Claude · Video

Stop Overloading Claude Code: Fix Your CLAUDE.md & Skills

A local example shows how to remove duplicated Claude Code instructions while retaining the rules that matter. Mathias examines CLAUDE.md imports, task-specific skills, other sources of context, and the evidence passed from research to writing, with a clear distinction between file cleanup and measured model performance.

What you’ll take away

  • Remove duplication only after identifying the constraints and project facts every session still needs.
  • Use the loading behavior of imports and skills to decide where task-specific detail belongs.
  • Pass checked claims, source dates, and open questions to the writing step instead of an unfiltered research dump.

Keep a useful project entry point

A small CLAUDE.md should still explain the repository, its important working rules, and the checks needed for a useful result. The episode’s local cleanup removes repeated material without treating a shorter file as the only goal. Keep instructions that prevent a recurring mistake, and inspect the edited file to confirm that meaningful boundaries survived.

Moving a file does not necessarily defer its content

Anthropic’s documentation explains that CLAUDE.md imports load at startup. Splitting a large file into several imported files can improve organization while leaving the context cost in place. Task-specific skills and references serve a different purpose. Choose their scope and loading behavior deliberately, and review other instruction sources that may still duplicate the same rules.

Give the next stage checked evidence

The research-to-script example asks what a writer actually needs: a supported claim, its source, the relevant date, and the limitation that should travel with it. A large collection of notes can obscure those distinctions. A compact handoff is useful when it preserves the reasoning a reviewer needs to verify the final wording.

Measure the claim you want to make

The episode shows local file operations and reports from an authored example. It does not record live Claude inference or a doctor-command session, and it does not measure speed or billing savings. The vendor’s reported prompt reduction remains an attributed report. Check current commands and version requirements before trying the diagnostic steps in your own installation.

Jump to a chapter

Source material

Use the original documentation for changing product capabilities and account availability.

Published episode and chapter list Claude Code memory and instruction loading Claude Code skills Claude Code commands Anthropic’s context engineering lessons

Continue with a guide

Keep exploring

Pass it on

Share something useful.

Choose where to share, or copy the link.