Building Agent Skills - M3U Playlist Generator
The need for a M3U Generator
I wanted to demonstrate a situation in which an AI Coding Agent helped me solve a specific problem. I have a unique music playback setup with Emacs and MPV. I use a package called EMMS in Emacs to act as the frontend client I wrote about it here. Recently I have been trying build simple playlists to match certain moods and situations while I’m working.
With EMMS in Emacs you can have a playlist in an M3U file. The M3U file is a relatively straightforward format but I was writing them by hand. So it occured to me, why not have an AI Coding Agent create something for me?
Building an M3U Generator
I started by planning my idea of an M3U Generator. I began by writing down (yes by hand on a sheet of paper 🤓) what I envisioned and some simple logic.
In case you can’t read my handwritting, I am basically stating what I envision the workflow for generating the M3U playlists file would look like. I then continue on and write down the basic logic which the AI Coding Agent will build.
I outline the logic like so:
- Find songs in a text file
- Take songs and add M3U info
- set 0 to song length - easy cheat to not guess songs length
- Name M3U file based on 1st line of text file
- Save M3U file to root of playlist directory
- If cannot save to remote directory save to local “music” directory
With this simple logic I wrote a prompt and asked Pi create a Skill.
I loaded another Skill which assisted in creating the new Skill I wanted. In my prompt I also asked the coding agent to first research and learn about the M3U format from Wikipedia. Then I gave it the simple logic I had written. In this example I updated the logic to work with org-mode files instead of plaintext.
Pi wrote the skill and some python code to handle the heavy lifting. It even decided it needed to run a test for me. I tested the Skill myself and it worked as expected.
Learning from this experience
The Skill worked but I still failed. Even with an extremely simple example the LLM failed to understand my intent. The outcome may have been sucessful but the “process” was not as I intended it. I never wanted Pi to write more files on disk that it needed to. In order for Pi to “successfully” complete the task it wrote new org-mode files and simply left them for me to clean up.
This taught me one thing Specs and Planning are crucial. Other AI Coding Agents solve this by baking in a “Plan” mode which may have caught this. But that is still leaving to much for the LLM to try and understand. To solve the issue from before I simply added extra instructions in my AGENT.md file.
Conclusion
In summary this was a great learning experience. It helped reinforce some of the ideas and practices that I have previously experienced. In order to progress from here I have to sharpen my teaching and planning abilities.
Thank You
If you enjoyed or found any of the content on my site helpful, you can buy me a cup of coffee or send some bitcoin âš¡ so I can continue to bring you amazing content for free!
You can Buy Me A Coffee
