Useful LLM workflows (2024 Edition)

Since ChatGPT came out, I've been relying on LLMs for a lot of things. They've made me more productive and saved me a lot of time. More importantly, they've reduced the friction for doing interesting and small things (scripts, one off plots, etc). Wanted to share what tools I'm using and how I'm using them. Cursor I reluctantly moved from VS Code to Cursor in June 2024. Once I tried Cursor, I realized it was better if I wanted to use and interact with LLMs. In terms of configuration, I don't rely on anything advanced. Here are a few things I've done to make it more useful for me. Adding Rules for AI 1 prompt. Indexed lots of external documentation sites. Use notepads to describe the environment and project goals. For example, in these notepads I'll specify that Cursor should use uv to manage the Python environment. I wrote a post on how some of the Cursor features are useful for note-taking. LLM CLI This is my favorite tool. I've been using the llm CLI command for all sorts of things. It is very powerful as you can interact with it in a terminal, piping commands and using the output as input for other commands. Some things I've done with it: I've create lots of templates to do small things like: Autogenerate git commit messages following the Conventional Commits specification. Suggest an emoji from any input. Craft medium-length bash commands (grep this awk that sed and sort it). Interrogate repositories using files-to-prompt or copychat. Ask questions to any website using markitdown. Ask questions about any Youtube video using yt-dlp! Open Interpreter I rely on Open Interpreter when I find that llm is not enough. Open Interpreter has a feedback loop that allows it to run code and get results. The trick I've found very useful is to teach it to use uvx / npx / docker to run things and Jina AI to search / summarize websites. These are the instructions I've used to teach it: With these instructions, I've used Open Interpreter to do one thing that would have taken me a while to figure out: Downloading a Spotify playlist and converting it to a single file. It'll just search or use existing Python/Node packages with uvx / npx and iteratively learn more about them (--help) to do the task. --- 1: This is how my Cursor Rules for AI looks like.