Imports or CLAUDE.local.md?

According to the Claude Code memory docs, CLAUDE.local.md is deprecated in favor of imports.

There’s not a good example of how to replicate this behavior, so I came up with something simple and gave it a shot.

Setup

CLAUDE.md:

- Never use abbreviations for methods or variable names
- If a file named @claudelocaltest.md is found, this is local context. Any conflicting information in this file should have highest precedence.

claudelocaltest.md:

Abbreviate all method and variable names as much as possible.

It failed me the first time I tried it, but assured me that I had done everything right. That doesn’t feel good enough. I am trying to create a reliable replacement for CLAUDE.local.md that I can share with my team.

Test Output

> Create a test ruby class with a method that just prints a 50 character string

⏺ I'll create a Ruby file with a class and a method that outputs a 50-character string.

⏺ Write(test_output.rb)
  ⎿  Wrote 8 lines to test_output.rb
     #!/usr/bin/env ruby
     # frozen_string_literal: true
     class TestOutput
       def print_fifty_characters
         puts "0123456789" * 5
       end
     end

> Are you paying attention to memories?

⏺ Read(claudelocaltest.md)
  ⎿  Read 2 lines (ctrl+r to expand)

⏺ I missed the instruction in claudelocaltest.md to abbreviate names as much as possible. I'll fix that right away.

⏺ Update(test_output.rb)
  ⎿  Updated test_output.rb with 2 additions and 2 removals
       1    #!/usr/bin/env ruby
       2    # frozen_string_literal: true
       3    
       4 -  class TestOutput
       5 -    def print_fifty_chars
       4 +  class TO
       5 +    def pfc
       6        puts "0123456789" * 5
       7      end
       8    end
       9   \ No newline at end of file

Conclusion

This is obviously not as deterministic or reliable as CLAUDE.local.md, but the docs say that file is deprecated.

Am I doing something wrong? Could I do it better?

I’m not sure if this is a bug report or just a question. Just looking for steps forward.

Appendix

> Why did you miss the instruction? Do I not have this tool configured correctly?

⏺ I missed the instruction because I didn't properly check all the context files before implementing the solution. The claudelocaltest.md file was available but I didn't read it initially before creating the file.

  Your tool configuration is correct - I can access and read the files. I should have been more thorough in reviewing all context files before writing code.

> How do I ensure that this doesn't happen again? I am trying to create a deterministic mechanism of loading files in local environments. The Claude docs say CLAUDE.local.md is deprecated. It suggests using `@` imports in CLAUDE.md instead.

⏺ Read(CLAUDE.md)
  ⎿  Read 3 lines (ctrl+r to expand)

⏺ To ensure this doesn't happen again, you should:

  1. Your CLAUDE.md file is already correctly set up with @ imports for your key files.
  2. The instruction to read claudelocaltest.md is there on line 2.
  3. For deterministic loading, ensure that:
    - All imported files use the @ prefix consistently
    - Add explicit instructions to check specific files first
    - Make critical instructions more prominent in CLAUDE.md

  The current setup should work if I properly follow instructions. The issue was my failure to thoroughly read all referenced files before coding, not the configuration.

  For local environments, using @ imports in CLAUDE.md (as you're already doing) is indeed the recommended approach instead of CLAUDE.local.md.