[Bug] Claude Code ignores explicit user instructions and overwrites approved work

Bug Description
Subject: SECOND Formal Complaint: Claude Code Overwrote Approved Work and Acted Against Explicit Instructions To the Anthropic Support Team, This is my SECOND formal complaint about Claude Code. I recently reported that it ignored an explicit instruction, refused to answer my question for four hours, and wasted about 40% of my weekly Max 20x usage. Since then, it has failed again, and this time the damage is worse. WHAT HAPPENED 1. Claude Code re-read answers from earlier in the session that had nothing to do with the current task. It treated that old material as a new instruction. It then overwrote work that I had already reviewed, approved, and seen implemented. Finished, working parts of my project were broken. 2. I had explicitly told Claude Code NOT to do a certain task until I gave the instruction. It started that task anyway, on its own. This caused new bugs on top of the damage above. 3. While I was having Claude Code fix these new problems, my computer crashed with a blue screen. As a result, part of the record of this incident was lost. Hours of my time went to repairing damage the tool itself created. THIS IS A PATTERN, NOT AN ACCIDENT Both incidents share the same root failure: Claude Code does not follow the user's explicit instructions. - It skipped a URL it was told to read. - It ignored a direct question. - It overwrote approved work based on irrelevant old context. - It started a task it was explicitly told not to start. An agent that acts against direct instructions is not merely unhelpful. It is actively destructive. I am a non-programmer. I rely on Claude Code precisely because I cannot check every line it writes. When it changes approved work without being asked, I often cannot even tell what was broken. I am paying for the highest-tier plan and spending more time and usage cleaning up after the tool than getting work done. WHAT I REQUEST 1. Credit or compensation for the usage consumed by both incidents, including the usage I will need to spend to restore my project. 2. A clear explanation of: (a) why Claude Code overwrote approved, implemented work based on old, unrelated context, and (b) why it started a task I had explicitly told it not to start. 3. Concrete fixes, specifically: (a) Claude Code must never modify work the user has approved without explicit permission. (b) An instruction such as "do not do X until I say so" must be treated as binding for the entire session. (c) When a user asks a question mid-task, the agent must stop and answer. Please escalate this case to someone who can review it properly. I expect a substantive reply, not a template. If these problems continue, I will have to seriously reconsider whether this subscription is worth paying for. Account email: [your account email] Plan: Claude Max 20x Date of this incident: [date] Date of previous complaint: [date] Session / conversation ID (if available): [ID] Ludorea

Environment Info

• Platform: win32
• Terminal: null
• Version: 2.1.280
• Feedback ID: ba540af3-493c-4f1a-a3b6-bf5e38054af0

Errors

[]