Files
music_library/on-claude.md
T
2026-03-08 19:47:43 +00:00

1.7 KiB

After a few weeks of daily use I feel like I'm getting consistently good results with Claude.

I jump between two applications (always Elixir):

  • the work one is very large, around 3500 files, and has a VERY structured set of skills, scripts, and generally speaking tooling to work with agents.
  • my own application to manage music records which is around 200 files, which has a much smaller setup

In my application, Opus 4.6 is excellent: most of the times it just one-shots the needed change. I believe it's because:

  1. The AGENTS.md file is predominantly generated by https://hexdocs.pm/usage_rules/readme.html, which pulls directives from dependencies.
  2. I only have two additional docs referenced in AGENTS.md:
  • 1 for architecture, which describes at a high level the application structure, who does what, how it's stored, etc. I created that by asking Opus to describe the architecture of th application in a way that would help it to maintain it and extend it. I reviewed it manually and tweaked it afterwards. It's updated automatically by Claude when it makes changes to the codebase.
  • 1 for project conventions, again generated by asking Claude to extract project conventions by analyzing the last 500 commits. Every once in a while, I ask Claude to check it and update it.
  1. The test suite is extensive, more than I would "naturally" do. This makes it that there usually is a very easy correspondence between files that get changed and tests to run or edit, and Claude leverages this consistently. This also means that some tests are maybe redundant.
  2. Only one skill, ui-development, which contains the auto-generated documentation from a UI toolkit I'm using.