Sep 1, 2024 · 1h 7m · startup-ideas
I can't believe we coded an app with AI in 67 mins (V0, Cursor AI, Replit, Claude AI)
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this collaborative live-coding session, Greg Isenberg and guest Riley demonstrate how to build and deploy a full-stack, AI-powered startup idea extraction application from scratch using V0, Cursor AI, Replit, and Claude. Through persistent prompt iteration and high-agency problem solving, they illustrate how non-traditional programmers can turn conceptual ideas into production-ready software in under an hour.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Greg holds 12% of the talking time here. How this is scored →
Greg as informed peer 2.1 Guest teaching 3.9 Guest disagreement 1.3 Greg pushing back 1.1
speaking balance: gold is Greg, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 4:27 Steering away from faceless video app idea
Riley pushes back against Greg's suggestion of building a complex faceless video generator, steering the live demo toward a realistic transcript parser instead.
Hardest push from Greg ▶ 7:34 Pushing for flat design terminologyGreg interrupts the prompting sequence to insist on including flat design as a stylistic keyword to see how v0 handles design jargon.
Biggest teaching moment ▶ 45:50 Using Perplexity to update Cursor's API docs contextRiley demonstrates how to overcome model hallucination by using Perplexity AI to fetch exact modern API docs and pasting them directly into Cursor.
Greg holds their own ▶ 16:48 Comparing CS degree coding to natural language promptingGreg draws upon his computer science background to highlight the paradigm shift between manual syntax programming and natural language code composition.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Greg as informed peer | Guest teaching | Guest disagreement | Greg pushing back | Why |
|---|---|---|---|---|---|---|
| High Agency Versus Low Agency in AI Development | 3 | 5 | 2 | 1 | Riley explains the high-agency mindset required to build apps with AI, educating Greg on how multi-million dollar app skeletons like Notion can be cloned without writing code. Greg engages attentively and frames the concept in business valuation terms. | |
| Ideating the Startup Idea Presentation Generator | 3 | 4 | 2 | 2 | Greg suggests building a faceless video creator, but Riley gently pushes back toward a transcript-to-presentation generator suitable for live demoing. Greg contributes audience analysis criteria to the prompt. | |
| Greg Isenberg Promotion for The Idea Bank | 2 | 3 | 1 | 1 | Greg inserts a promo for The Idea Bank, then suggests trying flat design terminology on v0. Riley explains how v0 handles frontend Next.js plumbing and live code compilation. | |
| Designing the Interactive Sip or Spit Mechanic | 3 | 4 | 1 | 1 | Greg introduces his signature Sip or Spit mechanic, and Riley immediately structures a voice prompt that contextualizes purpose and Tinder-style swipe mechanics for the LLM. | |
| Iterating and Refining the Slider UI Controls | 3 | 3 | 1 | 1 | Riley iterates on the slider controls while Greg reflects on his computer science background versus natural language composing. Riley defines the concept of composing code. | |
| Structuring Multi-Slide Decks and Visual Polish | 3 | 3 | 1 | 1 | Greg requests subtle visual adjustments to the dot texture and slider labels. Riley agrees on the importance of taste and explains how AI building heightens design awareness. | |
| Visual Prompting with Annotated Screenshots | 2 | 4 | 1 | 1 | Riley demonstrates annotated screenshot prompting to fix UI inversion bugs in v0, showing how visual mockups allow rapid prototyping without deep code editing. | |
| Sponsor Message from BoringMarketing.com | 1 | 4 | 1 | 1 | Following Greg's ad read for BoringMarketing, Riley introduces the workflow linking Cursor, Next.js templates, and Replit deployment hosting. | |
| Configuring SSH Keys to Connect Cursor and Replit | 1 | 5 | 1 | 1 | Riley performs a detailed live tutorial on configuring SSH keys between terminal and Replit, admitting the process looks intimidating while guiding the viewer step by step. | |
| Generating Backend Logic with Cursor Composer | 1 | 4 | 1 | 1 | Riley configures Cursor Composer to edit multiple files across Next.js paths, explaining how context templates guide the model in writing backend logic. | |
| Installing Dependencies and Diagnosing Silent Execution Failures | 2 | 4 | 1 | 1 | After installing OpenAI dependencies via Replit shell, the test input fails silently. Riley explains the concept of API wrappers while troubleshooting the lack of error output. | |
| Debugging API Environment Variables and Logging Errors | 2 | 4 | 2 | 1 | Riley emphasizes persistence when dealing with failures and instructs Cursor to explicitly add error logging to the input fields so the bug can be identified. | |
| Refactoring Components and Fixing Framer Motion Errors | 2 | 4 | 1 | 1 | Riley pivots strategy by building manual input fields first before automating AI parsing, and resolves missing Framer Motion package errors in Replit. | |
