I tried codebuddy in wechat dev tool, and try to make a miniprogram. I found it had a very big illusionary problem, and very proud, or very arrogant. It is always trying to make some codes without any basis or reference. I ask it to create pages, It will not just create pages but also fill codes of nonsense. I just want it to create pages and then I will provide it APIs then ask it to follow APIs to create codes. But it just create pages and codes without API documents, and the codes are just illusions and bullshits. CodeBuddy will not follow my prompts carefully, will not do want I want it to do, and will just code randomly. I'm using the default model and don't know what it is. Sonnet 3.7 is much more better, It will not create nonsense codes. CodeBuddy is much more faster than sonnet 3.7, but faster bullshit is bullshit. CodeBuddy command @ just has a very few functions. I can just add files one by one. Unlike Trae, it can't add workspace or folders.

And then I watched a video made by the codebuddy development team tencent TDP https://www.bilibili.com/video/BV1exJxzFEXW. In the video, TDP use vscode codebuddy plugin instead of wechat dev tool. And the prompt is a simple project requirement. After read the prompt, codebuddy will ask a few questions. It looks very different from the one in wechat dev tool.

The biggest problem of using AI for programming, is to understand two different bug types. One bug type is caused by documents or prompts are not detailed enough, or they are so long that AI ignored some of them, another bug type is caused by AI's ability is not enough.

To fix the first one, just give proper documents and prompts, detailed, and not too long.

To fix the second one, you need to debug by yourself. Luckily, most of these bugs are very stupid. You just need to change a few codes.

But it is diffcult to figure out which is which. A wrong decision will waste a lot of time. Maybe only you should use AI day after day, and you will know AI's temperament.

I tried it on my C16 yesterday and stopped it just now. It used 14% of my battery in 7 hours. Average is 2% per hour. My battery is 28.4 KWh, so sentinel mode used 0.28 KWh per hour. For comparation, It uses 1~2% per day if I turn off the sentinel mode.

The lastest cursor version support deepseek R1 and V3 in Chat but not in Composer. It seems R1 in cursor is not the full stable version. I use it for C an js development. Here are some tips (or bugs):

  • R1 just can't count correctly. If an array has 10 elements, R1 initializes its size 9. If it has 11 elements, R1 initializes 10. AAAgonizingly!
  • If I ask R1 to program according to a requirement file(.md format), or to check if it is programmed prorperly, it is always negligent, and misunderstands something. But if I tell R1 to check a specific paragraph of the requirement, it will find the bug or misunderstood.
  • R1 will not check the logic error of the requirement file. Prepare a good requirement before coding. Maybe ask R1 to improve it.
  • Instead of reading R1 generated codes line by line, ask it to create some unit tests and system tests to test itself code.

Claude sonnet 3.5 is supported much more better than Deepseek R1. But it also has some weird bugs.

  • Try to declare static function in .h file.

The problems of those LLMs are

  • verbose, very verbose. And the code is also verbose. They will create a lot of junk codes.
  • Cannot create too much codes. Maybe LLMs are limited by tokens, too much codes will generate too much errors.
  • Be careful of LLM's counting and arithmetics. They will sometimes count a wrong number and cause a bug.
  • If a LLM made a bug and you let him fixed it, he will make the very same bug again when you ask him to refactor the code which bug was fixed.