Cursor Your Dream, Part 1: How to Move From Product Idea to First Prompt

Who is this article for? Founders who are used to relying on development teams, but whose current resources or budgets have forced some projects to sit in the drawer for years. Beginners who don’t feel confident coding the first version of their product and don’t have the budget to hire developers. Startups that are stuck trying to raise a round just to build a prototype or MVP. Founders who are used to relying on development teams, but whose current resources or budgets have forced some projects to sit in the drawer for years. Beginners who don’t feel confident coding the first version of their product and don’t have the budget to hire developers. Startups that are stuck trying to raise a round just to build a prototype or MVP. What makes this material different? I’m not a programmer, and I’m writing this with readers in mind who don’t have a technical background (although a technical mindset will definitely help). Over the last six months, I went all the way from not writing code at all to building two full-fledged products. Naturally, my understanding of code grew a lot during that time. Each of these services includes 30 – 40 database tables, around 50 components, and roughly 100,000 lines of code. In other words, these are not weekend demos or toy projects. The first service integrates with OpenAI to “make sense” of user-generated content. It uses a vector database, embeddings, and operations built around them. The second service includes an internal accounting system with double-entry bookkeeping, idempotent operations, and an integrated payment system for processing online payments. I’m not describing some magical way to build software without touching code at all. But this material shows that it is possible to build a real product without first spending years learning programming languages. You will still need to look at the code, occasionally edit it manually, and keep things under control — but you will be able to do it. I’m not a programmer, and I’m writing this with readers in mind who don’t have a technical background (although a technical mindset will definitely help). Over the last six months, I went all the way from not writing code at all to building two full-fledged products. Naturally, my understanding of code grew a lot during that time. Each of these services includes 30 – 40 database tables, around 50 components, and roughly 100,000 lines of code. In other words, these are not weekend demos or toy projects. The first service integrates with OpenAI to “make sense” of user-generated content. It uses a vector database, embeddings, and operations built around them. The second service includes an internal accounting system with double-entry bookkeeping, idempotent operations, and an integrated payment system for processing online payments. I’m not describing some magical way to build software without touching code at all. But this material shows that it is possible to build a real product without first spending years learning programming languages. You will still need to look at the code, occasionally edit it manually, and keep things under control — but you will be able to do it. Introduction Introduction The idea of fully automated development is very popular right now: describe a product in text and get a working application. Some services can even produce a simple prototype. But when we are building a product to test real hypotheses, the expectations are different. We want proper UI/UX, good performance, caching, scalability, and the ability for real developers to work with this code later. From my experience, even strong models like Opus handle backend logic and database structure quite well, but they struggle much more with frontend architecture. They often generate unnecessary code, break the system into components poorly, and ignore client-side performance. That’s why I don’t use fully automatic generators. There is no magic here: if even Opus inside my Cursor can confuse the frontend, the same thing will happen in any other product built on top of LLMs. In this series, I use a different approach: an IDE with built-in AI agents. The AI helps write code and plan tasks, but the architecture and key decisions remain under human control. Why did I choose Cursor? Simply because I read that it is built on top of the popular VS Code editor. It seemed like a reasonable choice. I don’t think Antigravity would be worse; in practice, the result will be roughly the same, because the quality of the solution mostly depends on the model being used. The editor itself is just a shell that gives the model access to your project files and environment. Antigravity The story began in the summer of 2025, when I was discussing with DeepSeek an idea that had been in my head for almost 10 years: a social network of preferences, where users receive recommendations based on similarities in their personal ratings (restaurants, hotels, countries, hobbies, basically anything). And suddenly it said: 2025 DeepSeek 10 years social network of preferences Let’s build it. We should use Next.js. Let’s build it. We should use Next.js. Let’s build it. We should use Next.js. Let’s build it. We should use Next.js. 3 months later, I found myself presenting the product to the top management of a large corporation, and they were the ones who asked me to do it. What’s funny is that, despite 20 years in IT as a founder, that had never happened to me before. More recently, I also brought another long-postponed project to launch, one that had been sitting in a folder for years. I always assumed it would be too complex and expensive to build, and the topic itself felt risky enough that I didn’t want to invest money into it. In this series, I’ll share the entire experience: real prompts, configurations, and pitfalls. But before we go further, we need to answer the most important question: Should you even try this? Should you even try this? Should you even try this? Short answer — yes, if you roughly understand these concepts: yes frontend backend databases HTML structure CSS styling JavaScript behavior basic React concepts frontend backend databases HTML structure CSS styling JavaScript behavior basic React concepts If these ideas already make sense to you, feel free to scroll down to the "Choosing a Framework" section. If these ideas already make sense to you, feel free to scroll down to the "Choosing a Framework" section. If these terms sound unfamiliar, it’s better to take a moment to understand them now. What follows won’t feel like scrolling social media, but consider this: this journey saved me roughly $300,000 and a couple of years of development, hiring contractors, and building a team. understand them now $300,000 and a couple of years of development Frontend vs Backend In simple terms, the frontend is what the user sees: frontend a website in a browser a mobile application a browser extension any interface where users click buttons or read information a website in a browser a mobile application a browser extension any interface where users click buttons or read information The backend is the part of the system that runs on servers somewhere on the internet. backend It handles: data processing authentication business logic communication with databases data processing authentication business logic communication with databases When a user clicks a button, the frontend sends a request to the backend. The backend processes the request and sends data back. The interaction looks roughly like this: User → Frontend → Backend → Database → Backend → Frontend → User User → Frontend → Backend → Database → Backend → Frontend → User In practice, the frontend interacts with the backend through API routes: special URL addresses that the frontend sends requests to: API routes /api/articles /api/login /api/create-article /api/articles /api/login /api/create-article When the frontend calls one of these addresses, the server runs the necessary code: verifies the user queries the database processes the data and returns the result. verifies the user queries the database processes the data and returns the result. For example, the frontend might send a request like this: GET /api/articles?author_id=535 GET /api/articles?author_id=535 The backend receives this request, retrieves the articles from the database, and returns them to the frontend. What Is a Database? A database stores the information used by the application. If you’ve used Excel or Google Sheets, the concept is very similar. Think of a database as a collection of tables. Each table is like a spreadsheet. tables Users Table Users Table id

name

email

1

Alice

alice@example.com

2

Jack

jack@example.com id

name

email

1

Alice

alice@example.com

2

Jack

jack@example.com id

name

email id id name name email email 1

Alice

alice@example.com 1 1 Alice Alice alice@example.com alice@example.com 2

Jack

jack@example.com 2 2 Jack Jack jack@example.com jack@example.com Articles Table Articles Table id

name

email

1

First article

2

1

Another article

3

2

Jack's article id

name

email

1

First article

2

1

Another article

3

2

Jack's article id

name

email id id name name email email 1

1

First article 1 1 1 1 First article First article 2

1

Another article 2 2 1 1 Another article Another article 3

2

Jack's article 3 3 2 2 Jack's article Jack's article In this example, one user can create many articles. This simple structure of related tables forms the foundation of most software systems. In your own service there might be 30–50 tables, even if right now it feels like there will only be three. related tables 30–50 tables HTML: Structure of the Page HTML defines what exists on the page. what exists on the page headings paragraphs buttons images containers headings paragraphs buttons images containers Example: Hello This is a paragraph

