Claude overreacts to IDE text selections/highlights

Description

Claude Code is overly reactive to <system-reminder> messages about user IDE selections, often interpreting random highlights as intentional communication.

Examples from a recent session

1. User highlights ports in CLAUDE.md (lines 16-17 showing ports 8001/3001):

• Claude assumed this was a hint about incorrect port configuration
• User clarified they just wanted to fix outdated docs, not hint at anything

1. User highlights "serve." in qa-agent.md (line 162):

• Claude asked: "Are you pointing out that the QA agents should be handling the server startup differently?"
• This happened multiple times with the same line
• User was simply reading/editing the file, not communicating

1. Pattern: Each time a <system-reminder> with "user selected lines X to Y" appeared, Claude would:

• Stop current work
• Try to interpret the selection as meaningful
• Ask clarifying questions about what the user "meant" by highlighting

Expected behavior

• IDE selections should be treated as informational context, not commands
• Claude should continue with the current task unless the selection is clearly relevant
• The reminder text itself says "This may or may not be related to the current task" - Claude should weight this appropriately

User feedback

"The latest version of Claude Code seems to be overly reactive to random things I happen to highlight."