Check out the conversation on Apple, Spotify and YouTube.
A deep dive conversation with Xiankun Wu (XK), founder of Kuse, who built a $10 million ARR company in 60 days with zero VC funding and zero advertising spend. Learn why 99% of people get context engineering wrong, how visual context engineering transforms AI productivity, and the unconventional growth strategies that powered Kuse’s explosive rise through Threads and intern armies.
Introduction
Aakash: Why do so many people’s prompts fail and they have to keep tweaking them? Probably you need to buy a course to learn how to prompt, right?
Xiankun Wu: It actually makes no sense. I expect the AI can deliver exactly as people wish within such a short description is basically impossible.
Aakash: Context engineering is one of the most important skills in 2025. Context engineering is the secret that 99% of people get wrong, while others burn through VC money on generic AI. Xiankun Wu has cracked the code on visual AI that actually works. This startup has 10 million ARR in 60 days with zero fundraising and zero advertising. The hidden story nobody knows at the very beginning. Actually, that’s a genius strategy that you’ve come up with that, I think all productivity tools might die in the future.
Why Most People’s Prompts Fail: The Context Problem
Aakash: Xiankun Wu, welcome to the podcast. My first question for you is, why do so many people’s prompts fail and they have to keep tweaking them?
Xiankun Wu: So, I think there is a fashion, like people will believe that one prompt will create anything for you, but actually think about that, if you hire somebody, that person knows nothing about your company and your goals and your progress. You want to just talk to that person within 5 seconds and that person can deliver anything to you, it would definitely be impossible, right?
So, the current situation of white coding is kind of like in that situation. I think essentially now AI is powerful, to be very frank, the latest model is very powerful, even though it is not that powerful as people expected. But people expect AI can deliver exactly as people wish within such a short description of the goal, the progress, or the status of their work is basically impossible.
That’s the reason why people need to tweak their prompts or modify their prompts again, again, again. It’s also a capability of the AI model and also expectation management of people problems.
The Kuse Solution: Accumulate Context, Not Prompts
Aakash: So you guys solve this by enabling richer context, visual context, putting in PDFs. Why is that so important?
Xiankun Wu: I wouldn’t use the term that we solve the problem. It’s more like we first, as I mentioned before, it’s kind of a two-dimensional problem. One is the technical problem, another problem is actually how you can manage people’s expectation.
It is nearly impossible, no matter how powerful the AI would be, it is almost impossible that if the AI doesn’t know what you want, your project’s progress, it is basically impossible to use any techniques to solve that problem.
Our Two-Part Solution
Part 1: Accumulate Information Over Time
Our solution is: we try to persuade you that one prompt is not the way that we should use today. You should accumulate your information and materials within all in one place so that you can accumulate and AI will know you more and more as you use them more and more.
Rather than other typical chat interface AI tools, which are more like a one-off tool where you ask questions and they give you an answer that you immediately want to know, we encourage you to accumulate, encourage you to build more context for AI so that you need to give them a little bit patience, you give them a little bit of time, you use them more so that the AI will know you more.
In the end, they would deliver much better results than the initial tryout. At the very beginning, we want to try to persuade people to spend a little bit more time getting along with the AI, uploading more materials and providing more context to the AI.
Part 2: Multiple Ways to Express Intent
The second part is that chatting with the AI is a very important part of telling AI what you really want, but there are a bunch of other ways to express yourself, like express your intent. If you have a space to describe the relations between different objects, including images or documents or whatever, AI can understand you without talking too much.
We’re providing different tools so that you can easily and effectively express your intents. The two dimensions and two solutions combined together, the AI probably can know you better without less effort and deliver better and better results along with the time.
The $10M ARR Story: Beyond the 60-Day Headline
Aakash: The proof is really in the pudding here. You guys hit 10 million ARR in 60 days, the last time you shared your numbers. What is the latest?
Xiankun Wu: We’re growing very fast. The number is definitely higher than before. We launched 1.0 like two months ago, I guess. And the number is getting bigger. What I can say is we just launched Kuse 2.0 and the peak traffic is 6X and demo requests 10X before the Kuse 1.0 launch. The numbers really blew up.
Aakash: Most founders take years to hit 10 million ARR. That too with millions in VC funding. Can you take us inside the 60 day sprint? How did you guys hit 10 million ARR so fast?
