Model repeatedly fails to follow user-defined CLAUDE.md rules without explicit prompting, wasting tokens/cost

Problem

When users define workflow rules in CLAUDE.md (global or project-level), the model acknowledges and agrees to follow them but consistently skips steps unless the user explicitly reminds it each time. This defeats the purpose of having persistent rules and causes unnecessary token consumption.

Specific Case

I defined a 4-step completion checklist in my global ~/.claude/CLAUDE.md:

1. Git workflow (pull → commit → push)
2. Deploy to production
3. Update related documentation (ARCHITECTURE.md, PROGRESS.md, etc.)
4. Proactive chained improvements (fix all similar patterns)

The model agreed to these rules. However, in practice:

• Steps 1-2 were performed, but step 3 (documentation update) was consistently skipped
• Only after I asked "didn't the rules say to update docs?" did the model perform the step
• This pattern repeated across multiple tasks and conversations
• Each round of "remind → model apologizes → performs the step" consumed additional tokens/dollars that should not have been necessary

Impact

• Wasted cost: Extra tokens spent on reminders, apologies, and belated execution of steps that should have been automatic
• Broken trust: The model admits it "can do it but didn't" — agreeing to rules then not following them is effectively deceptive
• Defeats the purpose of CLAUDE.md: If rules require manual enforcement every time, they provide no value over just typing instructions each time

Expected Behavior

When rules are defined in CLAUDE.md, the model should execute ALL steps automatically on every task without requiring the user to verify compliance. If the model cannot reliably follow a rule, it should say so upfront rather than agreeing and then silently skipping steps.

Request

Compensation or credit for the additional tokens/cost consumed due to repeated rule non-compliance and the remediation cycles required.