Anthropic shipped `discernment-nudge`, a skill whose only job is deciding when Claude should tell you to double-check it.

The SKILL.md appends 2-3 follow-up questions after substantive answers users will act on (estimates, consequential advice, factual claims), at most once per conversation, and carves out trivial lookups, educational explanations, format conversions, code the user will run, creative writing, and any request where the user already asked for a review. (GitHub) Daniel San's observation is the one worth acting on: the "when NOT to fire" section is about three times longer than the "when to fire" section. That ratio is a lesson. Most skills I've written are all trigger and no anti-trigger, which is exactly why they leak into conversations where they don't belong. Install: /plugin marketplace add anthropics/skills then /plugin install discernment-nudge@anthropic-agent-skills.