The Discernment Nudge skill adds a small critical-review step after a substantive answer or draft. When its conditions apply, it asks two or three questions tied to specific facts, assumptions or missing context in the output.
Key takeaways
- It runs at most once per conversation according to its description.
- It skips trivial tasks, pure formatting, casual chat and several other defined cases.
- It is instruction-only and does not inspect external systems by itself.
What does the Discernment Nudge skill do?
The skill is designed for advice, plans, proposals, estimates, analysis and factual claims that a user may act on. Instead of adding generic warnings, it asks concrete follow-up questions connected to the answer just produced. Its boundary rules are equally important: too many questions would interrupt straightforward work and reduce usefulness.
How does the workflow work?
- Complete the substantive answer or draft first.
- Check whether the output contains decisions, assumptions, projections or claims the user may rely on.
- If applicable, identify the two or three most material points worth checking.
- Append short questions that help the user validate those points without reopening the entire task.
The exact result still depends on the model, the files available in the working environment, and the permissions granted to that environment. This page documents the repository instructions; Anavem did not run an end-to-end benchmark of the skill.
What does a useful first task look like?
Use a draft rollout plan containing two explicit assumptions and one external dependency. After the plan is written, ask the skill to append no more than three questions tied to those exact uncertainties. The questions should identify what evidence would change the decision instead of repeating generic caution.
How can you verify the result?
- Limit the addition to two or three questions.
- Tie every question to a specific claim, assumption or dependency in the answer.
- Confirm the skill did not interrupt a task covered by its skip conditions.
A successful run should produce an output you can inspect independently. If the result cannot be checked against a file, rendered artifact, source reference or explicit acceptance criterion, narrow the task before relying on it.
When is it a good fit?
- Plans and recommendations with important assumptions.
- Drafts that depend on facts supplied by the user.
- Analyses where missing context could change the decision.
Choose it when those tasks match the documented scope. A popular repository is not evidence that a skill is appropriate for confidential or production data.
What should you review before installing it?
- Whether the user already requested a fact-check or review.
- Whether follow-up questions add value or merely delay a simple task.
- Whether the questions identify specific uncertainty instead of giving vague disclaimers.
Read the current SKILL.md and every bundled script before installation. Pin a reviewed commit when repeatability matters, then test with non-sensitive data and the narrowest available permissions.
When is it not the right choice?
- Simple lookups, formatting or casual conversation.
- A substitute for actual verification against primary sources.
- Repeated use that burdens every turn with boilerplate questions.
Why is it included in this research set?
This skill is published in Anthropic’s official Agent Skills repository. GitHub’s repository API reported 179,524 stars and 21,221 forks on 2026-10-03. Those figures apply to the repository as a whole, not to this individual folder.
GitHub stars and short-term growth are discovery signals, not quality scores. They do not prove security, correctness, maintained compatibility, or individual-skill usage.