What I Learned from Making runyournumber.com
BLOT
Recently I built a personal finance calculator website, runyournumber.com, using vibe coding. Honestly, I just wanted to bring an idea in my head to life. But as I kept building, I ended up learning far more than I expected — not just about writing code, but about how to collaborate with AI, how to understand markets, how to promote a product, and even how to rethink the way I approach work. Here are the things from this journey that I think are most worth sharing.
Why I Built This Website
The project is a personal finance calculator website. The content comes from important personal finance concepts I’ve read about over the past few years. I’ve always had this idea: many of these concepts exist as static text, and when you actually want to put them into practice, you have to grab a pen and paper and calculate things yourself, which is incredibly time-consuming. What I wanted to do was pair each concept with a calculator that uses concrete examples to tell you whether something fits your situation, how well you’re doing, and where you need to improve — all of it computable. I think this is highly complementary to the concepts you find in books.
Why Calculators Haven’t Been Replaced by AI
When I was discussing calculator design with AI, it mentioned that calculators are still partially “AI-proof” — meaning they’re not easily replaced by AI. Why? For a simple compound interest calculator, AI can directly compute the correct answer — no problem there. But when you’re doing financial planning, you need to provide a lot of personal information: whether you get paid biweekly or monthly, your mortgage situation, and so on. As these situations get more complex, the number of input fields grows, the underlying math becomes more involved, and it’s hard to get AI to give you the right answer with a single search query, because there are many details that need to be confirmed back and forth. So AI told me that this area isn’t necessarily going to be fully replaced by AI anytime soon.
Understanding the Market and Competitors
While building the website, I wanted to understand what competitors existed in the market, so I asked AI to do a market analysis. The competitors that initially came up were massive calculator websites like NerdWallet and SmartAsset. AI also pointed out that you can’t make money from calculators alone — these websites make money by partnering with other companies on their site, such as lenders or home refinancing services. They actually profit by generating market leads, not from the calculators themselves. So it suggested I shouldn’t charge users. I thought about it — the logic was sound, but the comparison wasn’t right. I’m just one person, while NerdWallet and SmartAsset are large commercial operations. So I asked AI to find smaller-scale websites as references. One example was mortgagecalculators.org. This site shows up when you search for mortgage calculations, and from its design, it doesn’t look like it was carefully designed by a large team. AI also mentioned that this website was built by one person over twenty years. Its most valuable asset is the massive number of articles it has accumulated, all linking back to its calculators, so users naturally find the site through search — that’s the most valuable thing he built over those twenty years.
AI Quickly Exceeded My Knowledge
AI quickly went beyond what I knew. After the website was built, I had to think about next steps. The next step was getting more people to know about the site, so after discussing with AI, it recommended doing SEO — search engine optimization. I originally knew nothing about SEO, but AI suggested several approaches and we followed through. The problem was, during the process you don’t really know if the optimization is actually better than before. So I asked how to verify that the optimization was actually working. AI then recommended a Google tool where you can paste your website URL and it analyzes your search score. If the score improved after optimization, then AI’s suggestions were helpful.
Don’t Know How to Do Something? Just Ask AI
When we’re doing something unfamiliar, we can just ask AI: “What else should I be paying attention to?” I asked this question because while building the website, I wasn’t sure what tools were available or what areas needed attention — things like market research or monetization. I originally thought about putting up ads or a “Buy Me A Coffee” link, but AI suggested waiting until the website had a certain number of users before adding ads, since that would be better for user experience. Also, because this is a static website, my current operating costs are basically zero. I only spent $100/month on Claude Code’s premium tier (and I actually finished in two weeks), about $12 for the domain name, and my own time. So it’s fine not to know how to do something — just ask AI what you should do, let it give you suggestions, and then build from there. You don’t have to follow everything it tells you either. Instead, treat what it tells you as a starting point for learning, and then decide for yourself whether to act on it.
