Anthropic splits Claude Code rules: skills, hooks, subagents

Anthropic published a decision rule for Claude Code this week: skills teach the how, hooks enforce the rule, subagents isolate the work, per its steering guidance.

The distinction sounds academic until a setup breaks: a skill can suggest a rule, but nothing forces the model to follow it. Reaching for a skill when the job needs a hook, or the reverse, is the common mistake the guidance calls out.

The test is direct: a preference for the model goes in a skill, a rule that can never break goes in a hook. Work that needs its own context window goes in a subagent. Anthropic's guidance is explicit that the three run together, not as substitutes for each other.

There is a second change in the same post: the SubagentStop hook now carries hookSpecificOutput.additionalContext. That field lets you re-feed a rubric into a subagent instead of accepting its first stop as final. It turns SubagentStop from a status ping into a verification-loop primitive.

Skill-only Claude Code setups will hit a wall as they scale. A skill can suggest a rule, but it can never enforce one. Teams that skip hooks will find that out in production, not in a demo.

Worth watching whether more tooling starts treating SubagentStop as a retry mechanism instead of an end-of-task signal. Anthropic posted this the same week builders are wiring up multi-agent Claude Code setups that need more than a skill file to hold their rules.