Be like Marie Kondo, but for your agents

Agents really like tidy workspaces, and I think you constantly have to keep fighting the entropy of the universe (well, for now maybe just the entropy of repositories and computers) to maintain the tidiness.

So what does this mean in practice? Minimize:

• number of skills installed for your user/globally on your machine
• number of skills installed in your project
• number of plugins installed (for example in Claude Code)
• number of MCP server configured in your agent harness
• how up-to-date and long your AGENTS.md is
• how much documentation markdown files, README.md, etc. you have in your repository

Every token that is in the context of the LLM will distract it and makes it’s performance worse. Every redundant markdown file that “documents” stuff that is self-evident from your code will probably distract your agent + will eventually be out-of-date (this is a really big problem).

I think I am now cleaning up something of the aspects listed above at least once a week.