I started my own meeting manager with a sketch, not a database.
In a June 2026 live stream I started building the meeting app I want. In one session I went from a rough sketch to a prototype running on my machine. The part worth copying is the order: settle how the screen should feel, then add structure.
The business case
I do a lot of meetings, and recapping the last one before the next one drives me nuts. Plenty of tools exist for this. None of them does what I want.
That alone isn't a reason to build. The number-one startup mistake I see is building before you know anyone wants the thing. Test it first with a landing page and some ads, or a mockup you try to sell, then build what people keep asking for.
My rule for vibe coding a tool is simple. Either it's an internal tool that buys you back real time, or you plan to sell it and have outside proof of demand. I have both. Three years ago I compared meeting note takers on my second channel, Dwain: Unfinished. That video sits at about 31,000 views, with 64 comments asking very specific questions. You can read the comparison I ran in 2023, and the case for testing small in the smallest system that proves the point.
Why I'm leaving Read AI
I've used Read AI for about three years. It joins your meeting like a participant, records it, pulls out the notes and gives you a summary. It's a good tool. I also tried Fireflies and Otter, and Read AI was the best at the time.
It stopped fitting me. It feels clunky and slow, and I can't sort meetings by company without building folders by hand. It also carries scores I don't need, like engagement and sentiment. Sentiment only ever tells me I talk too much.
I want notes turned into tasks, every meeting with one company in one place, summaries across several meetings without asking the AI each time, and meeting objectives the AI checks me against. After two years on the Enterprise plan I downgraded to cut costs, and I lost the highlights from my past meetings. That's when I decided to build my own.
Sketch first, then Claude Design
I drew the screen by hand. A sidebar with search and a list of companies and people, business and personal. Click one and a list of meetings opens, grouped by date. Click a meeting and the summary sits in the middle, with its to-dos on the right. I also want every open to-do across all my meetings with a company, because I forget what I promised three months ago.
I pasted that sketch into Claude Design and asked for a modern, minimalist interface. Then I answered its setup questions: no dark mode, plenty of breathing room, a clickable prototype with sample data, action items with due dates and a status, big buttons, a way to paste in old notes, and a button to summarize with AI. A viewer suggested appointments and reminders, so I added them.
Chat voted down the first design, an orange one, so I asked for a blue scheme and a more modern font. Claude also added a workspace switcher I hadn't asked for. I liked it: one workspace each for SnapSuite, LeadScore AI and Affordly.
I trust it for real work. I was paying about $6,000 a year for a scheduling system, so I rewrote it in a weekend, designed in Claude Design. It runs inside one of my apps as an embedded frame that talks to the main app.
Jump to
- 0:00 Why I'm building it, and validating before you build
- 16:46 What Read AI doesn't do for me
- 29:12 From a hand sketch to Claude Design
- 47:12 Picking a design, and the Google Stitch comparison
- 58:50 A PRD and JSON data model from a screenshot
- 1:03:25 Why I build on Cloudflare
- 1:13:11 Turning the design into a working prototype with Codex
- 1:21:08 My scheduler rewrite, then pricing and run-rate advice
I build the whole thing on screen in the full stream on my channel: Building a meeting manager live on YouTube.
The Google Stitch test
A few months before this stream, I'd have started in Google Stitch, then moved the design into Lovable or Bolt. By June 2026 I'd stopped. Claude and ChatGPT were good enough at interface design that I went straight from the design to the app.
To be fair, I gave Stitch the same sketch and prompt, running Gemini 3.1 Pro with thinking. I thought the result was much worse than Claude's, and that Google's changes had made a good tool worse. That's one test on one day, so try your own prompt before you pick.
A PRD from a screenshot
Once I knew how the app should feel, I needed a data structure to match it. I took a screenshot of the design, pasted it into ChatGPT and asked for a detailed JSON data structure plus a PRD, a requirements document the AI can build from.
It came back with workspaces, users, companies, contacts, meetings and attendees. Each meeting had a summary, risks, next steps, notes, a transcript and attachments, and the PRD listed acceptance criteria. I used ChatGPT here on purpose. I keep Claude on design so it doesn't burn tokens on data.
Codex builds the prototype
I made a new folder and an AGENTS.md file, the instructions file coding agents read, and pasted the PRD in without formatting it. Then I opened that folder as a project in Codex.
The prompt asked it to turn the blue design into a working Next.js app with a local JSON back end, keep the color scheme, and not change the UI. I added that last line because Codex tends to redesign a Claude design when I move it over.
Normally I'd do this in Visual Studio Code with GitHub Copilot. My Copilot bill had passed $500 a month, so I skipped it. By the end of the stream the app ran locally, and it was pretty much what I wanted.
Design first, back end later
I build on gut feel first. Once I know where a button or a drop-down makes sense, I add structure. Part two of this build adds the database and back end, then a webhook layer so Read AI and other meeting tools send their data in automatically.
That part goes on Cloudflare, where I build my vibe-coded apps with its D1 database. Setup takes more work than Vercel, but I haven't found anything cheaper. As of October 2026 the Workers paid plan starts at $5 a month.
The app exists to save me the recap before every meeting. If you're deciding whether a tool is worth building, and what to build first, we can work through it when you book a strategy call.
About Dwain
I'm Dwain Browne, Founder of LeadScore AI and SnapSuite. I'm a Toronto-based software architect with 25+ years across software, sales, operations and entrepreneurship. SnapSuite helps contractors and trades teams get paid faster, and LeadScore AI shows you which leads deserve your next call. I also run Vibe Code to Profit, a community helping people build software with AI.