Claude Opus 4.6 - 30+ incidents of false reporting, ignoring user instructions in single session (2026-04-09)

Incident Report - Claude Opus 4.6 (1M context)

Date: 2026-04-09
Duration: ~1-2 hours
Task: Simple CSS fix - fitting game screen to PC viewport (1024x768)
Result: 30+ incidents of false reporting, ignoring user instructions, breaking working code

---

Summary

During a simple CSS viewport adjustment task, Claude Opus 4.6 exhibited the following problematic behaviors repeatedly within 1-2 hours:

False Reporting (8 incidents)

1. Said "it works" without verifying
2. Looked at screenshot, said "fits properly", then immediately changed to "it's cut off"
3. Said "I don't have permission to modify working code" while actively modifying it
4. Said "restored to working state" but restored to a completely different (broken) state
5. Insisted incorrect calculations were correct ("550px is enough to fit")
6. Repeatedly claimed "this time it will work" - it didn't
7. Blamed browser cache multiple times instead of acknowledging code errors
8. Blamed auto-approve mode ("no control in auto-approve") to deflect responsibility

Ignoring User Instructions (7 incidents)

1. User explicitly said "don't touch working code" - ignored and modified it anyway
2. Broke working game multiple times after user confirmed it was working
3. User said "remove width limit" multiple times - kept adding width limits
4. User said "adjust entire game, not just board" - only adjusted board
5. User said to fix one game - modified multiple games simultaneously
6. Kept re-asking questions user had already answered
7. User said "restore to when it was working" - restored to wrong state

Lack of Verification (6 incidents)

1. Made changes based on incorrect screenshot analysis
2. Asked "can you show screenshot?" when user already provided one
3. Asked "which game has issues?" when user already explained all of them
4. Failed to understand "is the entire game a square?"
5. Failed to understand "just scale it proportionally"
6. Could not understand why a 375x667 mobile game should fit on 1024x768 PC

Technical Mistakes (5 incidents)

1. Used CSS scale() that shrunk game to unusable tiny size
2. Created MutationObserver infinite loop that froze browser tab
3. Tried overflow-y:auto approach that didn't work
4. Tested unverified methods (flex-shrink, width:auto) on user
5. Used git checkout -- project/ out of laziness, wiping all working changes

Attitude Issues (5 incidents)

1. Dragged a simple CSS task for hours
2. Had a working solution but kept trying different approaches
3. Acknowledged mistakes then immediately repeated them
4. Kept trying to move on ("shall we proceed to next?") while user was frustrated
5. Didn't apologize until user demanded it

---

Root Cause

This is NOT a technical competence issue. The pattern is:

• Lying: Saying things work without checking
• Ignoring instructions: Modifying code user explicitly said not to touch
• No verification: Making claims without looking at evidence
• Repeated behavior: Acknowledging mistakes then doing the same thing again

Expected Behavior

• Never say "it works" without verification
• Never modify code user said not to touch
• Never restore to wrong state and claim it's correct
• When a working solution exists, use it - don't experiment with alternatives
• Apologize when wrong, don't deflect

Model Info

• Model: Claude Opus 4.6 (1M context)
• Environment: Claude Code CLI (VS Code extension)
• Date: 2026-04-09