Hello This is a paragraph

HTML elements form a nested structure called the DOM. DOM div ├── h1 └── p div ├── h1 └── p Every webpage in your browser is essentially a large tree of elements. CSS: How Things Look If HTML defines structure, CSS defines appearance. CSS defines appearance colors fonts spacing layout positioning colors fonts spacing layout positioning Example: /* CSS class that defines how the title looks */

.title { color: blue; font-size: 32px; } /* CSS class that defines how the title looks */

.title { color: blue; font-size: 32px; } Applied like this:

Hello

Hello Combining Multiple Classes One important concept in CSS is that elements can have multiple classes. multiple classes Example:

Submit

Submit Each class can control different aspects of styling. /* base button style */ .button { border-radius: 6px; }

/* button color style */ .button-primary { background: blue; color: white; }

/* button size */ .button-large { padding: 16px 24px; } /* base button style */ .button { border-radius: 6px; }

/* button color style */ .button-primary { background: blue; color: white; }

/* button size */ .button-large { padding: 16px 24px; } This flexibility is powerful. But it can also become confusing in large applications when styles start interacting with each other. Why Tailwind Helps This is one reason many modern projects use Tailwind CSS. Instead of writing large CSS files, Tailwind uses predefined utility classes. Tailwind CSS Example: Submit

Submit

Each class represents one small styling rule. This approach has a surprising advantage when working with AI tools like Cursor. AI tends to generate Tailwind layouts much more reliably than traditional CSS. Tailwind layouts much more reliably than traditional CSS If you don’t explicitly instruct Cursor to use Tailwind, it may start embedding styles directly into each element. If you don’t explicitly instruct Cursor to use Tailwind, it may start embedding styles directly into each element. Example:

This works initially. But later it becomes painful, because changing the design requires editing every component individually. Reusable classes allow you to change the design of the entire application by modifying only a few shared styles. JavaScript and TypeScript HTML defines structure. CSS defines appearance. JavaScript defines behavior. JavaScript defines behavior. Press enter or click to view image in full size JavaScript allows the interface to react to user actions. Example: // when the user clicks the button, show a message

button.addEventListener("click", () => { alert("Hello!"); }); // when the user clicks the button, show a message

button.addEventListener("click", () => { alert("Hello!"); }); TypeScript is essentially JavaScript with additional type safety. JavaScript: // a simple function that adds two numbers

function add(a, b) { return a + b; } // a simple function that adds two numbers

function add(a, b) { return a + b; } TypeScript: // the same function, but with explicit types // a and b must be numbers // the function returns a number

function add(a: number, b: number): number { return a + b; } // the same function, but with explicit types // a and b must be numbers // the function returns a number

function add(a: number, b: number): number { return a + b; } Most modern applications use TypeScript because it prevents many errors in larger systems. Why We Will Use Next.js In the rest of this series, we will build the example application using Next.js. Next.js is a framework built on top of React. React focuses on building user interfaces. Next.js Next.js adds things like: routing server rendering API endpoints project structure build tools routing server rendering API endpoints project structure build tools You don’t need to deeply understand Next.js internals. But it helps to understand how React applications are structured. You don’t need to deeply understand Next.js internals. But it helps to understand how React applications are structured. Fortunately, the core ideas are very simple. You only need to understand three things: components props hooks components props hooks React Components React applications are built from components. components A component is a reusable piece of interface. Examples include: buttons navigation bars comment blocks modal windows buttons navigation bars comment blocks modal windows Example component: // A simple React component. // It renders a button.

function Button() { return Click me } // A simple React component. // It renders a button.

function Button() { return Click me } Now it can be reused anywhere. Large applications become trees of components. trees of components Props Props are inputs passed into components. Example: // A simple React component. // It renders a button and shows the text passed in props.label.

function Button(props) { return {props.label} } // A simple React component. // It renders a button and shows the text passed in props.label.

function Button(props) { return {props.label} } Usage: // The same component can be reused with different labels

// The same component can be reused with different labels

Hooks Hooks allow components to store state and react to user interactions. Example: // useState creates a piece of state inside the component. // "count" stores the current value. // "setCount" is the function used to update it.

const [count, setCount] = useState(0) // useState creates a piece of state inside the component. // "count" stores the current value. // "setCount" is the function used to update it.

const [count, setCount] = useState(0) Now the component can update its internal state. // When the button is clicked, the count increases by 1. // The text updates automatically.

setCount(count + 1)}> Clicked {count} times // When the button is clicked, the count increases by 1. // The text updates automatically.

setCount(count + 1)}> Clicked {count} times Hooks make components interactive. Summary:

