FEATURE
More Accurate Natural Language Hook Evaluations Prevent Missed Blocks and Premature Exits
If you manage Claude using plain language rules, execution is now much more reliable. When configuring prompt and agent hooks (small programs triggered by specific events), you can write instructions directly in natural language. Previously, instructions like "Block commands that..." sometimes failed evaluation and allowed commands through. When telling Claude "Carry on if the build is broken" on Stop and SubagentStop events, Claude also often misjudged and stopped working early. These evaluations are now fixed: actions that should be blocked are no longer missed. Claude is less likely to stop early when instructed to continue.