Personally, I keep a text file as a "debugging journal" and just append a line every time I try a new step. For example, this is what I wrote down after recently trying out a local LLM:

  llama.cpp
  - idk why there's so many llama versions to install on yay
  - i went with llama.cpp-bin, because it was built with libcurl and the first one i tried apparently was not
  - but i had to remove llama.cpp-git-debug from a previous installation
  - remember yay -Q | grep ... to check for installed packages
  - the cli interface changes; i ended up with --hf-repo ggml-org/qwen2.5... --hf-file qwen25....
  - the huggingface.com page probably has the most accurate and up-to-date instructions
  - my goal: fast, offline, generalized/automatic autocomplete
  - localhost:8080 to access web ui after running llama-server

They're quick notes, and they actually help me problem solve, not disrupt me. I'm casual about it, though. I'm not copying every input and output verbatim. I think the idea is to leave yourself enough breadcrumbs so that you can reproduce, grab screenshots, and copy error messages later when you're not in the zone. Hope this helps.

Also, note that I'm most likely to publish a blog post in the following days while the problem is still fresh in my mind. If I wait months or years, it's pretty much doomed to stay in /drafts forever.