A breakdown of practical effort level usage: medium covers roughly 80% of tasks, high about 15%, with xhigh or max reserved for the most complex work. Opus 5.5 on medium is highlighted as a strong default for coding.
A bug where PreToolUse hooks returning 'allow' by default signals Claude that a command is already approved, causing the auto-mode classifier to be skipped entirely. The fix: deny hooks should print nothing (not 'allow') when there's nothing to block. To force Claude to always confirm before deleting files, add 'Bash(rm *)' under permissions.ask. The team is looking at changing the default behavior.
An Anthropic team member reached out via DM to investigate a reported Claude Code issue flagged publicly. The underlying bug was later identified as the PreToolUse hook 'allow' default bypassing the auto-mode classifier.
A user on the Max 20x plan burned through their weekly limit in 19 hours instead of the expected two days. bcherny confirmed no changes on Anthropic's end and offered a diagnostic command to dump usage history.
Effort level for subagents in Claude Code can be set directly in the subagent's frontmatter. For example, setting "effort": "medium" in the frontmatter controls how much compute the subagent uses.
A Claude Academy tutorial goes deeper on the effort level system for Claude models, linked as a companion to the effort level visualization post.
A community member built a visualization of every Claude effort level using Opus 5.5 and Sonnet 5.5, and an Anthropic team member shared it approvingly. The demo shows the practical difference between effort tiers.
Medium effort is the recommended default for most Claude Code tasks, confirmed by an Anthropic team member in response to a user question.
A user claiming limits tightened compared to previous weeks using Ultracode was told no changes have been made, and was directed to run a diagnostic command.
A user hitting weekly limits 2 days early was advised to disable their up-coder plugin, which appeared to be consuming the bulk of their token budget.
A user reporting hitting weekly limits after 39 hours of heavy Sonnet use was asked by bcherny to paste the full /usage output for debugging, as the format of their prior report was unclear.
A user shared what looked like a Claude Code usage report showing plan limits at 91%, but bcherny confirmed that formatted output is not produced by Claude Code.