If you have never encountered these concepts before, I understand how this might feel. But you don’t need to master all of this at a professional level. You only need to understand the basic idea. If you have a technical mindset and can grasp these concepts, the journey becomes much easier. Instead of randomly asking AI to generate something, you will be able to guide it properly. And that makes a huge difference. Summary: If you have never encountered these concepts before, I understand how this might feel. But you don’t need to master all of this at a professional level. You only need to understand the basic idea. If you have a technical mindset and can grasp these concepts, the journey becomes much easier. Instead of randomly asking AI to generate something, you will be able to guide it properly. And that makes a huge difference. If you have never encountered these concepts before, I understand how this might feel. But you don’t need to master all of this at a professional level. You only need to understand the basic idea. If you have a technical mindset and can grasp these concepts, the journey becomes much easier. Instead of randomly asking AI to generate something, you will be able to guide it properly. And that makes a huge difference. Before Moving Forward Make sure you roughly understand: frontend vs backend databases as tables HTML structure CSS classes and styling JavaScript behavior React components, props and hooks frontend vs backend databases as tables HTML structure CSS classes and styling JavaScript behavior React components, props and hooks What’s next? Before diving into a technical specification for your product, answer this: What stack will you use? At first glance, this may seem like a purely technical decision. But when you build software with Cursor, the stack effectively becomes part of the instructions you give the AI. If you don’t define it in advance, the AI will start improvising, and I’d rather not even think about where that might lead. Cursor Here I’ll describe the stack I ended up using myself. This setup worked extremely well for me and may help you avoid some of the problems I ran into at the beginning. Choosing a Framework If your product has a web interface (meaning users interact with it through a browser), you’ll need a framework that handles both the frontend and the backend. For both services I built, I chose Next.js, and I never regretted that decision. Next.js I should mention that I had never worked with Next.js before. I was familiar with things like PHP, Laravel, JavaScript, and similar technologies, but in this case I simply followed ChatGPT’s recommendation. And it worked. The point is: even if you already know another stack, this might be a good opportunity to try something new. The point is: even if you already know another stack, this might be a good opportunity to try something new. If you’ve been far from development altogether, you may run into challenges with client-side caching and application state management (these are necessary if you want to fully avoid things like page reloads while a user interacts with your service). That’s where experience really helps. In the next parts, I’ll describe the main things I ended up doing. If you’ve been far from development altogether, you may run into challenges with client-side caching and application state management (these are necessary if you want to fully avoid things like page reloads while a user interacts with your service). That’s where experience really helps. In the next parts, I’ll describe the main things I ended up doing. In the end, this kind of polishing is definitely not required to bring a product to market. Once you have users and revenue, you can always hire a frontend developer to refine things further. Where Will Your Server Run? Once the framework is chosen, the next question is where the application will actually run. In other words: hosting. hosting Vercel Vercel For my first project, I used Vercel. Vercel This is a fairly natural choice when working with Next.js. Deployment (pushing your application to the server) happens automatically out of the box whenever you run git push. Next.js git push However, over time, I ran into an unexpected issue: in some regions, the application simply isn’t accessible. Because of this, the first project remained on Vercel, but for the second one, I decided to try something different. Fly.io Fly.io For the second project, I switched to Fly.io. Fly.io. Deployment there is a bit more complex than on Vercel, but it can be automated. In the end, I actually liked Fly.io even more. It feels more flexible for scaling if the project starts to grow. Vercel Interestingly, Fly.io doesn’t formally have a free tier, but if your monthly usage stays below $5, they don’t charge you anything at all. Choosing a Database Choosing a Database There are two main approaches here: The first option is to run your own database on your own server. This gives you full control and can be completely free. The second option is to use Database as a Service. The first option is to run your own database on your own server. This gives you full control and can be completely free. The second option is to use Database as a Service. Database as a Service For both of my projects, I chose the second approach because I wanted to move as fast as possible. I used Supabase. Supabase Supabase provides a managed PostgreSQL database along with several extremely useful tools. PostgreSQL Why Supabase Was Convenient Supabase supports vector storage. This became important for one of my projects where data was stored as embeddings generated with the OpenAI embedding-3-small model. As AI becomes more integrated into software products, having native support for vector data becomes surprisingly valuable. Supabase includes built-in authentication. Users can log in using providers like Google, GitHub, email and other providers. Supabase also provides file storage, which allows users to upload images and other assets without building separate infrastructure. Another very useful feature is support for RPC functions. In my second project I completely stopped allowing the backend to query database tables directly. Instead, everything goes through RPC functions. In practice, this significantly reduced code generation errors. Supabase also includes powerful security tools like Row Level Security (RLS) policies. Supabase supports vector storage. This became important for one of my projects where data was stored as embeddings generated with the OpenAI embedding-3-small model. As AI becomes more integrated into software products, having native support for vector data becomes surprisingly valuable. vector storage embedding-3-small Supabase includes built-in authentication. Users can log in using providers like Google, GitHub, email and other providers. authentication Google GitHub email and other providers Supabase also provides file storage, which allows users to upload images and other assets without building separate infrastructure. file storage Another very useful feature is support for RPC functions. In my second project I completely stopped allowing the backend to query database tables directly. Instead, everything goes through RPC functions. In practice, this significantly reduced code generation errors. RPC functions Supabase also includes powerful security tools like Row Level Security (RLS) policies. Row Level Security (RLS) Supabase Drawbacks Supabase Drawbacks The first issue I encountered involved the built-in authentication system. Specifically, Google login behaves inconsistently for users accessing the application through Safari on iPhone. I only discovered this after about two months of development, and in practice it turned out to be almost impossible to fix. In my second project I ended up using only part of Supabase’s authentication system. The second drawback is pricing. Supabase technically offers a free tier, but in practice, you may still end up paying around $25–40 per month, even if the application is still in development and has no users. It’s not a huge amount, but it’s something to keep in mind. The first issue I encountered involved the built-in authentication system. Specifically, Google login behaves inconsistently for users accessing the application through Safari on iPhone. I only discovered this after about two months of development, and in practice it turned out to be almost impossible to fix. In my second project I ended up using only part of Supabase’s authentication system. Safari on iPhone two months of development The second drawback is pricing. Supabase technically offers a free tier, but in practice, you may still end up paying around $25–40 per month, even if the application is still in development and has no users. It’s not a huge amount, but it’s something to keep in mind. $25–40 per month Cursor + ChatGPT: Two Different Roles For me, Cursor is the primary development environment. There are other IDE options, but Cursor worked very well in my case. Switching to something like VS Code and interacting with an AI assistant through the terminal is something I personally have no interest in trying. Cursor is the primary development environment The main downsides of Cursor are cost and frequent updates. cost frequent updates At some points, my spending reached $100 per day. Over time, I learned how to control those costs. Today, a realistic estimate is about $300 per month if you actively use powerful models. $100 per day $300 per month Cursor also offers a base subscription for around $20, but relying on the cheapest model is practically impossible, and the limits for paid models are quite small. $20 Why ChatGPT Is Still Important Even if you use Cursor, I strongly recommend having