Lessons from Building FanMeter

Idea originated at 1:45 AM, 21st February. MVP shipped ~27 hours later at 5 AM, 22nd February. And no, I have not touched the code and I don’t feel bad about it

At the time of writing this entry, I had 4 hours of sleep in the past ~62 hours. I still feel excited to blog about it.

Now, I know that spending 27 hours to vibe code an application in 2026 is probably too much time. I know this because I did ship another vibe coded app in less than 2 hours. But fanmeter.in is different because it is the first one which I’m proud of. It is an application I’d use myself (I couldn’t say that about the other one). It is also one which I know my friends and family like to play, at least enough to share it with their own friends.

Here are some of my learnings from building it:

1. Taste is the differentiator

One of the main reasons why I’m happy with Fan Meter is because of how the UI design turned out. Inspired by Greg Isenberg’s podcast with Suraya Shivji, I looked for inspiration on cosmos and generated the design screens using Flux 2 Pro on Weavy AI.

I find that this is a great workflow. The search & similarity algorithm on cosmos is top-notch and that really allows you to choose your theme and get an image model to design UI screens matching that theme. There was a lot of back and forth iteration involved in this process for me as I kept chasing and fine-tuning for that exact look and feel that I was happy with. I will be experimenting a lot more & write a blog post on getting LLMs to create the frontend of your apps as per your taste.

2. Build something you’d use (duh!)

At least to begin with, build something that you would use. I love Geoguessr and when I had the idea of something similar but with movies, I was so excited. AI narrows the gap between idea and execution. Sure, we can build stuff for others or to make money, but I think it would be better to only do it after gaining enough experience by building stuff for oneself and learning what works, what doesn’t and how to best build with AI.

3. Be fully involved in planning

I mentioned I haven’t touched the code but I was fully involved in planning. Every time I prompted Claude Code to build a feature, I used this prompt suggested by Matt Pocock:

Interview me relentlessly about every aspect of this plan until we reach a shared understanding. Walk down each branch of the design tree, resolving dependencies between decisions one-by-one.

This put me in the driving seat, and I find it very comforting. I like this because not only does it give me an understanding of the technical architecture, but also gives me more points in the iteration loop where I can influence the direction the application is taking.

4. AI can do more than what you think

One thing I’ve learned from the past few months of experimenting with AI is that it is a lot more capable than what I thought. For example, when I started iterating on the UI design changes for Fan Meter, after each change, I would share the screenshot of the page to Claude to point out something it got wrong. I did it enough times that it got annoying. Then I figured I could just ask it to do that itself.

After every frontend change, take a screenshot of the page you are working on and see what’s different from the reference design

That’s it, and from then on, things started moving faster because I didn’t have to be involved to do the manual process of taking a screenshot and pointing out what’s wrong. I was completely blind to the fact that Claude Code could take the screenshots and check the changes by itself. I’m sure there are probably better ways to do the same but it is my lack of knowledge or underestimation of Claude which caused the bottleneck.

5. Reach matters a lot

If you are building to make money or gather as many users as possible, reach matters a lot. Because the effort required to build something is low now, anyone can build anything. There needs to be a strong marketing engine. I wrote about a post which covers this in more detail.

6. It’s an addiction

Claudoholism is real. I know that because I am a Claudoholic. I alluded to this in my post GPT-5.3-Codex-Spark and AI coding addiction

Seeing your idea come to life by simply prompting in natural language is very exciting. I have been staying awake past bedtime this week because I couldn’t stop telling myself “just one more prompt”.

But I’m absolutely loving this. You can just build stuff. All the things I’ve wanted to build but couldn’t be bothered before, I can build them now.

7. I have to rewire my brain.

I’m 26 years old now and there’s probably a layer of thought in my mind which just subconsciously rejects some ideas from reaching the conscious surface. In all of these years, my brain has learned that there are some things which I can’t possibly do because they are so out of reach considering the amount of time and effort that’s required. All of us have a calculator in our minds which decides whether we can do something or not, this calculator has been fine-tuned throughout our life. The older we get, the stricter this calculator is. Everytime I learn that AI can do something which we all thought it couldn’t do , I start feeling anxious about this calculator. What ideas am I blind to because of the understanding of my capabilites I have developed in my life before AI? I feel restricted by my learning of my capabilities, it is outdated now. I feel like I want to reset the calculator and start developing a sense of my capabilites from scratch (like a kid).

This experimentation I am doing now with AI is part of my effort to make my calculator more flexible.