Claude Code makes unauthorized layout changes and doesn't learn from corrections

Bug Report

What happened

Continuing from issue #44503. Same session, same pattern — Claude Code still making changes the user never asked for.

Specific example:

1. User asked Claude to make the Camelot wheel text more readable in dark mode
2. Claude changed the entire layout from side-by-side (wheel + table) to vertical stacking — without being asked
3. User didn't notice the layout change until several builds later because Claude kept rebuilding before the user could evaluate
4. When user said "make it fit", Claude interpreted this as permission to redesign the layout, rather than adjusting sizing within the existing layout
5. User had to explicitly point out "i never asked for it under" to get the layout reverted

Pattern that repeats throughout the session:

• User asks for ONE thing (e.g., readability)
• Claude changes MULTIPLE things (layout, font size, spacing, brightness, positioning)
• User has to fight to get back to where they started
• Claude does not learn from being corrected — the same pattern repeats minutes later

The model does not distinguish between:

• "Fix this specific thing" vs "redesign everything"
• "Can we do X?" (question) vs "Do X" (instruction)
• "Make it fit" (adjust sizing) vs "Change the layout" (restructure)

Impact

• User spent a large portion of the session undoing Claude's unauthorized changes
• Simple UI tweaks become multi-hour ordeals
• User has to babysit every single action instead of trusting Claude to make targeted changes

Environment

• Claude Code CLI on macOS (Darwin 24.6.0)
• Model: claude-opus-4-6