Week 1: I built an app I can't launch
Published 5 months ago
Three days live on YouTube. One app built from scratch. And I can't even launch it yet. Let me explain. Last week I kicked off the $10K MRR Challenge. The plan was simple: stream, build, ship, repeat. I went live three times. And during those streams, I actually built a complete app: CoachAI. But here's the thing. I built it for the Shipyard hackathon. Shipyard is a hackathon organized by RevenueCat where they partnered with 7 creators. Each creator suggested an app idea. I picked one that's also been sitting in my notes for months: "Simple AI coaching app." The idea? Get mentored by the best coaches in the world. Well, not directly, but by an AI trained on all their public materials. Imagine having a question about marketing and asking Hormozi AI for his opinion. Or getting Tony Robbins to review your business plan. You basically build your own board of advisors. I broke my own rules Before building anything, I do keyword research. I use Astro to find keywords with high popularity (people are actually searching) and low competition (possible to rank). The results for this app? Not promising. Keywords like "business," "coach," and "learning" are way too competitive. Others have zero popularity. There was one keyword: "mentors". It was showing 30 popularity and 11 competition. Looks too good to be true. Probably a bug. If a founder from the Incubator came to me with this data, I'd tell them to find another idea. But it's always easier to give advice than to follow it yourself. I was live. Had to make a decision. So I said fuck it, I'm doing it. My reasoning:
3 days of building Day 1: Set up the project using my boilerplate (onboarding, analytics, RevenueCat, CI/CD—all preconfigured). Copied components from the AI Girlfriend app. Set up Supabase, anonymous auth, and Stream Chat. By end of day, users could send messages. But the coaches were silent. No AI agent yet. Day 2: Built the AI agent using Langgraph. The part that listens to messages, generates responses with an LMM (claude in my case), and sends them back. Initially, the answers felt like ChatGPT. Long walls of text in single messages. I wanted it to feel like chatting with an actual coach. The fix? I just told it in the prompt: "split long messages into smaller messages." That's it. The agent figured out the rest. Day 3: Polish. Onboarding screens (used a new AI tool from the guys at screendesigns. It exports files optimized for Claude Code, and I basically one-prompted the whole flow). Icon was a struggle. Couldn't find inspiration, so I made a simple "C" in Canva. No time to waste on an icon. Finally, I set up the paywall with RevenueCat and uploaded to TestFlight. Done. Here is a visual demo if you are curious. The problem The app is ready. But I can't publish it until the hackathon ends in late February. Which means it's not moving the MRR needle right now. So I have to get back to the other apps in the portfolio. The ones that can actually generate revenue while Coach AI sits in TestFlight. This is just the beginning. Follow me on X for daily updates. Subscribe on YouTube. I'll keep streaming, maybe not every day (that's harder than I thought), but often. What would you like to know more about? Hit reply. Cheers, Vadim PS. I really hope nobody receives flirty messages from their business coach. I copied a lot of code from the AI Girlfriend app. Pretty sure I cleaned it all up. Pretty sure. |