Home

What Do I Build Now?

Does your idea deserve its own interface? A framework for founders and product engineers in the age of agents.

So, I almost built another app. It was a trainer for mental models: short lessons, drills, spaced repetition, you know, all the bells and whistles. I had screens sketched and everything. Then I asked the question every builder now has to ask first: does this idea deserve its own interface?

Software is unbundling

For most of the smartphone era, software came bundled: a capability (what it can do) welded to an interface (where you meet it). Apple's App Store opened on July 10, 2008 with about 500 apps. By 2025, Apple's own transparency report counted more than 2.1 million. Every one of them follows the same deal: to use the capability, you go to the interface. You download the app, learn its menus, and move your data in.

That bundle is coming apart, and you can see it in the dates. ChatGPT reached 900 million weekly users in February 2026 and 1.2 billion by late September. That's a single interface where people now ask for things instead of opening apps. Your assistant already has memory, a scheduler, and access to your files. Anyone can create their own "capabilities" now by turning them into skills. A skill is a capability without an interface of its own, so it lives in the conversation you're already having.

Here's what that looks like in practice:

Skills became a format. Anthropic introduced Agent Skills on October 16, 2025: folders of instructions, scripts, and resources an agent loads only when a task needs them. Two months later it made them an open standard, with a partner directory that includes Atlassian, Figma, Canva, Stripe, Notion, and Zapier, and adoption by Microsoft in VS Code and GitHub and by coding agents like Cursor (VentureBeat).

The plumbing standardized. The Model Context Protocol, the open standard that connects agents to outside tools and data, had more than 10,000 active public servers and 97 million monthly SDK downloads a year after launch, and it's supported in ChatGPT, Gemini, Microsoft Copilot, Cursor, and VS Code (Anthropic).

Apps moved inside the chat. In October 2025, OpenAI launched apps in ChatGPT with Booking.com, Expedia, Spotify, Zillow, Canva, Figma, and Coursera. You say "Spotify, make a playlist for my party this Friday," and Spotify shows up in the conversation. Talk about buying a home, and ChatGPT can offer Zillow listings on a map without you ever opening Zillow.

Even checkout moved inside the chat. With Stripe, OpenAI launched Instant Checkout, which lets US users buy from Etsy sellers without leaving ChatGPT, with Shopify merchants announced next. Discovery, comparison, and payment, the whole shopping funnel that used to need an app, now happen in one conversation.

Greg Isenberg, a startup founder and investor who is one of the most-followed voices on X on where AI products are heading, puts it bluntly: skills are the new apps, and the harness is the new OS (post). He's right about a lot of software, but not all of it. For founders, telling the difference comes down to two questions. How should this software meet people? And who gets paid when it does?

Question one: how should it meet people?

Software can meet people in three ways. It can wait for you to tell it what you want (AI made this possible), it can give you an experience words can't, or it can sit beside you while you work. Each is a core frame your idea lives in, and each points to a different thing to build: a skill, a traditional app, or an ambient app. Here are the types of consumer software I'm seeing now: