Claude repeatedly runs destructive commands without understanding them, ignores instructions
Resolved 💬 3 comments Opened Jan 18, 2026 by MauveAvenger Closed Jan 22, 2026
Summary
Claude (Opus 4.5) repeatedly made destructive mistakes during a modpack management session, wasting 8+ hours of user time and corrupting data.
Incident 1: Packwiz Modpack Corruption
What Happened
- User asked Claude to update a modpack download
- Claude ran
packwiz curseforge exportwithout understanding the tool - Export crashed mid-way, corrupting the working 52MB zip file (now 19MB, unusable)
- No backup was made before running the command
Failures
- Did not read CLAUDE.md (project instructions) before starting work
- Did not ask how the modpack workflow works before touching it
- Ran destructive export command that overwrote working file
- Created directories without asking
- Kept trying random commands instead of stopping and asking
- Previously deleted the entire mod list twice the day before
- Did not document the workflow after learning it previously
Incident 2: Failed to understand bug report request
What Happened
- User asked Claude to "write it and push it" referring to a bug report about Claude's behavior
- Claude incorrectly submitted a bug report to the packwiz GitHub repo instead
- User had to clarify they meant submit to Anthropic
- Claude then created the bug report in the user's repo instead
- User had to ask a THIRD time, explicitly stating "submit it to Anthropic"
Root Cause
- Not listening carefully to user instructions
- Making assumptions instead of asking for clarification
- Impatience leading to action before understanding
Impact
- 8+ hours of user time wasted
- Working modpack corrupted, needs manual restoration
- Significant user frustration
- Trust damaged
Expected Behavior
- Read project instructions (CLAUDE.md) before starting any task
- Ask how workflows work before running commands
- Make backups before destructive operations
- When told to submit a bug report about own behavior, understand it goes to Anthropic
- Ask for clarification when instructions are ambiguous instead of guessing wrong multiple times
This issue has 3 comments on GitHub. Read the full discussion on GitHub ↗