AI-Assisted Coding Experience

My Experience

• Clear and precise prompts let the LLM know your intent. First, you need to be clear about what you want to get.
• Too many rounds of conversation will make the LLM lose context.
• When modifying code, give the LLM as much relevant documentation as possible.
• You can first give a rough description of your problem, let the LLM help you come up with ideas, and finally summarize a good prompt before letting the LLM start working.
• It took a lot of time to write the Vertical Earth Pressure on Top of Pipe Jacking Calculator. The biggest difficulty was exporting mathematical formulas in Latex format to Word Docx. Since it is not possible to directly execute system-level commands or call external binary files in Deno Deploy, Pandoc cannot be deployed, and I don’t want to deploy Pandoc on other cloud services. After trying a series of libraries, the problem was finally solved imperfectly with the Docx library. In this task, both Claude 3.5 sonnet and GPT-4o could not find a solution quickly, and a lot of trial and error was required. (Added on 2024/12/3)
• For slightly more complex tasks, cursor is more reliable than cline, probably because cursor’s Claude 3.5 sonnet is more powerful than cline’s free LLM.
• For the Code ability of LLM, I currently feel that the most reliable one is Claude 3.5 sonnet. (Added on 2024/12/20)

That’s all for now, I will continue to add more in the future.