| Retrieving API Documentation with Perplexity AI | 1 | 5 | 1 | 1 | Riley uses Perplexity AI to retrieve the latest Next.js documentation for OpenAI structured output, educating the audience on feeding up-to-date syntax to Cursor. | |
| Testing YouTube Transcripts and Switching to Anthropic | 2 | 4 | 2 | 1 | Facing continued silent failures with OpenAI API calls, Riley switches the backend to Anthropic Claude. Greg expresses surprise at how easily models can be swapped. | |
| Debugging the Anthropic API Response and Parsing Pipeline | 2 | 4 | 2 | 1 | Riley inspects new Anthropic error logs and instructs Cursor to be lenient when parsing JSON responses. Greg jokes about having to plead with the model. | |
| Achieving Successful AI Transcript Extraction and Evaluation | 3 | 3 | 1 | 1 | The app successfully extracts startup ideas from podcast transcripts. Greg and Riley verify the extracted data points and test the Sip or Spit swipe interaction. | |
| Stress Testing with Full-Length Podcast Transcripts | 2 | 3 | 1 | 1 | Riley feeds an 82-minute episode transcript into the app, successfully parsing 15 distinct startup ideas. Greg jokes about paying Riley 100k for the tool. | |
| Post-Session Walkthrough of the Deployed Full-Stack App | 2 | 4 | 1 | 1 | Riley showcases post-session updates including authentication, link parsing, and saving cards to profiles, and pitches Software Composers as a learning community. |
Statements from this episode (18)
Brown: AI tools can build a full Notion clone in seven hours
“It takes six or seven hours if you like really want to go ham to create something that looks a lot like Notion, but with a full database storage and everything fully with AI without writing a single line of code.”
Brown: Non-coders can build complex apps with 15 hours of AI practice
“If you have the grit to push through the errors, because there's a lot of errors once you start dealing with databases, but usually Claude will figure it out, and the more times you do it, the better you will be at, like, fixing those errors, and you get bette…”
Riley Brown had never written code in his life until 50 days ago
“50 days ago I had never coded in my life.”
Brown: Vercel's v0 is like Claude Artifacts on steroids
“And so this is a lot like clawed artifacts, but like on steroids.”
Brown: Explaining a feature's underlying purpose boosts AI prompt creativity
“I like to make sure I give it, I tell it the purpose because it will actually use that in the prompt and it might add to its creativity.”
Brown: AI development is about composing code rather than writing it
“I call it composing code. Like, you're kind of You're just kind of guiding it. You're telling it what you want, and you know what things are feasible and what types of things aren't feasible the more you do it.”
Isenberg: Knowing design terminology remains essential when building apps with AI
“I will say, I think knowing how to design, or knowing some of the terminology, is still gonna be super helpful.”
Brown: Anyone can generate a working website prototype in one minute
“We can go to any website you think is cool and literally screenshot it and then describe like what you should do with it. And you can in one minute have a working prototype for it. It's not a full app because it's not storing any of the data, but it is a proto…”
Brown: App deployment in Cursor is significantly harder than in Replit
“Because cursor, it's significantly harder to deploy apps. For beginners, Replit makes taking whatever you've created and putting it on the internet with a domain that you can share”
Riley Brown spends over $1,000 per month on AI tools
“My AI budget is over a thousand dollars a month.”
Brown: Cursor AI is the best software he has ever used
“Cursor is by far the most fun tool I've ever used is the best software I've ever used in my life.”
Brown: Cursor's multi-file editing Composer feature drives the tool's virality
“And now we can use the composer, which is the feature that is why cursor is going so viral. This basically allows us to edit multiple pages at once.”
Brown: Choose 'Save All' over 'Accept All' in Cursor for easier undoing
“If you're using cursor, hit save all instead of accept all. Accept all is not reversible. You have to ask the AI to reverse it instead of just being able to, like, hit reject.”
Brown: Non-coders should build workflows manually before using AI to automate
“If you don't know how to code, this is a good way to do it is to actually create it manually and then say, I want to automate this process and have that whole process happen in the background.”
Brown: Explicit runtime errors are preferable to blank screens for AI agents
“Error messages are good. Blank screens are bad, right? This is what should have popped up in the dev log, right? We should see the errors right here, and it actually knows what to do and we can actually just see how it reacts.”
Brown: Feeding Perplexity API docs into Cursor improves generated code quality
“Perplexity will actually find the latest documentation. It's just going to give better instructions. We can feed it to cursor, which will allow it to make better decisions, write better code.”