The 30-Minute Prototype That Replaces Your PRD (+ Could Help You Build a Billion-Dollar App)

➿ Grab the full SME playbook. Snag the prompts + handholding.

the problem: You had an idea three weeks ago. It was crystal clear—the flow, the interaction, the way users would use it.

So you wrote the PRD—or "brief," whatever your org calls it. Someone interpreted it into mockups. Someone else interpreted those into code. By the time it ships, the thing has been rebuilt three times through telephone, and half of what you wrote was fiction anyway—features scoped out, flows changed mid-build, edge cases that never materialized.

And somewhere in there, your sharp, clear idea got belt-sanded down into something you're not quite sure about anymore.

The average PRD cycle burns weeks of salary across your org. Tens—sometimes hundreds—of thousands of dollars of fanfiction about a product that won't exist as described.

the solution: Alex Berger, COO of Bolt.new, remembered something we all forgot after second grade: showing beats telling. So, instead of writing PRDs, he builds prototypes.

The basic idea is that in the time it takes to do your morning standup, you could copy code straight from a live product, paste it into an AI tool (think: Bolt, Lovable, Replit, and v0), and modify it. All while keeping the same fonts, spacing, and everything else. That way, your stakeholders can click through + review your big idea quickly—instead of squinting at a document or waiting for it to trudge down the pipeline.

Ideally, you’re not rebuilding a net-new product (and if you are, more on that later). Ideally, you’re changing one thing about the product you’re currently building or selling. This can be:

Navigate to the exact component you want to improve and focus there. If you can't point to a single pixel on the screen that you want to fix, you're probably trying to build too much at once.

Your first instinct might be to take a screenshot of the page you're on and throw that into an AI tool. Good. Do that, then steal the code behind the element. Here’s how to do that in three clicks and two pastes:

common mistake: Don't copy the entire page, either. Just grab the element you care about + its immediate container. The more surgical you are, the cleaner the output of the next step becomes.

Paste your screenshot, element code, and styles into Bolt (or wherever you’re working). Ask it to match everything without loss.

pro tip: If the output looks off, don't wrestle with it. AI stutters sometimes. Start a fresh chat and try again. Debugging a bad foundation will cost you more time than just regenerating a clean one.

You’ve got your “before.” Now tweak it. Tell the AI what you want changed—and just as important, what you don’t want it to touch.

If you’re unsure how to get your idea out of your head and into this AI tool, just describe your idea to ChatGPT and ask it to draft a prompt for you. Then, once it’s in, click around. Make sure it actually works as you envisioned. Then move to step 5.

This is the hard part for most people. You’ve built your idea. Now, send a Slack message or email to your stakeholders, your team, whoever needs to see it: “Hey. I built a quick prototype to show what I mean. Click around. I’m curious what you think.”

If they have questions, they’ll ask. If they don’t get it, you’ll know immediately—and you can iterate in the time it would’ve taken to schedule another sync.

some 2026 inspo: If you can clone a screen from any product, what’s stopping you from building something bigger? These tools have already minted founders who took the same approach at scale: clone what works, add what's missing, ship it.

➿ Grab the full playbook for the complete walkthrough, additional prompts for adding interactions, and Alex's full breakdown of why this approach kills PRD culture.

Three years. $10 billion valuation. Their business model? Recruiting ex-bankers, ex-consultants, and ex-BigLaw attorneys to teach AI systems everything they know.

… and pay can reach $200/hour.

the catch: Those insights feed directly into models designed to automate the industries these contractors just left. Think: people who spent a decade mastering finance, strategy, or corporate law are packaging that expertise for machines.

Mercor’s CEO says a small fraction of contributors—roughly the top 10–20%—generate most of the training value. Meaning that the sharpest humans are helping build the sharpest models.

what to do: If you’re in consulting, finance, or legal, your tacit, unspoken knowledge is becoming a commodity. Either you’re the one capturing that value, or you're the one getting replaced by it.

Claude Skills are customizable workflows that teach the AI how to do specific tasks your way. Think of them as saved playbooks—repeatable, standardized, and portable across Claude.ai, Claude Code, and the API.

Someone compiled 60+ of them into a single repo. The categories include:

There’s also a meeting insights analysis skill that surfaces behavioral patterns like who's avoiding conflict + who's dominating airtime during your calls.

The Financial Times reports on a Morgan Stanley analysis forecasting that 35 major European banks will reduce their combined workforce by roughly 10% by 2030.

With those lenders employing about 2.12 million people, that’s 200,000+ roles eliminated as they expand AI use, close branches, and push further into digital operations.

The bloodletting will hit the hardest in central services, mainly:

tldw: Google DeepMind dropped a 5-year documentary called The Thinking Game. CEO, Demis Hassabis, and his team filmed their entire run—from beating Go and Chess world champions to cracking protein folding with AlphaFold. It’s a behind-the-scenes look at the closest thing we have to an AGI moonshot. Demis cares so little about profits and so much about cracking AGI to help humanity—it's a refreshing reset.

eli5: When you talk to ChatGPT or Claude, you’re talking to a language model (or LLM). It reads text, predicts the next word, and continues until it sounds intelligent and does what it thinks you want. That’s it.

A world model is different. It understands how physical things work in 3D space—objects, movement, cause and effect. Not because someone hard-coded physics rules, but because it watched millions of videos of things happening and learned the patterns.

Think about how you learned as a baby. You didn’t read a book about gravity. You dropped things, watched them fall, and built an internal model of how the world works. That’s what these AI systems are trying to do.

what these models are for: They’re built for robots, self-driving cars, warehouse automation, video games, and simulations—any environment where AI has to move through and interact with physical space.

A bunch of heavy hitters are betting this is the breakthrough moment for world models:

apply it: Unless you work in robotics, logistics, or manufacturing, this won’t change your workflow tomorrow. But it explains why people are suddenly talking about “physical AI.” It’s also something you’ll see a lot of coming out of CES.

📧 Reply with your favorite AI use case. If it’s awesome, we’ll invite you to guest-star on our weekly show.