[](
url
)After I figured out my direction, I used Claude to build a tarot site. In 40 days I opened 287 tasks and finished 222 PRs. I tried to rely on AI completely and drop the traditional way of doing engineering, so I only did two things: tell it what I want, and look at the page (nod or shake my head).
The biggest thing I learned in the whole process is that for building websites or apps today, AI is really good at imitating. You only need to give it a good example, a good thing from first principles, and it breaks it down faster than I do and copies it in more detail than I do.
So what's left for the human brain is just two things: knowing clearly what a good result is, and judging where a good result will lead. I feel using AI now is like being a boss. You can ask it for results directly, but only if you know yourself what a good result looks like.
Go through TrustMRR and look at sites that already make money
Here's one suggestion for everyone, especially if you're building for a global market: go through the startups listed on TrustMRR. Their revenue is read directly from the Stripe payment backend. So it's a library of answers the market has already verified, which means correct, good results. From first principles, a good result means it got something right. Whatever category you want to build in, I think it's worth going there and looking around.
TrustMRR's homepage. The revenue on it is read from payment backends
For my category, for example, I could look for fortune-telling, divination and other related apps, and study their site shape, homepage title, page structure, sitemap, llms.txt and structured data, and figure out how each step is written, so the model can learn from it and do the same. At the same time I let it analyze the business model behind them.
Browsing by category, the top earner in Entertainment is an AI fortune-telling and astrology app
Break down the sites that got it right
You can also look at the sites that are already verified:
- how their homepage and small tools are built
- how they pass information to AI, and whether they even have MCP hooked up
- how they organize information into a fact sheet (how many visitors, who the founder is, and so on)
- whether each project page has a Markdown version for AI to read, and whether they expose an MCP endpoint so AI can fetch it
Click into any project and the revenue, subscriptions and founder are on the same page
TrustMRR's own llms.txt, which is basically a fact sheet written for AI
Then you sum up these sites that got it right and break down their backlinks and traffic sources. Here I used Treg's API, which lets me pull DataForSEO data in bulk to understand the situation and see where they submit backlinks.
For example, my biggest benchmark is Tarotap, and I looked at how it runs with just one person plus AI: from first principles, if the site does E-E-A-T well and the result is something people want to share, links come back naturally. And if you write the page content clearly, Google's AI Overviews and Perplexity will often show that clear content.
Tarotap's homepage, the site I benchmarked against at the start
But all of this needs your own understanding and judgment. AI can quickly get you to 60 out of 100, but getting to 90 has to come from your own judgment and review.
My habit is to have AI generate a PDF that puts all the research together so I can understand it quickly, with clickable links in it, and when I don't understand something I keep asking. I go by my own product taste, and I push myself to really become a product manager by building the product.
The automation part
So that's how I use AI to build fast and train my product thinking, and next is the automation part.
Once the core judgment is done, what's left is repetitive work, and I recommend everyone use scheduled tasks. Claude Code, Codex, Grok and similar tools all support scheduled tasks now, and you can let AI think on its own about how to turn every workflow into something that runs on a schedule.
- Backlinks and indexing: Every day at a fixed time, AI uses Google and Bing indexing data to pick directories to submit to and looks for new submission opportunities. It also writes outreach emails to article authors, and after a human reviews them once, they can go out in bulk.
- Tracking and improving: Every week, check search clicks and see whether the new backlinks are working, and let AI do its own research and suggest what to improve. You can even have AI ask tarot questions every week, see which content gets cited, and then analyze how the cited sites structure their content.
- Social media: Pull the viral posts from tarot accounts, analyze why they went viral, and work out the posting structure and the kinds of good content on each platform. Make the assets in one batch, review them together, and publish on a schedule every week. In the same way, AI can find related posts by itself and draft replies for a human to look over.
This work looks tedious, but it can run fully automated. Every day at 12 noon, AI generates a PDF and sends it to Telegram to keep me in the loop. It sums up what happened on and off the site in the last 24 hours and lists what I need to handle, and it keeps a progress sheet in the background so I can look back anytime.
The 7 scheduled tasks my tarot site runs right now
Using AI now is like being a boss. AI keeps getting stronger, and you don't need to go back and forth with it in chat like an employee, you get the workflow running through automation instead. Your time should go into understanding the product and finding growth: for the parts you don't know, let AI get you to 60 first so you learn fast, pull it up to 90 yourself, and then let AI replicate it at scale.
I really like something Guo Yu said: the time when you could make outsized returns by selling shovels in AI was roughly before the first half of 2025, and that's over now, and from the second half on, the time of using AI to build compounding returns is coming. If you start building up from now, there's really a lot of hope.
Finally, if you're interested, you can come and play on my site. It's free, there's no sign-up, and card draws are unlimited, and through this link you also get an extra 50 points of AI credit.
deararcana.com, first pick what you want to ask, then draw
Top comments (0)