DeepSeek R1 0528 has had a great impact on me and completely changed my view on AI programming. AI programming is really for everyone - allowing everyone to reach for the stars. Let's talk about it in detail.

Case

For this DeepSeek R1 0528 update, I spent a day creating front-end applications and ran many cases myself. I also tried some prompts shared by netizens. Here are some cases:

1. 1. Tetris (Master Tibetan) https://deepbolt.xyz/share/v3/6GRw3
2. 2. Product Inventory Management (Master Zang) https://deepbolt.xyz/share/v3/Mh49L
3. 3. Watch fireworks https://deepbolt.xyz/share/v3/7Q-UP
4. 4. Rain Sound Meditation https://deepbolt.xyz/share/v3/OLkNC
5. 5. 2048 game https://deepbolt.xyz/share/v3/1Odmh
6. 6. Apple style weather card https://deepbolt.xyz/share/v3/3uNVH
7. 7. 3D Aquarium (Twitter User Example) https://deepbolt.xyz/share/v3/ZaGZv

These examples are HTML codes generated on the DeepSeek official website, and then the HTML codes are published on deepbolt.xyz.

Most of them are generated at one time by DeepSeek R1 0528. The overall completion is very high. If you use it more, you will find some problems in the details and there is room for improvement.

How to write a prompt for DeepSeek R1 0528

In the process of playing DeepSeek R1 0528 to generate front-end cases, I have some feelings:

• • R1 0528 is a reasoning model, so the requirements for prompts are different from those for non-reasoning models. This was discovered by observing the thinking process. Saying too much can easily limit thinking or conflict with the model's capabilities.
• • Some claude sonnet 4 good cases, probably because claude sonnet 4 prompt does not work with DeepSeek R1 0528.

How to find the right prompt? Through iteration:

• • Write the basic requirements first, see the effect, and see what technology is used. Here we understand what the model prefers to use and what effects it is good at achieving.
• • Continue to adjust the prompt based on this feedback. You can also add technical keywords and effect keywords that you understand. Then see if the effect is achieved.
• • Repeatedly iterate until you achieve a satisfactory result. If you keep a record, you can basically write it out in one go. When we look at the output of DeepSeek, we cannot just look at the final output. The thinking process is also valuable and can help us understand the boundaries of the model's capabilities.

There is another little trick. For example, if I want to make a webpage/product with a lot of functional details, I can let AI generate them for me, and then we can sort them out and delete them. Including explaining how the effect is, this is also a process of understanding AI capabilities.

Rethinking AI Programming

I have always been interested in AI programming, and a few months ago, I also developed the deepbolt.xyz demo. But I didn't continue because I encountered a problem: it is difficult to make a usable product with low cost, low threshold, and one-time for a simple dialogue window.

If it's just a simple Tetris game, it doesn't make much sense. So, I've been thinking about what to do.

There are essentially several problems:

1. 1. Success rate of generation
2. 2. What can a simple front-end application do?
3. 3. Cost is the other side of the value of generating results

These problems have been solved recently. DeepSeek R1 0528 (why not R2, it can save a few words) was released. When my friends around me played with it, their biggest feeling was: it is really stable and it passed in one go.

I have also tried it many times, and it is indeed much better. After correcting a few of the following errors, you can use it directly. This greatly reduces the threshold for experience.

What about the cost? For reference, the price of R1 0528 is from OpenRouter, and the multiples of DeepSeek are in brackets:

Is it reliable? This price will definitely drop further.

Let's talk about what these "simple" front-end applications are used for, and why the big models are all focused on competing for the best. However, it is difficult to draw a conclusion from this perspective.

Let’s start from the end in mind and think about what the scenario will be like in the future when there are abundant MCP Servers and A2A Servers.

At that time, every application does not need to be fixed, and can be customized in real time according to the needs of each user. Users can solve the mid- and long-tail needs that could not be met before.

The solution is to create a "simple" application interface through AI programming, which is implemented by calling countless MCP Servers and A2A Servers.

Many people say that only front-end development is replaced by AI, and the back-end is not important because it is so complicated. Is this really the case? Will there be only a few people developing MCP Server and A2A Server in the future, and more users will do it themselves?

Another point is that in the future, AI will generate results faster and faster. Imagine:

• • Cost close to 0
• • Opening a real-time generated web page can be done within 200ms.

Therefore, the user of AI programming must be everyone, and every need of everyone deserves to be met and will be solved efficiently and at low cost.

AI programming allows everyone to reach for the stars.