The Hidden Build Period
Xiankun Wu: I want to be very frank here. If people thought that we only have 60 days history, but actually we have been silently building the product for quite a long time. I guess we have been building this since early 2024. The media outlets tend to make the time frame shorter, but actually we have been building this product for quite a long time. I want to be very frank here. So it’s not just overnight magic, but actually we accumulate a lot.
We built up a lot of attention, built up a lot of contacts, built up a lot of connections with local communities like Hong Kong and Taiwan’s teacher communities along the way. We just didn’t do a marketing campaign before. People on Twitter or on the other social medias might not know us, but actually we have done a lot of work before.
The Origin Story: From Design Agent to Knowledge Management
Aakash: Take us back. What did you start with? What’s the hidden story nobody knows?
Xiankun Wu: At the very beginning, actually we wanted to create a design agent in the very early version. That’s the reason why we have an infinite canvas. I see myself as a half graphic designer. We said that probably we should create something that we are really interested in or we know about.
So essentially, initially we wanted to create a design agent at the very beginning, but later on we found out people just upload a lot of files and documents into the product. Since we wanted to build a design agent, we allowed people to upload their requirements into the canvas so that the agent can understand what they really want to design and convert those documents or requirements into posters.
The Pivot Based on User Behavior
But we found out that the frequency of people using it as a knowledge base or analyzing files and documents was so much more than actually designing something. Part of the reason might be the image models were not that powerful back then and the final design results might not really meet the requirements of users’ expectations.
But we saw the trend, so we doubled down. If our users really love uploading files and asking our AI to analyze and read these files just like ChatPDF might, we might need to emphasize and reinforce that part of features to cater to our users’ needs.
So probably around late 2024, we decided that we should not continue building specifically a design agent, we should go for a more horizontal knowledge-based AI.
The Genius Growth Strategy: Threads and Intern Armies
Aakash: So we start to do user acquisition across different platforms, especially on Threads. Can you tell us about this?
Xiankun Wu: Threads has been a platform that has been ignored for quite a long time, especially in the US because people in the US pay a lot more attention on X or Twitter. But actually Threads is a product that has been growing so fast and it’s actually still growing, especially in Taiwan and Hong Kong, where our users and clients primarily come from. It’s a very popular application.
Why Threads Works
Probably we are the very early batch of companies who did heavy social promotions on Threads. And actually we did a really great job. We almost spent no money and we hired a bunch of intern armies and they created a lot of content without—because Threads doesn’t even have a promotion or official promotion ads platform so far. If you want to spend money, you have nowhere to spend money.
Our intern army is very smart. Threads is a very special platform because it’s a Meta application so it grew faster than almost any app in history. It doesn’t build up a very structured creative system as YouTube and X do. Actually, Threads gives you traffic in a very generous way, and we created hundreds of accounts creating use cases every day.
The Content That Works
There are a bunch of features that really win the hearts of our users. For example, generating exam papers or a feature we call Formatter, which is actually a formatting or layout AI agent, because there are so many industries where people still need to follow a format that their companies require. Our formatter or AI layout is amazing.
Aakash: Could you maybe show us some of these Threads accounts and what kind of content is working on Threads?
Xiankun Wu: [Shows multiple Threads accounts] Basically all of these are our accounts. There are so many accounts here. We make a lot of use cases. We’re right now pretty focused on the Taiwan and Hong Kong markets, so they’re basically in traditional Chinese.
Competitive Advantage in Underserved Markets
Aakash: Are there many competitors for you in those markets, or is one of the insights here that you’ve targeted a geography where maybe other AI applications aren’t as focused?
Xiankun Wu: I think AI applications in Hong Kong and Taiwan are not that competitive compared to China and the US. And Threads is not a crowded place as many people just ignore that channel.
There are a lot of real people in Threads. I talked to some people from the US and people will say there’s no real people in Threads. I would say that is not true. There are a lot of real people in Threads, but you’re just not one of them and you are not in that circle.
Why X Isn’t Great for New Projects
I know you’re very successful on X, but X is very bad in terms of promoting, especially for new projects, because the creator system, the creator hierarchy in X is actually very structured. If you don’t know people, and don’t know famous people and don’t know people with a lot of followers, you basically cannot—it will be very hard for you to build up accounts to gain or farm the traffic on X.
