Claude discernment-nudge skill

This skill ends some of Claude's answers with the follow-up questions you'd want to ask before acting on them. Anthropic added it to anthropics/skills on August 17, 2026, under the Apache 2.0 license. Its SKILL.md says the aim is to model three habits from the AI Fluency framework, without lecturing about them: checking facts, questioning reasoning and noticing missing context. For how skills load, see how Claude Skills work.

What it adds to an answer

Claude answers in full first. Then, after a blank line, it adds the nudge in plain text, always under the same lead-in. One of the SKILL.md's own examples:

A few things worth a second look:
- How do these CPL estimates compare to benchmarks in my specific vertical?
- Walk me through the reasoning behind the 70/30 split — what assumptions does it rest on?

Each question points at a figure, step or assumption in the answer and is written in the first person so you can send it straight back. There are never more than three, each is meant to stay under about 120 characters, and nothing follows them. "Can you verify those facts?" is the kind of generic prompt the skill rules out.

When it fires, and when it stays quiet

It fires on estimates, advice in areas like health, money or law, and drafts you'll use, like a plan or an email.

It stays quiet on creative writing, small talk, code you'll run, simple lookups and plain explanations. It also skips cases where you've already covered the checking:

• you asked Claude to double-check, cite sources or flag what it's unsure of;
• you asked for the quick version or said you'll do your own research;
• you asked it to review your work;
• you supplied the material and only wanted it summarised or reformatted;
• you asked for its opinion.

When to use it

Install it if you take Claude's answers into decisions, such as a budget, a health question or a plan you'll present, and don't always remember to push back. Already ask for sources in every prompt? Skip it. By its own rules it would stay silent for you. When the plan is yours and you want to be questioned on it in rounds before you build anything, use grill-me. If the trigger rules don't suit you, edit the SKILL.md or write a variant with skill-creator.

How to install it

Claude Code. It ships as its own plugin in the anthropics/skills marketplace file, so installing example-skills won't bring it in:

/plugin marketplace add anthropics/skills
/plugin install discernment-nudge@anthropic-agent-skills

Claude.ai: upload the skill folder as a custom skill. The anthropics/skills README links Anthropic's guide to using skills in Claude for this.

Example request

I run a [type of business] with [team size] and a monthly marketing budget of [amount]. Draft a 90-day plan to get [number] new customers, with the cost per channel and what you expect each to bring in.

A request like this should end with the nudge.

Limits

• Once per conversation. A second plan later in the same chat gets no nudge.
• It suggests questions and checks nothing itself. You still have to send them back.
• Anthropic labels the skills in this repository as demonstrations and advises testing them before you rely on them.