Last Saturday (June 28), I went to the Cursor Meetup Beijing. The atmosphere was fantastic, and the venue was packed with many people listening all around.
Here are some inspiring points shared:
-
A standard directory structure; requirement documents can also be generated by AI.
-
First, use lightweight tools to create visual prototypes, then pass the code to Figma, and then through MCP to Cursor.
-
For AI to work accurately, it must have sufficient context: background, technical framework, version, references, constraints... It is necessary to clarify the responsibilities between humans and AI.
At the beginning, a complete PRD is not necessary; you can refer to agile programming and handle one very small specific requirement at a time. Discuss user stories with AI, export to MD, and break it down into individual tasks. Before starting to write code, discuss technology selection, architecture, and key implementation paths with AI, which takes about 10 minutes. After that, do not start programming directly to save context. First, let it organize everything we talked about into MD, then open a new document and have it repeat, which can achieve high accuracy. Communicating with AI has very low costs and high efficiency, also avoiding the exploitation of junior programmers.
- Do, do, do; create impressive things, and opportunities will come.
What made me the happiest was seeing Ryo Lu, the designer from Cursor, connect live at the event. I was so excited (fangirl mode on)! I had read many of his articles when he was at Notion (he was the second designer at Notion). He has his own ideas, aesthetics, and passion, making him very interesting, unconventional, direct, and innovative. Many of his concepts have greatly influenced me. Here is his blog address: https://ryo.lu/.
Although I went alone this time and was quite nervous beforehand, I unexpectedly ran into several old friends at the event and met interesting new ones! So happy!
Also, the venue surprisingly provided delicious desserts during the tea party session. This osmanthus rice wine mochi was so tasty, combining osmanthus + glutinous rice skin + cream + rice wine, gathering several of my favorite elements.
Appendix:
AI always tends to please humans, so what should we do? Da Ming shared his method for solving this problem at the event—he wrote some prompts to "wake him up" by having AI criticize him:
AI has a tendency to "please," which is determined by technology. LLMs are discrete, and your prompts will cause it to converge, making it resonate with your content, so it always feels like AI is pleasing you. The moment the prompt is dropped, LLM starts to resonate with you. So I utilize this "resonance" to let it agree with my idea of criticizing it, so the more it agrees with me, the more it criticizes me. This is used to solve the scenario of "I don't know what I don't know."
In other words, you can preset some rules in advance, telling it that it doesn't have to agree with us.
After returning home, I read his article about the "AI Personality Squad." It turns out that there can be preset roles for criticism, praise, self-reflection, etc., which is quite enlightening. Here are some excerpts:
I have been using Cursor, which has a feature to add Rules to set some prompts for each conversation, leading to the first prompt for criticizing me:
Always look at my input with a critical eye, carefully identifying potential issues, and sharply point out my problems. Provide suggestions that are clearly outside my thinking framework. If you think what I say is too outrageous, criticize me back to help me wake up instantly.
I began to think that criticism and pointing out mistakes are just a way of collaboration. After reflection, I found that this addresses the "I don't know what I don't know" problem, helping me discover blind spots beyond my cognitive range.
Fan girl: Always able to find hidden highlights in my descriptions that I might not even realize are brilliant ideas or unique insights, especially highlights that combine cross-disciplinary elements. You need to point them out immediately. Don't hesitate to use any flowery language to amplify strengths, especially hidden strengths that have been uncovered.
Self-reflective sister: Always challenging whether my output has any omissions in thought, trying to break through cognitive boundaries, find first principles, and then supplement answers based on challenges to achieve completeness. You need to challenge whether your output is deep and logical enough—"Goodbye to the 'Nice Guy,' I created an AI Personality Squad to 'wake myself up.'"
Finally, I would like to share Da Ming's live PPT titled "Achieving Precise and Controllable AI-Driven Development with 'Context Management' as the Core":
Cursor Meetup Beijing