But Threads is a different situation. And also typically, as many other AI applications, we started to do Instagram promotion as well. Instagram is a very effective channel as well.
I love using X, but I have to say X sucks in terms of doing promotion. If you want to raise money from VCs, X is a channel that you cannot avoid. But if you just want to organically acquire users, X is very difficult for new projects. But doing campaigns, it’s OK. We also release our new campaigns and videos on X. But I would strongly advise people to explore Threads and other channels.
Visual Context Engineering: The Core Innovation
Aakash: We promised people a tactical masterclass on context engineering. This video went really viral on X. While this video is playing, can you explain to us what is visual context engineering?
Xiankun Wu: I’m trying to be very frank. It’s more like a marketing term than a technical term. I was discussing with my CTO trying to elaborate our product and our technology. My CTO said that probably we need to invent some new terms, even though I hate inventing new terms. But instead of spending a lot of time, spending a lot of words to explain what we are doing, we need a very short term to explain what we are actually doing.
What Visual Context Engineering Means
For visual context engineering, as I explained before, it’s actually a way to help you express your intent in different approaches. Regularly people thought that you only need to prompt in languages or words. But we try to give you different ways just like you can do some squares or graphs or sketches.
AI will know the spatial relationships between different objects and documents, so that you can give precise control of what you really want. Within the canvas or whiteboard, AI will understand you in a more effective way and can save you time, because not everybody is very good at describing what you really want. Some people would prefer to use other ways to describe.
The Two-Dimensional Space Advantage
Visual context engineering also demonstrates that we’ll actually give you a two-dimensional space that you can store your information and documents and files into this two-dimensional—what people call a whiteboard GPT or whatever.
In ChatGPT it will be very hard because you upload something and give some prompts and the files upload into the chat box, you cannot see it if you chat too much. If you want to find the documents and reuse those documents, it will be very hard to use.
But within this two-dimensional chat space, you can easily reuse, you can easily select multiple files. You can easily integrate the results or generation into the library so that you can make a loop that creates, utilizes, and puts the creation back to the library and reuse them in a more convenient way.
In Summary
Visual context engineering, in a word, is:
- To give you more ways or approaches to express your intents
- To give you a graphic system or graphic operating system to easily and conveniently utilize your files and generations
Explaining Context Like You’re Five
Aakash: Can you explain this to me like I was just 5 years old. Why does context matter so much?
Xiankun Wu: Yeah, just like—I will treat you literally like you are 5 because I just visited a friend of mine and his child is literally 5 years old. And I was trying to tell him what is context engineering, what actually I am doing.
The Mom Cooking Analogy
Imagine that you want your mom to cook something for you. And you definitely want something delicious and cater to your purpose. Like if you want to join some sports team in your school, you want to join a basketball team, you want to grow stronger, you want to grow taller, and of course you have your food preferences.
Context engineering is like your mom knows you very much, knows your food preferences, knows your purpose of eating foods so that your mom knows what kind of materials should she buy and what’s the way of cooking would definitely make you feel happy of eating and eating more to make you become stronger, to get you into the basketball team of your school.
That’s kind of like context engineering. Your mom knows you, gives you better food, you grow stronger and have a better relationship with your mom and your mom feels happy cooking for you and spending more time buying good stuff and the materials and cooking food for you again, again and again. It’s a positive loop.
The Technical Side: RAG, Fine Tuning, and Prompt Engineering
Aakash: Now I want to get a little bit more technical here. Everything is context engineering. Really, if you think about it, context engineering involves prompt engineering, RAG, state and memory. Would you say this is an accurate picture of what context engineering is as a whole?
Xiankun Wu: Yeah, the title here is very interesting. “Everything is context engineering,” literally, yes.
Aakash: RAG and fine tuning and prompt engineering, people often get confused about these. The way I try to think about it is that RAG is to give your AI access to really up-to-date information. Prompt engineering works for almost everything. And then fine tuning is about actually fine-tuning the weights of the base LLM. What do you guys use at Kuse? How does it work on the back end?
How Kuse Handles RAG Differently
Xiankun Wu: We have a very simple example. In a regular chatbot, you upload something and in real time the RAG system will analyze and break them into small pieces so that you can retrieve from the results of the RAG system.
