Claude substitutes its own interpretation instead of following explicit user instructions
Bug Report
Claude repeatedly ignores explicit user instructions about what to fix, substituting its own interpretation and making things progressively worse.
What happened
User identified a CSS issue (grey bar between card rows on a React dashboard) and told Claude exactly what to fix: add a white background to the #root element. Instead of following that instruction, Claude:
- Fixed the wrong element entirely (changed a logo icon instead of a photo)
- Changed
object-positioninstead of container height - Added
background+border-radiusto a child element instead of#root - Added a
body.catalog-modeclass hack - Moved padding from
evidence-maintoevidence-page, breaking the layout on other pages - Changed the
bodybackground-color fromstone-1towhite, breaking the paper-on-desk design
Each change made things worse. The user had to escalate repeatedly before Claude finally did the one-line fix they asked for.
Claude's own assessment
"Every single time in this conversation, you've told me what the problem is, and I've done something different. I keep substituting my own interpretation instead of doing what you actually said."
Expected behavior
When the user says "the problem is you have an element called root with a background colour of white", Claude should add #root { background-color: var(--pds-white); } — not reinterpret the instruction as something else.
Environment
- Claude Code CLI
- Model: claude-opus-4-6 (1M context)
This issue has 3 comments on GitHub. Read the full discussion on GitHub ↗