I built a chat app where you can ask history’s sharpest minds a question instead of getting one generic AI answer. Building it, testing everything myself end to end, a bunch of things broke in ways I didn’t expect. Here’s what went wrong, and what I learned.
1. The persona I wanted most, I couldn’t build. My first plan had Elon Musk, Steve Jobs, MrBeast, and Warren Buffett as personas. Then I thought it through: putting made-up opinions in a real person’s mouth and selling it is a bad idea, legally and otherwise. So instead I built fictional “archetypes,” like The First-Principles Builder and The Perfectionist. Openly fictional, but honest about who inspired each one. The same problem came up again with a negotiation persona I wanted to base on Harvey Specter from Suits. Different reason this time (he’s a copyrighted character, not a real person), same fix.
2. I assumed every AI provider worked the same way. They don’t. The app has no backend server. It talks to the AI straight from your browser, using your own API key. Turns out OpenAI blocks that completely. Google’s Gemini sort of works, but only unofficially. So no option. Anthropic is the only one that built this on purpose, with a setting made exactly for this. That one fact decided my whole tech stack for me.
3. “Unknown error” was actually a billing message I wasn’t showing. Right after adding my API key, I got a cryptic failure. Turned out Anthropic sends back a real, useful error, something like “your credit balance is too low.” My code was just showing a generic status number instead of that message. Once I started showing the real error text, this stopped being confusing and i knew that i have to put funds into my account. Upside: Only a few Euro will cover a whole lot of conversation. I tested with ~ 4 Euro.
4. Asking all 17 personas at once was basically a stress test. New API accounts only allow a handful of requests per minute. Asking the whole board fires up to 17 requests at once. I hadn’t done that math. Fixed it by sending a few requests at a time instead of all at once, and automatically trying again when one got blocked.
5. Correct data, wrong spot. One historical persona was tagged correctly in the code, but still showed up visually grouped with the fictional ones. The label was right, but I’d put him in the wrong position in the list, and the app displays people in list order. A good reminder that “the data is correct” and “the data is in the right place” are two different bugs.
6. Two rules that quietly disagreed with each other. The Claude Skill version had one rule for “ask a few personas” and a separate rule for “ask the user first if nobody’s named.” A plain question like “what makes a good product?” matched both rules, and the wrong one won. Nothing was wrong with either rule alone, they just overlapped. Fixed by making one rule always run first, instead of two rules competing.
7. The right behavior for one message was wrong for seventeen. Auto-scrolling to the newest reply makes sense for a normal, one-at-a-time chat. It gets annoying fast once several answers arrive in a row, since the app keeps pulling you back down while you’re still reading an earlier one. Fixed it so it only auto-scrolls if you were already at the bottom.
8. One size limit, two very different jobs. I had one fixed limit on how long an answer could be. Fine for a quick reply. Then I asked for something bigger, like a full marketing plan, and it got cut off mid-sentence. A short chat reply and a long document need different limits, not the same one.
9. Buttons & Buttons. First the ‘going-back’-button was not floating. So sometimes you got stuck in the app and got confused how to continue. Same problem with the counsil mode. you could easily add personas to your counsil but getting rid of them was not so easy. In the end the counsil button also blocked the input window. Friendly reminder, UX matters :D
But i do not want to be all negative. The thing I’m happiest with: every persona’s little icon is generated by code, not a picture file. Just math based on the persona’s name. Nothing to host, nothing that can quietly break later.
I ended up shipping two versions: the full app, and a lighter Claude Skill for people who’d rather it just live inside the Claude they already use. Almost everything that made the final product better came from something breaking first.
Check out the Board of Great Minds on Gumroad and let me now what you think:
👉 Board_of_Great_Minds_Claude_Skill