Skip to main content

CLAUDE.md Supremacy

We have all likely observed that the contents of the CLAUDE.md are adhered to much more strictly than the user prompt.

Adherence Hierarchy:

  • CLAUDE.md instructions: Treated as immutable system rules that define operational boundaries
  • User prompts: Interpreted as flexible requests that must work within those established rules

Behavioral Differences:

  • Process execution: CLAUDE.md steps followed sequentially vs user prompts adapted and optimized
  • Persistence: CLAUDE.md context maintained throughout session vs user prompts contextual to the moment
  • Override behavior: User prompts rarely override CLAUDE.md directives vs CLAUDE.md consistently overrides user preferences

In light of the above I opt to dexterously describe my processes in my CLAUDE.md and simply use the user prompt to provide parameters for those processes or to steer the model. It has been fruitful to tactically flood the CLAUDE.md with as much context as possible relating to the steps it should follow.

Modular CLAUDE.md design and length management:

I tend to break CLAUDE.md into modules of functionality. To ensure maximum adherence I format the information in markdown ensuring Claude can see the boundaries between instructions and modules, it also helps prevent instruction bleed.

As you add more workflow systems to your CLAUDE.md, you may receive warnings about the CLAUDE.md size potentially affecting performance. This is not necessarily a problem if you understand your token budget. It has been more effective for me to front-load the context (including providing multiple examples and denoting which files he can read and which files he is forbidden to read) in CLAUDE.md rather than having Claude whimsically reading files which may or may not poison him.

Mechanic Benefits:

  • Higher instruction adherence: CLAUDE.md content treated as authoritative system rules
  • Consistent execution: Sequential process steps followed systematically
  • Context persistence: Instructions maintained throughout entire session
  • Reduced context pollution: Controlled file access prevents unwanted information contamination
  • Modular organization: Clear markdown separations between functional areas prevent instruction bleeding
System Thinking

This mechanic works best when you thoroughly understand the system you're building. By providing complete context upfront, you minimize Claude's guesswork, leading to better adherence, faster task execution, and token savings.