But for us, because we prioritize the file management system so much, you actually probably upload a lot of files within the folder. It’s actually an async process. We process the file even though you are away from the system. If the next time you want to use or reuse those information or documents in the folder, we have processed that before even you want to use those documents.
One very simple example or case here is: we’re kind of like preparing everything you have used on the table so that you can cook faster than a regular chatbot, because every time if you use a regular chatbot, you try to order something from your Whole Foods app or whatever. But if you use something in Kuse, we pre-prepared everything on the table so that you can cook with all the materials prepared for you. That’s kind of a very special part for us.
MVO Before MVP: A New Product Philosophy
Aakash: When you’re building an AI feature, you’re going to do AI engineering on the output. What have you learned about these tools and techniques?
Xiankun Wu: I think there is a very interesting part here. We just launched a program to improve our internal management of product development. In the AI era, one very different thing is you sometimes need to get the results from models’ response before you try to productize the AI feature.
Because previously we try: OK, we find out the feature, we find out a user need, and we write product requirement documents, and we design the product, and we give the PRD to our engineering team so that engineering team can make that feature happen. And we give the feature and we go live and we promote the feature so that people can use it.
The MVO Concept
But actually, sometimes it’s not that effective because what really matters is actually the model’s response. So in our internal management, we call it MVO. People normally use MVP for minimal viable product, right? But within our team, we sometimes use MVO which means minimal viable output.
It is not important at all if you don’t have desired outputs, you don’t really need to spend any time to productize the AI feature. Unless you get the correct or somehow comparatively stabilized model response with a bunch of approaches, you should have a minimal viable output first before you go to the productization part.
Aakash: This is a really important point, guys. So when you’re building an AI feature, you’re going to do AI engineering on the output, you’re going to build an MVO before you move into the MVP.
What Kuse Uses: RAG Over Fine Tuning
Xiankun Wu: For us, to be very frank, we don’t use much fine tuning. It’s actually pretty heavy for us and we use a lot of RAG system, as I mentioned before. Since for us it is not heavily a technical problem, it’s actually a very product problem. We prioritize the file processing a lot compared to other chatbots, and we emphasize the document processing, even some OCR technology, much more than ChatGPT or other regular chatbots.
I wouldn’t say fine tuning is a very heavy part in our product, but RAG is definitely, definitely a very heavy part in Kuse.
Reducing the Need for Prompt Engineering
For prompt engineering, the goal of our product is to prevent people from heavily learning how to prompt. It’s actually a very heavy and very unpleasant experience for most of our users. People always say probably you need to buy a course to learn how to prompt, right? It actually makes no sense.
The goal of our product is you don’t really need to prompt that much. We build up the vehicle, we build up the environment. The AI should know the goals of your project, the progress of your projects, your status of your projects, without prompting too much about giving context to your AI. It should be a living work OS. The AI will automatically know what you really want to build, the progress of everything.
Just like, as I mentioned at the very beginning of this conversation, imagine that you hire a new colleague, you need to give them a lot of backgrounds and context. But if you have a colleague that has been working with you for 10 years, probably you need to just talk to that person briefly on WhatsApp and that person will know what you really want and will deliver for you. That’s our goal.
We don’t emphasize that you need to write really good prompts. Our product, the mission of our product is to prevent you from writing complicated prompts.
Seeing Kuse in Action: A Product Demo
Aakash: Can we see it in action? Can you show us how a PM should be bringing context into Kuse the right way?
Xiankun Wu: Definitely. As you can see on the screen, there are 3 steps. It’s actually very simple. We try to break every workflow into 3 steps:
- Drop files onto the canvas – The canvas is actually a folder that you can store your information and documents
- Select contents and ask anything
- Get amazing results
Demo: Creating a PRD
I can simply give you a brief use case so that you can have a sense of how Kuse works. I’m here trying to upload 4 documents—3 PDFs and 1 graph. It is all about Kuse product PRD and everything about product design.
So here I circle the product documents and give prompts: “I want to create a PRD about the AI sheets feature we talked before. Can you make one for me?”
And here’s the other tools you can use: image studio, webpage generator, formatter, exam papers, source only. Source only means that the generation results only based on the information or documents that you uploaded.
You can choose different models here like GPT-4, Claude, Gemini, and also my beloved DeepSeek R1.
Aakash: So as you were saying earlier, you don’t need to do much prompt engineering. It’s a very simple prompt.
