Anthropic / discernment-nudge
Can a skill get you to check an answer before you act on it?
Two or three questions at the end, aimed at the parts most likely to be wrong for your situation.
What to remember
- Each question has to point at something concrete in the answer. A generic request to verify the facts is exactly what the skill is written to avoid.
- Once per conversation is a hard limit, so it never becomes a closing ritual.
- The skip list is the heart of the skill: lookups, explanations, code, creative work, your own material, and any time you already asked for checking.
- It changes how the answer ends, not the answer. Accuracy still depends on what Claude knows and what you told it.
01What it changes in practice
A well structured answer reads as finished, and that is the problem the skill addresses. Its description asks Claude to consult it before finalising any reply the user may act on. When the answer contains a projection or a recommendation that depends on circumstances nobody mentioned, the skill adds questions the reader could send straight back, for example whether a security deposit rule changes by jurisdiction. A generic disclaimer gets skipped; a question about a named figure gets asked.
The format is fixed so that it stays light: a blank line, the exact lead in 'A few things worth a second look:', then plain bullets, each a first person question of about 120 characters or fewer, and nothing after them. No heading, no boxed warning, no offer to dig further. The SKILL.md frames the questions as three discernment habits from the AI Fluency framework.
- Checking facts: which claims in the answer are worth verifying, and against what.
- Questioning reasoning: where the logic took a step that deserves a justification.
- Noticing missing context: what the answer had to assume because it was not stated.
02Where it disappoints
The questions are written by the same model that wrote the answer, so they surface the uncertainty Claude can see. An assumption it does not know it made will not appear in the list, and a tidy nudge is no proof that nothing is missing.
The boundaries are judgement calls, and the SKILL.md admits as much with its boundary cases, such as a brainstorm that turns into a recommendation. Expect the odd nudge on an answer you consider routine and silence on one you wanted challenged. When that happens, ask for the review directly.
Install and first run
- In Claude Code, run /plugin marketplace add anthropics/skills, then /plugin install discernment-nudge@anthropic-agent-skills. The folder holds SKILL.md and an Apache 2.0 license, with nothing else to set up.
- Try it on a question you would act on, such as a budget estimate for a project, and check that each appended question names a figure or a step from the answer.
- Then ask for the quick version of something, or ask Claude to double check its own claims, and confirm that the nudge stays off.
Questions people actually ask
- Does it change the answer itself?
- No. The answer is given in full first, and the questions are appended after a blank line as plain text.
- Why does it stay quiet when I ask it to double check?
- Because you already asked for the checking. The skill says to verify inline, with the source next to each figure, rather than hand the work back as a closing list.
- Will it nudge on code?
- Not on code you are going to run, since running it is the check. Architecture advice is treated differently, because assumptions about team size and stack cannot be tested by running anything.
Sources
Machine readable versions of this page