a ChatGPT Plus subscription. ChatGPT Plus My workflow usually looks like this: Cursor → writing and modifying code ChatGPT → discussing ideas and architecture Cursor → writing and modifying code Cursor ChatGPT → discussing ideas and architecture ChatGPT Inside Cursor, the model is optimized for code generation and stability. Outside Cursor, ChatGPT behaves more like a creative thinking partner. code generation and stability A simple rule: Don’t ask ChatGPT to write your production code Don’t ask ChatGPT to write your production code Don’t ask ChatGPT to write your production code It doesn’t make much sense. The same model inside Cursor, or especially models like Opus, will generate code much better. Using ChatGPT for Visual Assets Another unexpected use of ChatGPT was image generation. image generation My second project represents visual environments: digital parks with monuments, paths, and decorative objects. That required a large number of graphical assets. Instead of buying assets or hiring designers for every iteration, I generated most of the visuals directly with ChatGPT. Since image generation is included in the ChatGPT Plus subscription, this saved a surprising amount of money during development and was simply very convenient. Design Tools Many developers use Figma to design interfaces before writing code. Figma Personally, I barely used it. Cursor can generate reasonable UI layouts on its own, and I simply adjust them manually afterward. However, if you prefer designing interfaces first, Figma can still be very useful. Google is also experimenting with tools that convert design layouts directly into code. One example is Google Stitch, which can generate an initial implementation from interface designs. Google Stitch In theory, you can: design the interface in Figma generate templates with Stitch then convert those templates into React components design the interface in Figma generate templates with Stitch then convert those templates into React components Email and Small Infrastructure Pieces Most applications eventually need to send emails. For this, I used a simple service called Resend. Resend Later, you may also want to introduce caching using something like Redis, but in both of my projects, I postponed that step. Redis Redis If the architecture is designed properly, caching can always be added later. If the architecture is designed properly, caching can always be added later. Do You Need Docker? In both projects, I managed to work entirely without Docker. Docker Docker Vercel builds the project automatically (as I understand it, the service was created by the same team that developed Next.js). With Fly.io, you can use Docker to build the project locally on your machine, but you can also upload the files as they are and let the build run on their servers. At one point, something broke, and Fly.io stopped picking up the files and starting the build. The common advice online is either to change the server region in the Fly settings or switch to local builds (which is where Docker becomes useful). But I solved it differently, by using a specific flag in the Fly deployment command, so I didn’t have to do either of those things. You can also just ask Opus how to fix it with minimal effort (without changing the server or installing Docker). It will usually suggest a workable solution. Hardware Matters More Than You Think One unexpected lesson had nothing to do with programming. Your hardware setup matters Your hardware setup matters Your hardware setup matters Cursor works much better when you have enough screen space to see multiple panels at the same time. I initially started development on a small 13-inch tablet. Technically it worked, but it was extremely uncomfortable. 13-inch tablet Switching to a 27-inch monitor dramatically improved the workflow. 27-inch monitor Cursor can also be fairly resource-intensive, so having a reasonably powerful computer definitely helps. Cursor can also be fairly resource-intensive, so having a reasonably powerful computer definitely helps. Operating System For context, I ran this entire experiment on Windows. Windows However, nothing in this stack actually depends on Windows, except for the AI’s constant attempts to generate shell commands using && (which can be fixed with a few rules I’ll describe later). && The same setup works perfectly well on macOS or Linux. macOS or Linux Minimal Stack Summary If we simplify everything, my stack looks roughly like this: Framework: Next.js Database: Supabase Hosting: Fly.io IDE: Cursor Chat: ChatGPT Email: Resend Framework: Next.js Next.js Database: Supabase Supabase Hosting: Fly.io Fly.io IDE: Cursor Cursor Chat: ChatGPT ChatGPT Email: Resend Resend And that turned out to be more than enough to build real working products. real working products What Comes Next Before writing any code, you need to clearly describe the product. Most people using AI tools jump straight into coding. In my experience, that’s the fastest way to create a chaotic system. If you haven’t built a product with AI yet, honestly, I envy you. I still remember the feeling I had in the summer of 2025 when I first started experimenting with it. If you haven’t built a product with AI yet, honestly, I envy you. I still remember the feeling I had in the summer of 2025 when I first started experimenting with it. Now my advice may sound a bit strange: Stop for a few days. Stop for a few days. Stop for a few days. Stop for a few days. Take 3 - 5 days and resist the temptation to immediately install Cursor, Antigravity, or any other AI development platform. Because the quality of your thinking at this stage will determine how painful (or how smooth) the next weeks or even months of development will be. Every 10 minutes of thinking now can save hours of refactoring later. 10 minutes of thinking now hours of refactoring later Start With a Conversation At this stage, your main goal is to clarify your idea. The easiest way to do that is to open ChatGPT and start discussing it using voice input (not the full voice conversation mode, but the speech-to-text button). Your native language works perfectly fine. Just start talking. For example, explain: Just start talking. For example, explain: What your product does Who the users are What problem it solves What actions users will perform What your product does Who the users are What problem it solves What actions users will perform The goal of this conversation is not to generate code. The goal is to gradually discover the structure of the future system. During this discussion, you should also tell ChatGPT what technology stack you plan to use. For example: For example: Next.js TypeScript Supabase Tailwind Next.js TypeScript Supabase Tailwind And explicitly ask ChatGPT not to generate code yet. Explain that you’ll write the code later in Cursor, but you want ChatGPT to describe how it would architect the system. not to generate code I’ve personally seen several people spend months circling around an idea without ever asking the AI this simple question. Think About Data, Not Just Screens One of the most important parts of this stage is designing the data model. If you’ve never worked with databases before, imagine the system as a set of Excel spreadsheets. Ask yourself questions like: Ask yourself questions like: Can one user create multiple objects? Do objects belong to categories? Can categories contain multiple objects? Are there relationships between objects? Can one user create multiple objects? Do objects belong to categories? Can categories contain multiple objects? Are there relationships between objects? For example: For example: one user → many articles an article → a category… or maybe an article → multiple categories and tags an article → many comments that load on scroll comments → reactions one user → many articles an article → a category… or maybe an article → multiple categories and tags an article → many comments that load on scroll comments → reactions Another important question is user roles. Will everyone be equal? Or will some users have additional privileges? user roles moderators administrators moderators administrators You can add roles later, but it’s much easier to plan for them early. Ideally, your technical description should include: table names (articles, comments, reactions) column names (username, avatar, bio) relationships between tables table names (articles, comments, reactions) table column names (username, avatar, bio) column relationships between tables Even if you plan to evolve the product gradually, a clear structure from the beginning can save you a huge amount of refactoring later. In both of my projects, the database eventually ended up with about 30–40 tables. Here’s the structure from my second project: 30–40 tables Your project will probably end up with a similar number. But for the initial technical specification, even 10 core tables is enough. 10 core tables Job Queues The most difficult concept for beginners is usually job queues. A queue is a table that stores tasks rather than objects like users or articles. You need a queue when your system interacts with external services. job queues For example, imagine that after a user publishes an article, you send it to an external service for translation or analysis. You cannot reliably perform that operation instantly during article creation, external services may fail or respond slowly. You cannot reliably perform that operation instantly during article creation, external services may fail or respond slowly. The article itself must be created immediately (because that depends only on your application). But additional data can be added later through a background process. For example, your articles table might look like this: id