A New Way of Working with AI
When discussing feature design approaches with AI, it would often propose three options and estimate how long each would take — for example, one option is simple and only takes a few days, another is complex and needs a few weeks, and some might take months. Here’s a big catch: AI estimates based on how long it would take a person to execute, but with AI’s help, you can actually finish in a day or two at most. At first, I assumed the options AI said would take months were off the table. But after trying them, I found that even plans it estimated at months could be produced quite quickly — you just need to spend time iterating back and forth to get the results to match what you want.
This also connects to another point: in the past, deciding on an approach required extensive discussion. You’d discuss, decide on a direction, then an engineer would take the plan and build it, verify it, and the project would be done. With AI, I think we should flip this thinking — because the cost and time of building have become so short, we don’t need to decide on a direction before we’ve seen anything. Instead, we can build out several approaches to try, and then decide which one to use afterward. For me, this is a huge shift, because now the time cost isn’t necessarily in the building — it should be in making sure you clearly know which direction is actually better before committing.
Practical Tips for Interacting with AI
When adjusting the website’s UI, there’s a small challenge: if you try to describe to AI in words that “the formatting is broken here,” it’s actually really hard to articulate exactly where it’s broken and how. What you can do instead is take a screenshot, circle the problematic area, and tell AI “the formatting is broken here.” It will then clearly understand where the problem is and attempt to fix it.
Also, AI can actually operate software on your computer, such as a browser. The reason I needed this was that when I had a calculator, I wanted to write a tutorial explaining what each field means and what to fill in. Previously, doing this required manually screenshotting every single field — extremely tedious. But after I installed the Chrome browser MCP Server, AI could operate the browser through this server. I gave it one instruction: “This is my calculator, write a tutorial for it.” It then screenshotted each field on its own, explained the meaning and usage of each field, and produced a complete tutorial. This was something I never imagined before starting this project — it could actually generate a tutorial on its own, and it was about as good as what I would have written myself.
AI’s Limitation: It Has No Memory
While building this website, something happened: I wrote a blog post and wanted AI to publish it a week later. Because my articles and code are in the same repository, when I push to version control it automatically deploys to the website. AI told me it remembered and would set it to go live on a certain day next week. But then it went ahead and pushed the article to version control, and the article was published immediately. When I asked, “Didn’t you say you’d wait until next week?” it immediately apologized and said it didn’t remember. I later understood that if you want AI to remember something, you must make sure it’s written in memory.md, because that file becomes part of the System Prompt. If it’s not written there, AI has no memory whatsoever. Once you start a new conversation, it completely forgets any promises it made — this is something to be especially careful about.
Some people online say they can have Claude Code run against a spec for several hours and write the entire program straight through. I haven’t been able to do that yet, because while AI writes fast, it quickly hits points of uncertainty and needs you to step in, redefine the problem, and re-clarify requirements. You need to interact with it frequently — you can’t just set a spec and let it run for two hours without intervening. Though this might also be because I’m still figuring out the right way to use it.
There’s Always More to Do After You Finish
After completing this project, I quickly realized there’s so much more that can be done. For example, I built another small project: a browser extension that solves a problem I’ve always had — when browsing web pages with lots of images, the images always get cut in half. In the past, doing something like this required knowing what browser APIs are available. But with AI, you only need to have the idea, and let AI figure out what APIs to use and implement the functionality you want. It was done in under thirty minutes. The description and documentation were all generated by AI too. It’s currently under review in Google’s Chrome Web Store, and I’m hoping it’ll be available soon.
The Biggest Takeaway
The biggest takeaway from completing this project is knowing that we’ve reached a point where you should spend your time thinking about what you want to do, rather than worrying about whether it can be done. At least in the realm of software, the building itself isn’t necessarily the biggest problem anymore. Of course, if what you’re trying to do is very complex, the amount of human involvement will inevitably increase — that’s a given. But the speed is already so much faster than writing everything yourself. I’m curious, and I plan to take some ideas I’ve had in the past and build them all this way. Before, when I ran into a problem, I’d often think “I don’t know how to do this, so forget it.” But now you don’t have to settle. You really do have the ability to change the things that have been bugging you.
Check out the website: runyournumber.com