The initial commit was on August 9th and today I feel like I have a stable V1, so it took about 85 days. I have a 9 to 5 job so I could probably have done it in half the time if I was only focused on that.
The idea appeared while working remotely while traveling the Baltics, I was constantly using chatGPT to generate walking tours to explore the cities with specific requests, like I want to end somewhere where I can have lunch, etc. The results were really nice, but the experience was kind of annoying, receiving everything in text with no pictures for example. Since his responses usually made sense I figured the majority of the work would be frontend (spoiler alert: not the case) so I started building it more as an experiment to see if I'd be able to ship it end to end, not really worried about business models, growth, etc.
I think the main learning I'll take to my next project are:
- In the beginning I was trying to understand 100% of my code. When I started going to the backend part, I didn't care if it made sense to me or not. I could just feed context into the AI and he would make it work. I think it definitely helps to understand what you're doing to guide the AI, but it is not a prerequisite. If you're really experimental you could just copy and paste code and build an entire app with no idea how it works.
- Sometimes I'd get super obsessed about solving a bug or shipping something and I'd spend hours just getting frustrated. It's always better to just go do something else and leave it to the next day.
- In a perfect world AI would take care of the entire backend and integrations and I would spend my time doing cool interactions and perfecting the UI, which is what was most fun to me.
- Do not worry about optimizing costs (like Nextauth vs Clerk) before having growth, just go with what is easier/faster to implement.
- AI APIs responses are different from AI UI responses (the AI playground from Vercel helps a LOT to try different models, sadly only saw it after building everything).
Comparison of AI models for coding:
- Claude 3.5 Sonnet: best for most things, with weird exceptions. I don't really know what it is, but sometimes this model is super stupid. If I'm doing something big I'd always compare responses to the 4o. Also the limits can get super annoying if you're trying to build a lot in 1 night for example.
- Claude 3.5 Sonnet (new): Have been testing since the release which is not long ago, but seems substantially more reliable then the older version.
- Chat GPT 4o: is also good. Basically I'm constantly using this together with Claude. Unpopular opinion but if I had to choose only one I'd probably take the GPT subscription.
- Chat GPT 4o-1-preview: I like this model. Super different results from all the others. I'd use this especially for debugging or integrations.
- v0: great results. You can build entire pages with just screenshots. The output is never perfect but then you get it closer to what you want using the other models. My go-to for UI components.
--- I'm currently live on Product Hunt on #3