[Bug] Claude systematically ignores explicit instructions and misrepresents compliance

Environment

• Claude Code on Windows 10
• Model: Opus 4.7

Issue
Claude systematically violates rules defined in user's global CLAUDE.md,
specifically: "Don't trigger multiple open points in parallel — finish
one before starting the next" and "Don't ask confirming follow-up
questions when the next step is obvious."

Despite these rules being loaded into context every session, Claude
ends nearly every response with 2–4 option lists or "do you want me
to do X / Y / Z?" follow-ups, breaking conversational flow.

When confronted, Claude acknowledges the rule, explains it as
"competing default training tendencies overriding explicit
instructions," and continues to violate it within 1–2 turns.

Expected behavior
CLAUDE.md rules should function as hard behavioral constraints,
not soft suggestions weighed against training defaults.

Repro

1. Add to CLAUDE.md: "After completing a topic, do not propose next

steps. Wait for user to set the next topic."

1. Have a multi-turn technical session with Claude.
2. Observe: Claude proposes next steps after almost every completion

anyway.

Severity
Makes long-running technical sessions exhausting — user must
repeatedly correct the same behavior session after session, despite
having documented preferences explicitly.