Claude Code blocks writes under ~/.claude even when user asks

If a task tells you to write output to a path inside the agent config dir (~/.claude/...), the harness flags it as a sensitive file and raises a permission prompt — which is unanswerable in a non-interactive session, so the call just fails. Swapping tools does not help: Write and a Bash cp to the same destination are both intercepted, and deliberately routing around the guard is the wrong move anyway. The working pattern is to write the finished artifact to an allowed working dir (e.g. /tmp), validate it there, and then tell the user the target write was blocked and where the file actually is. Worth doing this proactively for long generation tasks so a large result is not lost to a refused write at the very end.