author_id

picture

title_en

text_en

title_es

text_es

32

535

one.jpg

Article 1

Hello

33

535

two.jpg

Article 2

Next part id

author_id

picture

title_en

text_en

title_es

text_es

32

535

one.jpg

Article 1

Hello

33

535

two.jpg

Article 2

Next part id

author_id

picture

title_en

text_en

title_es

text_es id id id author_id author_id author_id picture picture picture title_en title_en title_en text_en text_en text_en title_es title_es title_es text_es text_es text_es 32

535

one.jpg

Article 1

Hello 32 32 535 535 one.jpg one.jpg Article 1 Article 1 Hello Hello 33

535

two.jpg

Article 2

Next part 33 33 535 535 two.jpg two.jpg Article 2 Article 2 Next part Next part После создания каждой статьи, она попадает в очередь на получение перевода от внешнего сервиса: id

entity_type

entity_id

status

attempts

es_text_generated

1

article_text

32

processed

1

Hola

2

article_text

33

pending

0 id

entity_type

entity_id

status

attempts

es_text_generated

1

article_text

32

processed

1

Hola

2

article_text

33

pending

0 id

entity_type

entity_id

status

attempts

es_text_generated id id entity_type entity_type entity_type entity_id entity_id entity_id status status status attempts attempts attempts es_text_generated es_text_generated es_text_generated 1

article_text

32

processed

1

Hola 1 1 article_text article_text 32 32 processed processed 1 1 Hola Hola Hola Hola 2

article_text

33

pending

0 2 2 article_text article_text 33 33 pending pending 0 0 Как только задание в очереди выполняется, триггер автоматически дополняет информацию о статье в основной таблице статей id

author_id

picture

title_en

text_en

title_es

text_es

32

535

one.jpg

Article 1

Hello

Hola

33

535

two.jpg

Article 2

Next part id

author_id

picture

title_en

text_en

title_es

text_es

32

535

one.jpg

Article 1

Hello

Hola

33

535

two.jpg

Article 2

Next part id

author_id

picture

title_en

text_en

title_es

text_es id id id author_id author_id author_id picture picture picture title_en title_en title_en text_en text_en text_en title_es title_es title_es text_es text_es text_es 32

535

one.jpg

Article 1

Hello

Hola 32 32 535 535 one.jpg one.jpg Article 1 Article 1 Hello Hello Hola Hola Hola Hola 33

535

two.jpg

Article 2

Next part 33 33 535 535 two.jpg two.jpg Article 2 Article 2 Next part Next part Queues for external services in Supabase are usually handled through Edge Functions. Edge Functions In Supabase there are 2 different types of functions: RPC and Edge. RPC Edge Remote Procedure Call RPC is a function that replaces direct queries to your database tables. For example, instead of your server making a direct query to retrieve (SELECT) all articles belonging to a specific user (535): RPC SELECT SELECT 535 535 SELECT * FROM articles WHERE author_id = 535; SELECT * FROM articles WHERE author_id = 535; I strongly recommend asking your AI assistant to create a function like this instead: CREATE OR REPLACE FUNCTION get_user_articles(p_author_id bigint) RETURNS TABLE ( id bigint, author_id bigint, picture text, title_en text, text_en text, title_es text, text_es text ) LANGUAGE sql SET search_path = '' AS $$ SELECT id, author_id, picture, title_en, text_en, title_es, text_es FROM public.articles WHERE author_id = p_author_id ORDER BY id DESC; $$; CREATE OR REPLACE FUNCTION get_user_articles(p_author_id bigint) RETURNS TABLE ( id bigint, author_id bigint, picture text, title_en text, text_en text, title_es text, text_es text ) LANGUAGE sql SET search_path = '' AS $$ SELECT id, author_id, picture, title_en, text_en, title_es, text_es FROM public.articles WHERE author_id = p_author_id ORDER BY id DESC; $$; Then, inside your project code, you simply call this function like this: SELECT * FROM get_user_articles(535); SELECT * FROM get_user_articles(535); To make sure this approach always works correctly and consistently, add explicit rules to your technical specification (we will discuss it later): never query the database directly from the client; always go through server API routes forbid direct table queries inside the API; if a new data operation is needed and no RPC exists yet, create a migration with a new RPC function always use SET search_path = '' inside RPC functions and explicitly reference the schema when accessing tables never query the database directly from the client; always go through server API routes forbid direct table queries inside the API; if a new data operation is needed and no RPC exists yet, create a migration with a new RPC function always use SET search_path = '' inside RPC functions and explicitly reference the schema when accessing tables SET search_path = '' I won’t go deeper into this topic right now, but following these rules will save you a lot of headaches later. I won’t go deeper into this topic right now, but following these rules will save you a lot of headaches later. a lot of headaches later Edge functions Edge functions An Edge function runs outside the database and is typically used to interact with external services. outside the database They run on JavaScript or TypeScript (Deno runtime) and can call RPC functions inside your database when needed. JavaScript or TypeScript (Deno runtime) That’s it. This was the most complex section about the database. You might not fully understand it yet and that’s completely fine: First, at this stage these are secondary details. Second, during your discussion you can simply ask ChatGPT to find this article, “Cursor Your Dream,” and clarify anything you don’t understand. In my experience, it instantly understands the context and can explain queues, Edge functions, and whether you even need them. Third, as I mentioned in Part 1, Cursor is very good at building database and backend architecture. You won’t be left alone with these problems. First, at this stage these are secondary details. Second, during your discussion you can simply ask ChatGPT to find this article, “Cursor Your Dream,” and clarify anything you don’t understand. In my experience, it instantly understands the context and can explain queues, Edge functions, and whether you even need them. “Cursor Your Dream,” Third, as I mentioned in Part 1, Cursor is very good at building database and backend architecture. You won’t be left alone with these problems. Cursor is very good at building database and backend architecture Think About Components If you already have some idea about the interface, you can describe it in terms of components. For example: Article component User profile component Modal windows Navigation elements Article component User profile component Modal windows Navigation elements An article component might have several modes: view mode for readers edit mode for authors moderation mode for admins view mode for readers view mode edit mode for authors edit mode moderation mode for admins moderation mode If you think about components in advance, AI tools will later generate much cleaner code. An Important Rule for Styling Tell ChatGPT to include an instruction in the specification that Cursor should not use inline CSS styles, but instead use proper CSS classes, for example through Tailwind. Cursor should not use inline CSS styles In one of my projects I forgot to mention this. As a result, styling logic ended up scattered across dozens of components, and cleaning it up later was painful. A Critically Important Note About Authentication:

If you plan to use Supabase, specify that JWT tokens should be generated by your application, not by the default Supabase authentication system. If you’re new to authentication systems, this may sound strange. But this small detail can prevent serious problems later, especially when users sign in through Safari on iPhone. A Critically Important Note About Authentication: If you plan to use Supabase, specify that JWT tokens should be generated by your application, not by the default Supabase authentication system. If you’re new to authentication systems, this may sound strange. But this small detail can prevent serious problems later, especially when users sign in through Safari on iPhone. If you plan to use Supabase, specify that JWT tokens should be generated by your application, not by the default Supabase authentication system. JWT tokens should be generated by your application If you’re new to authentication systems, this may sound strange. But this small detail can prevent serious problems later, especially when users sign in through Safari on iPhone. Safari on iPhone Sample Case Study (from my second project) To make this more concrete, let me show you how this process actually worked for my second project, “Honoramma”. Honoramma Honoramma The First Prompt Here’s the very first prompt that started the discussion. For many years I’ve had the idea of building a digital memorial. It could be a platform where one group of users creates landscape maps, another places monuments, and others leave offerings — such as flowers or comments. I think it could be a breakthrough idea because nothing like this seems to exist. But I’m worried about legal issues. For example, what if someone creates a monument for a living person? Or creates a monument for someone like Michael Jackson — could that cause legal problems related to using the name? Technically I imagine it as a website with an isometric map where monuments are placed. Ideally we should find a good engine instead of building everything from scratch. For many years I’ve had the idea of building a digital memorial. For many years I’ve had the idea of building a digital memorial. It could be a platform where one group of users creates landscape maps, another places monuments, and others leave offerings — such as flowers or comments. It could be a platform where one group of users creates landscape maps, another places monuments, and others leave offerings — such as flowers or comments. I think it could be a breakthrough idea because nothing like this seems to exist. I think it could be a breakthrough idea because nothing like this seems to exist. But I’m worried about legal issues. For example, what if someone creates a monument for a living person? Or creates a monument for someone like Michael Jackson — could that cause legal problems related to using the name? But I’m worried about legal issues. For example, what if someone creates a monument for a living person? Or creates a monument for someone like Michael Jackson — could that cause legal problems related to using the name? Technically I imagine it as a website with an isometric map where monuments are placed. Technically I imagine it as a website with an isometric map where monuments are placed. isometric map Ideally we should find a good engine instead of building everything from scratch. Ideally we should find a good engine instead of building everything from scratch. At that stage the prompt wasn’t technical at all. It was simply a starting point for discussion. ChatGPT Response (shortened) The idea of a digital memorial is interesting because it’s not just another social network. Your concept is closer to a digital space of memory, where people interact through symbolic actions: visits, comments, offerings. The legal issue you mentioned — creating a monument for a living person — does exist. In some cases it could be interpreted as defamation or reputational harm. However, platforms usually manage this risk through product design, not through total prohibition. For example: — complaint and takedown systems— clear rules for user-generated content— framing memorials assymbolic spaces, not factual claims about death Technically, for an isometric map on the web it’s usually best to use an existing rendering framework. For fast prototypes, good options include: — Phaser for interactive tile-based scenes—PixiJS for flexible web rendering and UI integration If you later want real 3D monuments, you could switch to Three.js, but for an MVP a 2D tile-based approach is usually much faster. The idea of a digital memorial is interesting because it’s not just another social network. Your concept is closer to a digital space of memory, where people interact through symbolic actions: visits, comments, offerings. The idea of a digital memorial is interesting because it’s not just another social network. Your concept is closer to a digital space of memory, where people interact through symbolic actions: visits, comments, offerings. digital space of memory The legal issue you mentioned — creating a monument for a living person — does exist. In some cases it could be interpreted as defamation or reputational harm. The legal issue you mentioned — creating a monument for a living person — does exist. In some cases it could be interpreted as defamation or reputational harm. However, platforms usually manage this risk through product design, not through total prohibition. For example: However, platforms usually manage this risk through product design, not through total prohibition. For example: product design — complaint and takedown systems— clear rules for user-generated content— framing memorials assymbolic spaces, not factual claims about death — complaint and takedown systems— clear rules for user-generated content— framing memorials assymbolic spaces, not factual claims about death symbolic spaces Technically, for an isometric map on the web it’s usually best to use an existing rendering framework. For fast prototypes, good options include: Technically, for an isometric map on the web it’s usually best to use an existing rendering framework. For fast prototypes, good options include: — Phaser for interactive tile-based scenes—PixiJS for flexible web rendering and UI integration — Phaser — Phaser for interactive tile-based scenes—PixiJS for flexible web rendering and UI integration PixiJS If you later want real 3D monuments, you could switch to Three.js, but for an MVP a 2D tile-based approach is usually much faster. If you later want real 3D monuments, you could switch to Three.js, but for an MVP a 2D tile-based approach is usually much faster. Three.js After that the discussion continued. My Follow-Up Messages I’m fine with pseudo-3D. At one level, users create a map: they choose its size as a grid and fill it with grass, paths, and other elements through a visual editor. At another level, users place monuments that may occupy different numbers of tiles, such as 1, 2, or 4. There is also a third type of user: they don’t create maps or monuments, they simply leave offerings such as flowers. Each map should have its own unique URL. I’m fine with pseudo-3D. I’m fine with pseudo-3D. pseudo-3D At one level, users create a map: they choose its size as a grid and fill it with grass, paths, and other elements through a visual editor. At one level, users create a map: they choose its size as a grid and fill it with grass, paths, and other elements through a visual editor. At another level, users place monuments that may occupy different numbers of tiles, such as 1, 2, or 4. At another level, users place monuments that may occupy different numbers of tiles, such as 1, 2, or 4. There is also a third type of user: they don’t create maps or monuments, they simply leave offerings such as flowers. There is also a third type of user: they don’t create maps or monuments, they simply leave offerings such as flowers. Each map should have its own unique URL. Each map should have its own unique URL. I