Optimizing Claude Code

I was chatting with one of my colleagues and I realized they weren’t getting the most out of the CLAUDE.md files for Claude Code.  I thought I’d take a minute to share four tips that have served me very well.

1. Create a hierarchy of CLAUDE.md files.  Your personal file is used for all projects, so it should have your personal work style and preferences.  The one in the top level of any project dirtree has project-specific information.  Then you can have them in lower level directories: I typically code microservices in a monorepo, so each of those directories has one specific to that service.
2. Make it modular. You don’t have to have everything in the CLAUDE.md, it can contain guidance to read other .md files.  Claude understands “modular documentation” so you can ask it to do this, creating a high level file with guidance on when to consult detailed doc files.  This saves you tokens.  Again, this can be hierarchical.
3. Regularly capture learnings from a “good session”.  When I see I’m getting close to the compaction limit in a particularly productive session, I use this prompt: “Review all of the relevant CLAUDE.md and other .md files and compare them to what you know now.  If you can improve them, please do so.”  This works surprisingly well, though over time the files get pretty long, which leads to my final tip.
4. Occasionally ask Claude to optimize the CLAUDE.md files.  Tell it to review the file and compact it for ready use but preserve all of the critical information.  This also works quite well.

I hope that helps.  If anyone has any other tips, I’d love to hear them.  I’m always trying to get better at driving LLMs!