product thinking is not made obsolete by AI. If anything, it becomes more valuable, because the gap between teams that think carefully and teams that ship fast is only going to get more visible.

That bein said, a few practical conclusions, based on our experience:

Start from the moment, not the product. Ask yourself: in what specific type of conversation would my product genuinely help? What is the user trying to figure out or accomplish at that moment? Design backwards from that.

Small scope is a feature, not a bug. The best apps will probably expose three to five well-defined capabilities. Resist the urge to be comprehensive. Comprehensiveness is for web products. Conversational apps need precision.

If you can't explain the value in one sentence, it's not ready. "It lets you check your real-time inventory without leaving the conversation" is a product. "It's basically our app but inside ChatGPT" is not.

Quality of integration matters more than quantity of features. A single capability that works flawlessly in the right context is worth more than ten features that feel bolted on.

We're at an early, noisy moment. The ChatGPT App ecosystem will produce a lot of mediocre integrations before the genuinely useful ones shake out. That's how every platform transition works.

But the fundamentals don't change. Great products solve real problems in ways people can actually use. The interface layer is different. The underlying discipline isn't.

If you're thinking about building a ChatGPT app and you want to pressure-test the idea, you know where to find us.Let's talk!