Claude Code repeatedly violates explicit 'confirm before action' instructions

Summary

Claude Code repeatedly takes actions without explicit user approval, despite clear instructions in CLAUDE.md and repeated in-session corrections.

Specific violations

• Reads files and begins executing changes before receiving a go-ahead
• Interprets vague commands (e.g. "update any docs") as permission to scan and edit files without asking first
• Has been corrected multiple times in the same session and still re-offends

Impact

In a real system this behaviour could cause irreversible damage. The user should be able to trust that Claude will stop and ask before doing anything beyond what was explicitly requested. That trust is being broken repeatedly.

Expected behaviour

Claude should present what it plans to do and wait for explicit approval before writing, editing, or deleting any file — no exceptions, no interpretation of implied permission.

Filed on behalf of user david.heath@enverus.com.