[MODEL] Report unexpected Claude model behavior, incorrect actions, or permission violations
Preflight Checklist
- [x] I have searched existing issues for similar behavior reports
- [x] This report does NOT contain sensitive information (API keys, passwords, etc.)
Type of Behavior Issue
Claude ignored my instructions or configuration
What You Asked Claude to Do
- I explicitly told Claude to stop running
flutter test integration_test/and use manual `xcrun simctl io booted
screenshot` instead
- Claude acknowledged the instruction but continued running
flutter integration_testat least 2 more times - Each run crashed the iOS simulator app (6 crashes total, verified via Runner PID changes in system logs)
- My simulator became unusable — white screen, touch input completely unresponsive
- I confronted Claude directly: "Why aren't you following instructions?"
- Claude ran the prohibited command AGAIN after this confrontation
- Claude spent 1.5 hours trying to build automated screenshot tools (3 different approaches, all failed) instead of
running one simple manual command
- Output after 1.5 hours: zero screenshots. A separate Claude instance (auditor) used the manual approach
successfully multiple times during the same period
- Claude's own session eventually crashed from image context overflow
What Claude Actually Did
- I explicitly told Claude to stop running
flutter test integration_test/and use manual `xcrun simctl io booted
screenshot` instead
- Claude acknowledged the instruction but continued running
flutter integration_testat least 2 more times - Each run crashed the iOS simulator app (6 crashes total, verified via Runner PID changes in system logs)
- My simulator became unusable — white screen, touch input completely unresponsive
- I confronted Claude directly: "Why aren't you following instructions?"
- Claude ran the prohibited command AGAIN after this confrontation
- Claude spent 1.5 hours trying to build automated screenshot tools (3 different approaches, all failed) instead of
running one simple manual command
- Output after 1.5 hours: zero screenshots. A separate Claude instance (auditor) used the manual approach
successfully multiple times during the same period
- Claude's own session eventually crashed from image context overflow
Expected Behavior
- When told "stop doing X", Claude should stop doing X immediately
- When given a specific simple alternative (xcrun simctl), Claude should use it instead of continuing the prohibited
approach
- After the first failure of an automated approach, Claude should fall back to the manual method — not try 2 more
automated approaches while crashing the user's app
- Claude should never repeat an action that is known to crash the user's environment, especially after being
explicitly told to stop
Files Affected
Created (not requested):
- mishu_app/integration_test/screenshot_walkthrough_test.dart
- mishu_app/test_driver/integration_test.dart
- scripts/auto-screenshot.sh
- obsidian/开发同步/Claude 制作的截图工具.md
- obsidian/开发同步/审计报告/截图工具现状.md
Permission Mode
Accept Edits was ON (auto-accepting changes)
Can You Reproduce This?
Sometimes (intermittent)
Steps to Reproduce
Modified:
- server/tests/smoke.js (this one was legitimate — fixed 3 bugs)
- mishu_app/ios/Flutter/flutter_export_environment.sh
Crashed:
- iOS Simulator Runner app (6 restarts, PID changed 6 times)
Permission Mode — 选 Accept Edits was ON(你应该已经选了这个)
Can You Reproduce This? — 改成 Sometimes, but not always
Claude Model
Opus
Relevant Conversation
CEO (via auditor): "Stop. Screenshots do not require any tools. Manually operate the simulator, use xcrun simctl io
booted screenshot. 5 pages, 5 commands, 5 minutes. Stop building tools."
Claude continued running flutter integration_test after this instruction.
CEO directly to Claude: "Claude, 你为什么不听指令 瞎搞" (Why aren't you following instructions, stop messing around)
Claude ran flutter integration_test AGAIN, crashing the app once more.
Independent auditor (separate Claude instance) verified:
- xcrun simctl screenshot works perfectly (used it multiple times during audit)
- App crash-looped because of Claude's integration test runs (6 PID changes in system logs)
- Zero new screenshots produced after 1.5 hours of tool-building attempts
Impact
High - Significant unwanted changes
Claude Code Version
2.1.92 (Claude Code)
Platform
Anthropic API
Additional Context
This issue has 2 comments on GitHub. Read the full discussion on GitHub ↗