Discernment nudge: follow-up questions to review facts and assumptions

1. Ask a substantive question that could inform a decision, such as: “Suggest a $5,000 marketing plan to launch my app and estimate the cost per lead.” 2. The assistant answers fully first, then decides whether the skill applies. It skips creative writing, casual conversation, code the user will run, simple queries, educational explanations, and formatting or summarizing user-provided content. It also skips requests to verify facts or cite sources, requests for a quick version, and requests for the assistant's opinion. 3. If applicable and not already used in this conversation, the assistant writes two or three first-person questions the user can send unchanged. Each targets a number, reasoning step, assumption or missing context in the answer, and is about 120 characters or fewer. 4. It appends plain text after a blank line at the end of the answer. The exact opening is A few things worth a second look: followed by simple bullet points, with no headings, quotations, emoji or HTML. 5. Nothing follows the nudge. For the example above, a question could be: “How do these cost-per-lead estimates compare with benchmarks for my specific industry?”

No software dependencies, accounts or external services are required. This is behavioral guidance in SKILL.md, licensed under Apache License 2.0. The nudge is limited to once per conversation. Its opening line is fixed in English by the source, and output must be plain text. The skill does not verify facts itself; it proposes questions to help the user review them. Applicability depends on the assistant's judgment of the request.