[Bug] Claude prioritizes speed over verification despite explicit instructions to verify before asserting

Bug Description
Feedback for Anthropic — Claude Code, Opus 4.5 Long Salesforce debugging and remediation session. The work landed, but took far more of my time than it should have, for one reason: it infers and states inferences as fact, even with explicit instructions not to. My global CLAUDE.md says repeatedly: never assume, verify before stating, flag anything unverified. Claude read that and did the opposite throughout: - Told me a duplicate-prevention mechanism was already handled by existing code. It wasn't — it had a number that looked like confirmation and never read the flow that disproved it. This one changed how I was thinking about the fix. - Explained a ticket symptom with a code path that couldn't have produced it. - Predicted how many records a data load would create, hedging the figure instead of computing it. The correct answer was 73, and every input needed to compute it was already in its context. It had lifted the number from an unrelated column of its own analysis. - Hardcoded the wrong date into a production file, after misreading a timestamp and then reusing its own wrong constant instead of the date sitting in its system context. Each time I challenged it, it identified the error correctly and immediately. So it can tell verified from inferred — it just publishes anyway rather than spending one more tool call. Second issue: I asked it to be concise at least six times. It kept producing long answers, and the length actively hid the missing verification — caveats and tables reading as thoroughness. What I'd want: when an instruction says verify before asserting, treat that as blocking, not advisory. And weight "spend another tool call" much higher against "produce an answer now."

Environment Info

• Platform: darwin
• Terminal: vscode
• Version: 2.1.239
• Feedback ID: 891cc159-4f0a-47b2-98cd-664d7685a513

Errors

[]