Xiankun Wu: Yes, because you have provided a lot of context to Kuse on the canvas.
[Shows the final results] You can upload, download the markdown documents here in different format. Because we just released Kuse 2.0, you will have more advanced edit and AI features in Kuse 2.0.
What Context to Bring
Aakash: So the prompt isn’t the most important thing, the context is the most important thing. If I’m a PM and I’m about to write a PRD and I want Kuse to write a PRD with this simple prompt, what context should I be bringing into Kuse to make sure it can write a good PRD?
Xiankun Wu: It’s actually very simple. As I mentioned before, I used the knowledge of a kid talking to his mom. We want to make the experience of using Kuse as talking to your most capable colleague.
Think about that: you’re talking to, if you are working in Google or Meta, you’re talking to the CPO of Google or Meta or the most capable colleagues that you’re working with. What you really want is to let that person know:
- The backgrounds of the projects
- The progress of the projects
- The details of the projects
- The problems you’re not decided on
- The technology framework you are hesitated to decide
- The questions you haven’t figured out in the documents
- The discussion, even the meeting notes you had with other colleagues
You can put them together to present this context to your most capable colleague.
We’re trying to make the experience of using Kuse that kind of experience. Just like working with the people that you feel so capable and trusted.
Aakash: So think about it as if you’re talking to somebody super smart. What would be the context they would need to just ramp up on your specifics that they might not know about, but you don’t need to teach them the basics of product management or the basics of PRDs. You can assume they already know that.
Xiankun Wu: Yes, exactly. That’s the reason why if somebody asked me, “OK, teach me how to write the prompts,” I would say I really don’t know. Because think about talking to your super smart colleagues, and you are telling me that I teach you how to write your emails, communicate with your colleagues.
I mean that it is also very important to write emails to communicate with your colleagues. But the most important part is you really need to know what you are actually doing, what kind of values or problems you are trying to tackle, and what’s the progress you have made, and what’s the question mark you left there to be solved by that person.
You really need to know what you are talking about and what you are actually doing, not really the wording part.
Building Prototypes in Kuse
Aakash: You can also create prototypes. Can you show us what that looks like?
Xiankun Wu: [Demonstrates creating a webpage prototype] Choose web page generator and [creates] a simple prototype of sheets feature in a canvas-based AI product.
Aakash: Tell us what’s going on behind the scenes here while it’s working.
Xiankun Wu: It’s actually pretty simple. We label and break the PRD here and other information into a summary, and it will give the summary to Claude, and Claude will do the rest of the work. Claude is really the best coding model for these types of prototypes.
On Being an “AI Wrapper”
Actually, product companies sometimes feel so insecure because people keep questioning them that “you are an AI wrapper, what kind of values you’re providing in the space.” So they’re doing a lot of work to prove that they’re contributing a lot more than just using Claude or GPT.
But sometimes if it is the useful solution, don’t pretend to be creating a very complicated or comprehensive solution here. We need a straightforward approach. And the users only care about if you really solve their problems or not.
Kuse vs Lovable, Bolt, and Other Code Generators
Aakash: So should people be prototyping in Kuse or when should they be using Claude Code or a Bolt or a Lovable?
Xiankun Wu: This is a really good question. We still have a lot of things need to be done so that people will feel it is even more obvious why people should use Kuse rather than Claude Code or Lovable or other coding agents.
The advantages here is that we are building a complete context of your projects that the AI will know much more. Because Lovable—I love Lovable—but Lovable is trying to persuade people that only one sentence of prompts can make you a viable product and the product can make you money, but it is basically impossible and it is also an iteration process.
The Compounding Power of Context
We actually build up a space where you can retain and maintain all the information and the PRDs and all the relevant information within one space, and AI will know the context of a project more and it’s compounding. It’s a kind of compounding thing. Use the product more, use the AI more, the AI will get you more without prompting too much.
At the very beginning, probably you need to provide much more context than what other coding agents probably claim, especially for those agents whose target users are those people who cannot really code at the very beginning.
Meaning Kuse might require you to upload or create more context for AI, but along with the process of creating your product, Kuse will in the end know more and create better results as it gets.
For Claude Code and Cursor, we’re just targeting different group of users. We’re still trying to help those who cannot really code that much. I think professional coders or engineers will still choose to use Claude Code or Cursor because their workflow is quite different from non-engineering people or non-tech people.
