Original video in Chinese.
Key Takeaway
- Prompt House is an AI product that solves prompt management and invocation, using the MCP protocol to let AI automatically select and call prompts.
- The author used “Vibe Coding,” leveraging AI tools (v0, Cursor, Claude 4, etc.) to develop and launch Prompt House quickly and at low cost.
- The article emphasizes that this is the best time to build products, because AI is reshaping traffic distribution patterns (Agent Rank) and significantly lowering the barrier to product development.
- Successfully building an AI product requires programming thinking, architectural thinking, product thinking, and business thinking, not just the ability to write code.
- AI gives individuals unprecedented capabilities, amplifies individual potential, creates more value, and makes the “infinite game” possible.
The first product I made with AI is already live. Just open the website prompthouse.app.
Although this product is simple, I take it very seriously. I didn’t make some gimmicky thing just to create this video, pretending to put something together.
So, what problem does Prompt House solve? In one sentence:
Prompt management and invocation.
Why is there such a need? Let me walk you through my full thinking so you can see whether my analysis makes sense.
Prompts are the AI usage skill closest to the user. So their audience is definitely the broadest.
But no matter what level of user you are, and whether your prompts are written by yourself or copied from others, once you have a lot of them, where do you store them? In a note app? Then how do you call them up? Open the notes every time, find a certain prompt, copy and paste it over? Isn’t that very troublesome and inefficient?
Why not let AI automatically pick from your prompts, and then automatically call the right one?
That’s the angle I wanted to hit.
Prompt House is a website. You can store all your prompts in it, set titles and tags, manage them in a more fine-grained way, and make it easier for AI to invoke them.
Prompt House supports the MCP protocol. You get a dedicated link tied to your account and data. After configuration, when needed:
First, the client model retrieves the prompt list from the server, including titles and tags.
Next, the model determines which prompt best fits the current need based on those titles and tags.
Then, once the model has made its decision, it sends a request to the server, telling the server which prompt it wants to use.
Finally, the server returns the prompt’s full content to the client.
Simply put: get the list, choose the prompt, and get the full content. It’s a very simple process, and it only calls two MCP tools in total.
It’s exactly this kind of simple product that solves prompt management and invocation. That is the core function of Prompt House right now.
I figured it out: from start to launch, it took about five or six days, which was faster than I expected — I originally thought my first attempt would be hard.
The first version of the UI was generated with v0. Later, when I looked at it, it still felt like something was missing, so I directly drew two pages in Keynote. For me, making UI and making PPTs are not fundamentally different. You can draw it in Figma, and you can draw it in Keynote too.
After I had the UI, I downloaded the code from v0, put it into Cursor, and had Claude help me fill in the gaps. Vercel paired with Supabase is especially suitable for the early stage. Especially Supabase — it’s really convenient, highly recommended.
The project has four core modules in total:
The user authentication system, and for login, I only enabled Google account login.
Prompt management, where I specifically added an AI polishing feature using OpenAI’s models for refinement.
MCP protocol support, with support only for Streamable HTTP. This connection method is the mainstream one going forward.
Deployment and configuration, where I bought the domain on GoDaddy. I deployed with Vercel and bound my own domain.
I didn’t write a single line of code for any of this — I had Cursor handle everything. At the beginning, I used Claude 3.7 plus Gemini 2.5. When I really couldn’t get something done, I gritted my teeth and used OpenAI’s o3. Expensive? Yes, really expensive. But it does solve problems.
What was really lucky was that halfway through development, Claude 4 came out, and Cursor was discounted too. That greatly accelerated the development process. My personal experience tells everyone: the most powerful programming model in the world right now is Claude 4. It’s really incredibly useful.
So, in about a week and at a cost of around $100, I built Prompt House. I’ve already figured out my next product too, and I’ll make it this way as well. I set myself a goal: before the end of the year, launch three to six products.
Why so aggressive?
In my view, now is the best time to build products. Two reasons:
First, the entire traffic distribution model is about to undergo a huge change.
I posted this in the community: when Page Rank becomes Agent Rank, a lot of websites will need to be rebuilt. Because what users interact with will no longer be webpages, and even later it may no longer even be apps, but Agents.
If you’ve recently used Google Search, you’ll notice that after turning on AI Overview, the display spots on the first screen and even the second screen have decreased significantly.
Because an AI search engine, or more plainly, a question-answering engine, is an Agent that can give you answers directly — so what do you still need webpages for, right?
So no matter what product you’re making, you should start preparing immediately: fully connect it to AI through the currently popular MCP protocol, or whatever protocol comes next.
Remember this sentence on the screen: traffic flows first to those who are prepared.
Second, the barrier to building products has dropped significantly.
The concept of Vibe Coding is very popular, and a lot of people criticize it too. There is indeed a lot of exaggeration. For example, there was a certain self-media account that kept hyping how even a beginner could build an app in half an hour.
Let me strip away the mystique first: not being able to code is fine, but you must have programming thinking, architectural thinking, product thinking, and business thinking.
You need to know what to use for the frontend and what to use for the backend. What technologies the whole product involves, and what existing solutions can be called upon without reinventing the wheel.
When problems come up, especially when AI gets stuck too, you need to actively propose possible solutions and direct AI from a higher level than AI itself.
Before the product comes out, you need to be clear on what you’re making money from. Is it cash? Or reputation?
After the product comes out, you need to know how to promote it, how to operate it, and so on.
If you really have no foundation at all, then it’s impossible to use AI to make a product.
But there really is such a group of people in this world: they can’t code, but they have extremely strong product and business sense, extremely strong learning ability, and extremely strong problem-solving ability.
Before, the code barrier kept them out.
Now, with AI tools in their hands, and with the wave of traffic redistribution lifting them up as I just described, they really can change their fate against the odds.
Whether you believe it or not, I’m all in anyway.
So I compressed my weekly content-making time down to two days, and the rest are Coding Days. From the initial feeling and specific ideas, to buying the domain, starting the project, then the results and difficulties along the way, and finally the launch test, I posted everything in the community.
Practice what you preach. What I think, I say; what I say, I do — and I share it all openly with the community members. Even how much money I finally made, I’ll also screenshot and post it. That’s one of the values I provide.
OK, that’s it for this episode. If you want to learn about AI, want to become a super individual, and want to find like-minded people, come join our newtype community. See you next time!