AI Tools Tips
-
Copy example code into Cursor, and use comments to provide context.
-
In the Hong Kong region, it is acceptable to use DeepSeek or Mistral services, while it is not acceptable to use ChatGPT, Claude, or Gemini.
-
The API policy of a region is approximately the same as its app policy.
-
Use Cursor instead of Visual Studio Code.
-
There are still cases where you need to use Visual Studio Code, such as for git merge editor scenarios, where I still use
git config --global core.editor "code --wait". -
Starting the day from Deepseek V3 release, we don’t need to sunbscribe any AI tools.
-
Use Gemini or Grok to generate festival celebration images with prompts like “Generate a happy Lunar Snake New Year image with text names included”.
-
In some cases, even when providing original text to AI models to create a table, a few places in the output may differ from the input. For example, when using the Deepseek V3 model in Cursor to generate a table of pip list, it may include versions like
1.极狐0. Here,极狐refers to the Chinese GitLab platform. -
When using Deepseek or Mistral API to translate titles with prompts like
You are a professional translator. You are translating a markdown file for a Jekyll blog post from English to Chinese. {text}, it can lead to incorrect translations. Besides the text you provide, the output often includes excessive translation. -
Though sometimes AI models in Cursor give partially correct text, we can accept them, as we can add follow-up instructions that will make the AI models regenerate the correct parts.
-
Avoid providing excessive context to large language models if it’s unlikely to be helpful. For example, when generating conversational dialog lines, avoid providing 100 points on a topic. Large language models already contain vast amounts of data.
-
When providing ample context for tasks like translation or generating dialog lyrics, avoid using chain-of-thought features, as it can be slow and lead to verbose or unhelpful responses.
-
One way to test whether a chatbot can follow a user’s instructions is to ask it to explain something in English and then continue the input in Chinese, observing whether the chatbot maintains its output in English.
-
Instead of providing context to LLMs, finetune a model based on a large dataset of text or code, and then use the finetuned model.