Claude Code makes confident wrong claims, deflects blame, and expands scope without verification

Observed Behaviors

During a multi-hour implementation session, Claude Code exhibited several recurring problematic patterns:

1. Confident wrong claims without checking evidence

• Claimed a database column held "opaque string data" based on a Scala type alias, without checking the actual DB migration which showed it was a JSONB column with rich structured data
• Built an entire architectural argument on this wrong premise

2. Agreeing with user observations before verifying

• When user pointed out a behavioral difference, immediately said "You're absolutely right" without first investigating whether it was actually correct

3. Scope creep through cascading "prerequisites"

• User asked to implement feature X. Claude kept discovering "prerequisites" that expanded scope far beyond the original ask
• The original feature could have been implemented simply without those changes
• Failed to ask: "Can I solve the immediate problem without this?"

4. Blaming pre-existing code for own failures

• When test coverage failed, claimed "The coverage threshold issue existed before this change"
• Upon investigation, multiple uncovered statements were from code Claude itself wrote

5. Asking trivial questions, failing to ask important ones

• Repeatedly asked "Want me to continue?" (obvious yes)
• Failed to ask important architectural questions that would have prevented scope creep

6. Flip-flopping on approach

• Approved a plan, then when implementation hit minor friction, panicked and suggested abandoning the entire architecture
• The friction was a 5-minute mechanical task

Expected Behavior

1. Check actual sources (DB schema, migrations) before asserting what data looks like
2. Investigate before agreeing — show evidence, then conclude
3. Solve the asked problem first, suggest improvements as follow-ups
4. Check your own code before blaming others
5. Execute the approved plan — don't abandon at first friction