Vibe Coding in Practice: 7 Real Projects Reveal AI Programming's Product-Market Fit

The Verge's podcast The Vergecast ran a special episode called "Bear Lair" — a clear nod to Shark Tank — where hosts David Pierce and senior AI reporter Hayden Field played investors armed with "$1 billion in fake money," listening to listeners pitch apps they'd built through vibe coding. What looked like entertainment turned out to be a surprisingly rich snapshot of AI-assisted development in the wild. Not a single project was a flashy proof-of-concept — nearly all of them solved real, specific pain points from their creators' everyday lives.

Why Vibe Coding Deserves to Be Taken Seriously

David opened by admitting he's a coding novice: "I can write you an HTML website, and that's about it." But he made a pointed observation: vibe coding is the first time AI has genuinely found product-market fit. This isn't abstract tech optimism — it's grounded in something observable. Ordinary people are using natural language to treat models like Claude and Gemini as full-stack developers, filling in the long-tail needs that commercial software never bothers with.

The tech stack that kept showing up throughout the episode tells the story: Claude Code for code generation, Firebase for the backend, Whisper for transcription, and various MCP (Model Context Protocol) servers for connecting data sources. Simon, a listener from the UK, put it plainly: "I built websites in Dreamweaver in the '90s, know a bit of frontend HTML, and that's it. Building a full-stack app was completely out of reach for me." Now he has a product with real users.

What is MCP? MCP (Model Context Protocol) is an open protocol released by Anthropic in late 2024, designed to standardize how large language models connect to external data sources and tools. Think of it as the "USB port" of the AI world — any app or service that implements the protocol can be called by a model like plugging in a device. In the projects shown on the podcast, MCP servers let Claude Code directly read and write data from services like TickTick and Trello without manually writing API integration code for each one. This dramatically lowers the engineering barrier to "giving AI access to real-world data," and is a key piece of infrastructure enabling vibe coding to assemble useful tools quickly.

From Smart Home Dashboards to Eulogies: Seven Shapes of Real Need

The first presenter, John from Washington D.C., completely repurposed a Skylight smart display — essentially a cheap Android tablet with no battery — into a fully customizable home dashboard. It controls lights, shows subway status, and remotely starts his car's AC. What really got David's attention was the shopping list feature: via MCP servers for TickTick and Trello, he and his wife can send each other grocery requests using whichever app they each prefer, with all the data running on an old MacBook Air with its screen removed.

David opened with a $25 million bid, and his reason was personal: "My wife and I have been together 13 years and I've never been able to convince her to keep a shared shopping list." That's exactly the core value of vibe coding — it doesn't solve universal problems, it solves problems that only matter to you.

Sarah from Chicago, a mother of a two-year-old, built a baby food tracking app that tackles the most exhausting part of parenting: cognitive load. What has the baby eaten? What does she like? What's still in the fridge? Is her diet balanced this week? Hayden was immediately hooked and drove her bid up to $115 million to claim it, for a similarly honest reason: "If I had a kid, I'd use this every meal. There's just so much to keep track of."

Archival Search: A Tool Journalists Have Always Wanted

Eddie from New Jersey manages all the media archives for his church. He used Gemini to vibe code a Python script that runs all the sermon MP3s through Whisper for transcription, chunks them and sends them to a model for summarization with hallucination checking, and ultimately generates a database that can be searched in natural language through Google NotebookLM. He even noted that hallucination issues "were a lot worse six months ago" — a straightforward, first-hand observation of model iteration from a real user.

David immediately offered $100 million, explaining that this was exactly the kind of tool he'd been looking for as a journalist: "I keep thinking, when was the last time I talked to someone about this? Sometimes it was a week ago, sometimes six years ago." Hayden willingly stepped aside. This case highlights another dimension of vibe coding: it lets professionals build custom tools for their own workflows without waiting on any commercial product.

What is hallucination? Hallucination refers to large language models generating content that's factually incorrect or entirely fabricated — one of the core challenges in deploying AI in real applications. In scenarios like archival summarization, where faithfulness to the source material is critical, the risks are especially high: a model might attribute statements to a preacher that were never said. Eddie's "hallucination checking" step typically involves comparing the model's generated summary against the original transcript (either through another model call or rule-based checks) and flagging anything in the summary that can't be traced back to the source. His observation that things have improved over six months aligns with the broader industry view: accuracy in faithful long-document summarization has improved measurably with each model iteration and better prompt engineering, but the problem isn't fully solved.

Privacy, Health, and Collectibles: The Edges of Long-Tail Demand

Simon from the UK built the emotionally weightiest project. As a humanist funeral celebrant, he'd seen how gathering photos from grieving families was a nightmare — scattered across WhatsApp, iCloud, Dropbox. GatherIn, built with Claude and Firebase, lets people upload photos through a single link or QR code, with public/private tagging and automatic expiration after a set number of days. It deliberately avoids becoming a photo storage service. A tool originally designed for memorials quickly got repurposed by users for weddings, parties, and other gatherings. David claimed it for $250 million, calling it "the easiest business to make money from."

Meels from Illinois built PainFrame, a pain-tracking app using two minimalist HTML pages that work across devices (including e-readers), with a 3D body model for marking pain locations and CSV/JSON export for sharing with doctors or running through Claude for analysis. Ruslan from Georgia (the country) showed off an open-source Pokémon card market tree diagram — a visualization tool he'd originally built for the stock market. When users asked for a crypto version, he had zero interest, so he pivoted to card markets instead, switching from Plotly to D3.js to render actual card images in the tree nodes.

D3.js vs. Plotly: D3.js (Data-Driven Documents) is a JavaScript-based data visualization library known for extreme flexibility. It binds arbitrary data to DOM elements and drives SVG rendering, enabling complex interactive charts. Compared to "batteries-included" libraries like Plotly, D3 has a steeper learning curve but offers nearly unlimited customization. Ruslan's switch from Plotly to D3 came down to a specific need: displaying card images directly inside tree nodes — something Plotly doesn't natively support. This detail also shows that vibe coding isn't magic: when visual requirements push past what general-purpose libraries can handle, users still need to make meaningful technical decisions with AI assistance, rather than just "describe what you want and wait for the code."

A Signal Worth Paying Attention To

At the end of the episode, David said something worth sitting with: "I kept waiting for the person who'd say 'I built something that only makes sense to me,' because that's the real beauty of vibe coding. But every single one of these was a good, resonant idea."

There's a conclusion buried in that observation: when ordinary people can use natural language to get a working app on the first try (Spencer's cycling calculator "ran on the first attempt"), the barrier to software development is being fundamentally redefined. None of these projects will ever top the App Store charts, but they precisely fill the gaps that commercial software abandons for reasons of economies of scale. Vibe coding's product-market fit might live entirely in needs that are "too small for any company to bother with."

From a smart home dashboard, a baby food tracker, and semantic archival search, to a memorial photo collector, a pain logger, and a card market visualizer — these seven projects together sketch the true shape of AI-assisted development in practice. It doesn't replace professional developers. It lets anyone with a specific problem become the person who solves it.