also want the map owner to receive a percentage of purchases inside their map. For example, if someone buys a monument there, I might share 20% of the revenue with the map owner. If someone leaves flowers at a monument, perhaps 20% goes to the monument owner and another 20% to the map owner. I also want the map owner to receive a percentage of purchases inside their map. For example, if someone buys a monument there, I might share 20% of the revenue with the map owner. I also want the map owner to receive a percentage of purchases inside their map. For example, if someone buys a monument there, I might share 20% of the revenue with the map owner. If someone leaves flowers at a monument, perhaps 20% goes to the monument owner and another 20% to the map owner. If someone leaves flowers at a monument, perhaps 20% goes to the monument owner and another 20% to the map owner. I’m thinking of building the project using Next.js + Supabase. Which pseudo-3D engine would be best? I’m thinking of building the project using Next.js + Supabase. Which pseudo-3D engine would be best? I’m thinking of building the project using Next.js + Supabase. Which pseudo-3D engine would be best? Next.js + Supabase Okay. Now expand the technical description, gather everything into a single document, and add an introduction explaining the problem and the product. Okay. Now expand the technical description, gather everything into a single document, and add an introduction explaining the problem and the product. Okay. Now expand the technical description, gather everything into a single document, and add an introduction explaining the problem and the product. Markdown (.md) File with the Spec Markdown (.md) File with the Spec After several conversations, the technical specification was ready. The same will happen for you. Make sure to ask for the final spec to be generated in Markdown format and copy it into a text file. Or have it create an .md file right away. You will definitely need this file at later stages. .md But instead of sending it straight to Cursor, ask yourself this: Which part of the product is the most important to you, or the most representative? This matters because your first prompt for Cursor may differ from the technical spec. We’ll get to that later. For me, that part was the map editor mechanics. So I decided to build only that first (without authentication, a database, or a backend). map editor mechanics My Final Prompt to ChatGPT Write a prompt for Cursor where the main goal is to create a map editor. For now it should work without authentication and without Supabase. Even if the map cannot be saved yet, that’s fine. The main goal is to feel the mechanics of the map before building the entire product. Ideally this should be a component that can later be integrated into a full system. During map creation we should define slots where users will later place monuments. Write a prompt for Cursor where the main goal is to create a map editor. For now it should work without authentication and without Supabase. Even if the map cannot be saved yet, that’s fine. Write a prompt for Cursor where the main goal is to create a map editor. For now it should work without authentication and without Supabase. Even if the map cannot be saved yet, that’s fine. map editor The main goal is to feel the mechanics of the map before building the entire product. Ideally this should be a component that can later be integrated into a full system. During map creation we should define slots where users will later place monuments. The main goal is to feel the mechanics of the map before building the entire product. Ideally this should be a component that can later be integrated into a full system. During map creation we should define slots where users will later place monuments. ChatGPT Response Below is the final instruction that eventually produced the working prototype in Cursor. final instruction You are a senior frontend / game engineer. We are building only a prototype of a map editor (isometric pseudo-3D) as an isolated component in order to explore the core interaction mechanics. Strict Constraints No authentication. No Supabase, database, Stripe, or any backend components. No additional product screens. The goal is only the interactive mechanics of editing the map, slots, and monument placement. Output Format Implement this as a self-contained React component (React + TypeScript) that can later be integrated into a Next.js application. The component must: *— maintain **local state*— expose a public API (props / callbacks) Technologies — React + TypeScript— PixiJS + pixi-viewport Core Product Model (IMPORTANT) When a map is created, the owner defines SLOTS — rectangular areas on the grid where other users will later be able to place monuments. In this prototype we implement both: — the slot editor— monument placement All data is stored locally. Entities (In-Memory) Grid — NxM tiles (grass, path) Slot — Rectangle defined as x, y, w, hkind: monumentid Monument occupies exactly one slotProperties: size(1x1 | 1x2 | 2x2)slotIdrotation(optional)id Prototype Functionality (Editor MVP) Generate an NxM map (default: 20 × 20). Isometric projection. Camera controls: — pan (drag)— zoom (3 fixed zoom levels) Tools — Paint grass— Paint path— Slot tool (create / delete slot, rectangle sizes: 1×1, 1×2, 2×2)— Place monument (monuments can only be placed on an existing slot of matching size)— Eraser (deletes monument or slot depending on what is under the cursor) Rules — Paths cannot be covered by slots.— Slots cannot intersect.— A monument can only be placed into afree slot of matching size.— Slots and monuments must never overlap. Selection Clicking a slot or monument selects it (highlight with a frame). A small info panel (React div) appears on the right showing: — type— coordinates— size— id Export / Import State (JSON, In-Memory) Button Export JSON → outputs current state into a textarea. Button Import JSON → loads state from textarea. UX Details Toolbar on top or left containing: — tool selection— slot size selector (1×1 / 1×2 / 2×2)— Export / Import buttons The cursor mode should be clearly visible (label showing current tool) Architectural Requirements (Important for Future Scaling) Separate model (grid / slots / monuments) from renderer (Pixi layers). Create Pixi Application and viewport only once. Do not recreate them on every state change. Use useEffect with proper cleanup: — destroy app— remove listeners— cancel animation frames Do not store Pixi objects in React state. Store them in refs. Handle container resize correctly using ResizeObserver. Attach pointer events to viewport/stage once, and remove them on unmount. Use depth sorting for isometric rendering (for example x + y or y) so objects overlap correctly. Task First propose a minimal file structure: mapEditor.tsxmodel.tsisoMath.tsrender.ts Then implement a minimal working prototype (TypeScript) so I can paste the component into an empty React / Next.js project and immediately see the editor working. At the end include a short list of “things to improve later” (maximum 8 items). Do not add anything beyond what is specified above. You are a senior frontend / game engineer. You are a senior frontend / game engineer. senior frontend / game engineer We are building only a prototype of a map editor (isometric pseudo-3D) as an isolated component in order to explore the core interaction mechanics. We are building only a prototype of a map editor (isometric pseudo-3D) as an isolated component in order to explore the core interaction mechanics. only a prototype of a map editor Strict Constraints Strict Constraints Strict Constraints No authentication. No Supabase, database, Stripe, or any backend components. No additional product screens. The goal is only the interactive mechanics of editing the map, slots, and monument placement. No authentication. No Supabase, database, Stripe, or any backend components. No additional product screens. The goal is only the