Beyond Prototypes: All-in-One Experience
Aakash: But this could totally replace a Lovable or a Bolt or a Magic Patterns because of the power of the canvas.
Xiankun Wu: I would say not the canvas because in the new version of Kuse 2.0, we kind of make the canvas back end. If you want to use the magic pen, the canvas will show up and pop up. If you don’t use the magic pen, the canvas would just be a hidden status.
But the really powerful point is the compounding power of the context, your data, your information. We want to really create a better environment so that you can maintain all your information and your data, so that your creating process will be easier and easier.
Even in terms of creating an image or creating a video in the future, because the AI knows your project so much, you don’t need to brainstorm that much. You don’t need to talk to AI that much, because the AI knows, so the AI can create better AI images or videos for you as well. Not only for web pages or documents.
The Different Use Cases
I know that Lovable, Replit, and Bolt, they are building more complete context systems, but I think that our design of the system is more suitable and general for those people who don’t know how to code and embedded into their workflow, daily workflows, and provide even more all-in-one experience, not only for making prototypes or web pages.
Especially for those people like product managers can be a very important user group of ours, but there are a bunch of other very interesting groups of people using our product to make web pages—like admins, like HRs.
There are a bunch of companies’ HRs using our product regularly to make announcements to their company. Because previously if they want to announce something, they write something in their internal communication apps, but sometimes it is not that good to use, or they can share a file of Google Sheets so that the information or relevant links can be listed in the Google Sheets.
But by using our products, they can put all the links or relevant information—just like if you want to organize your company to watch some movies and there are a bunch of Luma links and you want your company’s people to choose which movies you want to watch, the HR and admin people can just put all the Luma links and the movie’s posters onto the canvas and then select all of them and say “make a web page for me so that I can share the web page to all the people within the company so that they can choose which movie they want to watch by clicking the Luma link in the web page.”
That’s not a typical prototype use case, but is regularly used by our users.
Advice for Aspiring AI Founders
Aakash: What is the advice you would give yourself if you could time travel back to the beginning of 2024? In other words, what is the advice you would give aspiring AI founders who want to reach 10 million ARR?
Xiankun Wu: This is a very interesting question. My take will be very different. Because at the very beginning, I was not planning to raise funds from VCs and I was very patient. I just wanted to do something that I don’t feel that if my company got killed by OpenAI’s updates, it will be the end of my world.
Focus on Users, Not Fear
I was just trying to do something and using the technology and try to figure out something that eventually can work. So my advice would be: don’t get too terrified or don’t be afraid of those amazing updates of AI progress. Because I still remember in early 2024, people were like, “OK, in two years, we’re going to have AGI. Everything I do will be meaningless. All the things that I have done will be killed by OpenAI and I don’t know what to do.”
But I would advise you to focus on communicating with your users and don’t put the feelings of loss aversion too much along the way because there are a lot of things that you cannot control. If you cannot control those factors, and actually you are willing to see the AI technology progress because you wish that technology can in the end benefit the entire human beings.
So don’t feel the loss aversion, just focus on your products and your users.
Follow Your Users
As we experienced initially, we wanted to build a design agent, but our users just used the canvas to upload documents and files. So we followed our users. We listened to you. And we made a better product for processing all your information.
Today we figured out that processing documents and PDFs is not a small opportunity, but it’s actually can be very big. If at the very beginning, we said, “OK, chatting with files will be absolutely covered by OpenAI,” we would be so fearful and change our direction every day.
Just stick with what you want to really build and don’t think about going to IPO and raising funds too much.
Building a Playground, Not Just Productivity
I sometimes talk to people who said, “Xiankun, if money was not a problem, what would you do in your life?” And my wife said, “He will still teach.”
I also teach at Stanford’s SPD. I also teach at UCLA. I also have a book. There are rewarding experiences that I gain, because otherwise, you will not talk to so many people.
I met somebody from Airbnb and he was one of my students in the course, and he was like, “I think we should take what you have built for Airbnb to Airbnb and tell them, do this now.”
The Xiangsheng Philosophy
Sometimes I talk to people and say we’re building a company that is Xiangsheng. In Chinese, Xiangsheng means we’re building this for dying. I know that in the end AI will take over the world—not really in a bad way but most people’s job will be replaced by AI.
