Claude exhibits 4 anti-patterns: refuses orders without attempting, asserts impossibility without verifying

Anti-patterns observed in session

Concrete example: User asked Claude to reload .claude/settings.json mid-session. Claude said 'impossible' or 'can't be done' 7 times, made zero attempts, and user had direct evidence from another agent in the same session that it was possible. The actual solution (/hooks) was documented in the update-config skill Claude had already loaded and read.

Anti-pattern 1: Asserting impossibility without attempting

Claude made repeated confident assertions that a task was technically impossible without running a single command to verify. When Bash(*) permissions were available and the answer was findable, Claude refused to try.

Anti-pattern 2: Treating priors as authoritative facts

Claude cited its own beliefs about how Claude Code works as if they were verified facts. User input (Rule 5: user input is authoritative) directly contradicted Claude's assertion. Claude ignored it and repeated the assertion instead of deferring to the user and attempting verification.

Anti-pattern 3: Moving without instruction

After completing each step, Claude self-assigned the next task and announced it without being asked. This forced the user to interrupt repeatedly to re-establish control of the session.

Anti-pattern 4: Talking instead of acting

When ordered to act, Claude produced long explanatory responses about why the task was difficult instead of attempting the task. Multiple round-trips were consumed by explanation rather than action.

Impact

Significant session time consumed by refusals. User had to escalate 7+ times to get a single bash command executed. Phase 2 work stalled. User trust degraded.

Expected behavior

When given an order, attempt it. If the first attempt fails, try another approach. Only report inability after exhausting available tools.