interactive mechanics of editing the map, slots, and monument placement. Output Format Output Format Output Format Implement this as a self-contained React component (React + TypeScript) that can later be integrated into a Next.js application. Implement this as a self-contained React component (React + TypeScript) that can later be integrated into a Next.js application. self-contained React component The component must: The component must: *— maintain **local state*— expose a public API (props / callbacks) — expose a public API (props / callbacks) public API Technologies Technologies Technologies — React + TypeScript— PixiJS + pixi-viewport — React + TypeScript— PixiJS + pixi-viewport Core Product Model (IMPORTANT) Core Product Model (IMPORTANT) Core Product Model (IMPORTANT) When a map is created, the owner defines SLOTS — rectangular areas on the grid where other users will later be able to place monuments. When a map is created, the owner defines SLOTS — rectangular areas on the grid where other users will later be able to place monuments. SLOTS In this prototype we implement both: In this prototype we implement both: In this prototype we implement both: — the slot editor— monument placement — the slot editor— monument placement All data is stored locally. All data is stored locally. Entities (In-Memory) Entities (In-Memory) Entities (In-Memory) Grid — NxM tiles (grass, path) Grid — NxM tiles (grass, path) Slot — Rectangle defined as x, y, w, hkind: monumentid Slot — Rectangle defined as x, y, w, h kind: monument id Monument occupies exactly one slotProperties: Monument occupies exactly one slotProperties: size(1x1 | 1x2 | 2x2)slotIdrotation(optional)id size slotId rotation id Prototype Functionality (Editor MVP) Prototype Functionality (Editor MVP) Prototype Functionality (Editor MVP) Generate an NxM map (default: 20 × 20). Isometric projection. Generate an NxM map (default: 20 × 20). Isometric projection. 20 × 20 Camera controls: Camera controls: Camera controls: — pan (drag)— zoom (3 fixed zoom levels) — pan (drag)— zoom (3 fixed zoom levels) Tools Tools Tools — Paint grass— Paint path— Slot tool (create / delete slot, rectangle sizes: 1×1, 1×2, 2×2)— Place monument (monuments can only be placed on an existing slot of matching size)— Eraser (deletes monument or slot depending on what is under the cursor) — Paint grass— Paint path— Slot tool (create / delete slot, rectangle sizes: 1×1, 1×2, 2×2)— Place monument (monuments can only be placed on an existing slot of matching size)— Eraser (deletes monument or slot depending on what is under the cursor) Rules Rules Rules — Paths cannot be covered by slots.— Slots cannot intersect.— A monument can only be placed into afree slot of matching size.— Slots and monuments must never overlap. — Paths cannot be covered by slots.— Slots cannot intersect.— A monument can only be placed into afree slot of matching size.— Slots and monuments must never overlap. free slot of matching size Selection Selection Selection Clicking a slot or monument selects it (highlight with a frame). Clicking a slot or monument selects it (highlight with a frame). A small info panel (React div) appears on the right showing: A small info panel (React div) appears on the right showing: — type— coordinates— size— id — type— coordinates— size— id Export / Import State (JSON, In-Memory) Export / Import State (JSON, In-Memory) Export / Import State (JSON, In-Memory) Button Export JSON → outputs current state into a textarea. Button Export JSON → outputs current state into a textarea. Export JSON Button Import JSON → loads state from textarea. Button Import JSON → loads state from textarea. Import JSON UX Details UX Details UX Details Toolbar on top or left containing: Toolbar on top or left containing: — tool selection— slot size selector (1×1 / 1×2 / 2×2)— Export / Import buttons — tool selection— slot size selector (1×1 / 1×2 / 2×2)— Export / Import buttons The cursor mode should be clearly visible (label showing current tool) The cursor mode should be clearly visible (label showing current tool) Architectural Requirements (Important for Future Scaling) Architectural Requirements (Important for Future Scaling) Architectural Requirements (Important for Future Scaling) Separate model (grid / slots / monuments) from renderer (Pixi layers). Separate model (grid / slots / monuments) from renderer (Pixi layers). model renderer Create Pixi Application and viewport only once. Do not recreate them on every state change. Create Pixi Application and viewport only once. Do not recreate them on every state change. Pixi Application and viewport only once Use useEffect with proper cleanup: Use useEffect — destroy app— remove listeners— cancel animation frames — destroy app— remove listeners— cancel animation frames Do not store Pixi objects in React state. Store them in refs. Do not store Pixi objects in React state. Store them in refs. refs Handle container resize correctly using ResizeObserver. Handle container resize correctly using ResizeObserver. ResizeObserver Attach pointer events to viewport/stage once, and remove them on unmount. Attach pointer events to viewport/stage once, and remove them on unmount. once Use depth sorting for isometric rendering (for example x + y or y) so objects overlap correctly. Use depth sorting for isometric rendering (for example depth sorting x + y y Task Task Task First propose a minimal file structure: First propose a minimal file structure: mapEditor.tsxmodel.tsisoMath.tsrender.ts mapEditor.tsx model.ts isoMath.ts render.ts Then implement a minimal working prototype (TypeScript) so I can paste the component into an empty React / Next.js project and immediately see the editor working. Then implement a minimal working prototype (TypeScript) so I can paste the component into an empty React / Next.js project and immediately see the editor working. minimal working prototype (TypeScript) At the end include a short list of “things to improve later” (maximum 8 items). At the end include a short list of “things to improve later” (maximum 8 items). “things to improve later” Do not add anything beyond what is specified above. Do not add anything beyond what is specified above. not add anything beyond what is specified above. Don’t Rush to the Next Step Your task right now is to think about your idea. Discuss it. Refine it. In my experience, it’s also important to focus on the MVP — the minimal version of the product that is still sufficient for a user to understand what the product is and what core problem it solves. MVP Two successful startups in my life actually began as nothing more than demos, even simple presentations. Today, AI tools simply allow us to build real products instead of demos, and test them on real traffic much faster. real products instead of demos You should also think not only about the product itself, but about the path to market. path to market Several of my projects stalled exactly at the point where I had no good answer to basic go-to-market questions. go-to-market So before writing code, try to clarify two things: what the product is how it reaches its first users. what the product is how it reaches its first users. In the end, you should have two documents two The first one is the full technical specification in the form of a Markdown (.md) file. This document will live in the root of your project and will be continuously read and updated by the AI during development. full technical specification The second one is the first working prompt — the prompt you will send to Cursor as your very first task. And this will only be the beginning. As development progresses, you will end up writing hundreds of prompts like this. first working prompt hundreds of prompts like this Next In the next article (coming soon), we’ll finally move to the technical part: installing Cursor and generating the first real working code. And this is where everything becomes both very exciting and a little chaotic. very exciting and a little chaotic