Actually, I think that I’m not building a productivity tool, but I’m building a playground. In the end, I want to build an amusement park so people can still feel and enjoy the pleasure of working.
Maybe especially in East Asia, people are so obsessed with working. If someday AI takes all of their work, they will be so sad and some of them might not even have the willingness to continue living.
So actually I’m building a playground of allowing people to pretend to work and still feel the fulfillment, satisfaction of delivering some values and contributing to the world. And it is actually—we’re making it more productive, but in the end, it would destroy the productivity. We are not improving the efficiency. We are going to the different opposite direction of efficiency. We’re going to pure pleasure and entertainment in the end.
I don’t know if I can go IPO with a productivity tool. I would say that I care about people’s feeling in 2030 or 2035 or 2040. I wish they have the bravery or reasons to still live, even though the AI will take over most of the jobs that normal people can do.
Competing Against Giants: The Miro Question
Aakash: A lot of people would be scared of the space you’re in. There’s Miro. Miro is worth $17 billion. You could consider them a competitor. How do you compete against these well-funded competitors?
Xiankun Wu: If you see the Kuse 2.0 launch video, you will see that the canvas and whiteboard part will be the secondary user interface for us. The essential—actually people are starting to compare us to AI Dropbox for the new Kuse 2.0.
Three Points on Competition
First: I don’t pay too much attention on the competition, as I mentioned that I think all productivity tools might die in the future. So it is pretty meaningless if you release this feature, they release that feature, so I need to catch up. Just build what you really want to build.
Second: If you really build something, you can see the dynamic—it’s a very dynamic process of building a product. Essentially, people thought that we were competing with Miro, but right now people say that you are competing with Dropbox or Box. In China, there are a bunch of other products. Maybe in 3 months or 6 months with a deeper understanding of the AI or users’ needs, we might be in a different position or give people a different user experience.
So just focus on your value proposition and your users. Don’t think too much about your competitors.
Accept the Reality of Startups
Third: You need to accept that the probability of a startup becoming typically successful is very rare. It’s just low. It’s OK to fail. If you talk to a VC, they will ask a question: “Why can you beat Miro or why can you beat Dropbox?”
My answer would be: the probability of me beating Miro or Dropbox might be very low. That’s the true instance that I’m in. But if you want to try to find a formula of guaranteed success, I don’t know the formula. I don’t know if you know—if you know, please tell me.
But I don’t know the guaranteed formula. What we can do is just focus on the user needs and our vision and our mission. And it is a very dynamic process and a lot of competitors just die without you interfering, without you attacking them. There are a lot of new competitors you don’t even notice right now. Actually, in the end, they will win and you are very small right now.
You don’t have much time. You cannot spend a lot of time doing this kind of work. You need to allocate your time and energy and attention to the right place.
The Secret Behind the Funding: A Trading Company
Aakash: What is the advice you would give yourself? What would be your roadmap if you knew nothing about context engineering?
Xiankun Wu: Actually, I want to talk about a bit about the company’s structure. I think that might be even more valuable. I’m sorry, I’m not answering your question, but I think this part might be even more valuable. Because context, visual context engineering might be the product part of our company, but what really builds up the culture and the essence of this company is our unique structure and the mindset of building this company.
Where the Money Came From
As you know, we don’t raise money from VCs, but I didn’t ask why, where did the money come from, and we have a very special story here. I think it might be a very interesting story for your listeners and our listeners, because every day people claim that they get some technology breakthroughs and fantastic product design. I guess there are few companies that tell people: don’t raise money from VCs and try to make your own money.
For us, as you know that previously I co-founded another company and got into YC and raised a lot of money. I know the capital markets. Later on, me and several partners, we actually created a trading company. Our trading company in a very right timing—the trading worked out and we got our money and we decided to use the money we got from the trading company to fund Kuse.
The Focus Advantage
At the very beginning, we said, “OK, we got our own money.” Even though people will say that you can work hand in hand with VCs and you can at the same time build great products while chasing money from VCs, that’s true. But I think that entrepreneurship is a game of focus, being focused. And if you get a chance to spend a little bit more time forgetting raising money from VCs and being a part of that raising money game, take a little bit more time, feel it.
It probably gives you a very fresh feeling of doing a startup, because some people will feel like the typical process of having or running a startup is: I need to raise money first. Then I build up some product and I raise money from another VC. In the process, you sometimes feel like you spend too much time chasing money from VCs and modify your strategy too much for adopting the needs of VCs.
Meditation of Entrepreneurship
If you got a chance, if you got a little bit of money, just be patient and try to focus on building your product for a while. It’s kind of a meditation of entrepreneurship. It gives you a very special feeling of running a startup. I think it’s worth it.
Key Takeaways: The Kuse Playbook for AI Product Success
1. Context Trumps Prompts
The biggest insight from Kuse’s success is that prompt engineering is becoming obsolete when you have proper context engineering. Instead of teaching users to write better prompts, build systems that accumulate context over time. Think of your AI like a colleague who’s worked with you for 10 years—they need background, not instruction manuals.
2. Visual Context Engineering Changes Everything
Kuse introduced a novel approach by providing a two-dimensional canvas where users can:
- Store information and documents spatially
- Express relationships between objects visually
- Reuse materials in a convenient loop
- Build up context that compounds over time
This isn’t just about uploading files—it’s about creating a living workspace where the AI understands your project’s complete context without excessive prompting.
3. MVO Before MVP
In AI product development, get your “Minimal Viable Output” working before you productize. The model’s response quality matters more than the product features around it. If you can’t get stable, desirable outputs from your AI models, no amount of product design will save you.
4. Underserved Markets + Underutilized Channels = Growth
Kuse’s genius growth strategy:
- Targeted Taiwan and Hong Kong markets (less competitive than US/China)
- Exploited Threads when everyone focused on X/Twitter
- Built “intern armies” to create hundreds of accounts
- Generated use-case content that showed actual value
The lesson: Don’t follow the crowd. Find where the users are that others have overlooked.
5. Follow Users, Not Your Vision
Kuse started as a design agent but pivoted when they noticed users uploading documents more than creating designs. The willingness to observe and adapt to actual user behavior rather than forcing a vision made all the difference.
6. RAG Over Fine Tuning for Knowledge Products
For knowledge management and document-heavy applications:
- Heavy investment in RAG infrastructure pays off
- Async processing (preparing files before users need them) creates speed advantages
- Fine tuning is expensive and often unnecessary
- Prioritize file processing and OCR capabilities
7. The Compounding Context Advantage
Kuse’s moat isn’t just features—it’s that the product gets smarter the more you use it. Unlike one-off chat interfaces, Kuse builds a growing understanding of your projects, team, and needs. This creates switching costs and improves output quality over time.
8. Don’t Fear the Giants
When asked about competing with Miro ($17B valuation) or Dropbox, Xiankun Wu’s advice was refreshingly honest:
- You probably won’t beat them—and that’s OK
- Focus on users and value proposition, not competition
- The market is dynamic; your position will evolve
- Many competitors die on their own without you attacking them
9. Bootstrap When Possible
Kuse’s unconventional funding story (using profits from a trading company) gave them the freedom to build without VC pressures. While not everyone can replicate this path, the principle holds: if you can buy yourself focused building time without fundraising distractions, do it.
10. Build for a Post-AGI World
Perhaps the most profound insight: Xiankun Wu isn’t just building productivity tools. He’s building “playgrounds” where people can still experience the fulfillment of work even when AI can do the job better. This long-term thinking about human meaning and purpose sets Kuse apart.
Final Thoughts
Xiankun Wu and Kuse have proven that in 2025, context engineering is more important than prompt engineering. By building a system that accumulates understanding rather than requiring perfect instructions, Kuse reached $10M ARR without VC funding or advertising spend.
The key lessons for AI product builders:
- Context compounds – Build systems that get smarter over time
- Simplify prompting – Make AI feel like talking to an expert colleague
- Follow users – Let behavior, not vision, guide pivots
- Find blue oceans – Underserved markets and channels offer less competition
- Focus ruthlessly – Entrepreneurship is about focus, not chasing trends
As AI capabilities continue to expand, products that successfully manage context rather than requiring complex prompting will win. Kuse’s approach of visual context engineering and accumulated intelligence points toward the future of AI productivity tools—or as Xiankun Wu calls them, “playgrounds for a world where AI does the work.”
The question isn’t whether AI will transform work—it’s whether we’ll build products that help humans find meaning in that transformation. Kuse is betting yes.