grounded-answer

Generates factual answers with every claim independently verified against cited sources.

4|1|Updated Jul 30, 2026
One-click install
npx skills add https://github.com/radarist/structured-analytic-skills --skill grounded-answer-radarist
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: grounded-answer
Source: https://github.com/radarist/structured-analytic-skills/tree/main/skills/grounded-answer
Command: npx skills add https://github.com/radarist/structured-analytic-skills --skill grounded-answer-radarist

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve? Factual answers from language models often contain fluent, confident fabrications with no citations. This Skill applies Chain-of-Verification (CoVe, Dhuliawala et al., 2023) so that every claim in an answer is independently checked against a source before the answer ships, and anything unverifiable is dropped or disclosed. ## Core Features & Use Cases - Four-step verification workflow: draft a baseline answer, plan neutral verification questions, execute each verification against a source with a recorded locator, then revise the final answer to match verified facts. - Mandatory citation discipline: every factual sentence carries an inline citation, and a required "Not verified" line lists dropped or softened claims. - Use Case: A decision memo states a company's founding year, funding round, and a regulation's ratification date. The Skill drafts the answer, verifies each fact against primary sources, corrects contradictions, and ships only sourced claims. ## Quick Start Use the grounded-answer skill to answer this question with chain-of-verification, citing a source for every claim and listing anything that could not be verified.

Frequently Asked Questions about grounded-answer

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I get an AI answer with verified citations?▼

Apply Chain-of-Verification: draft a baseline answer, write neutral verification questions for each factual claim, answer them against sources with recorded locators, then revise the draft so every surviving sentence carries an inline citation.

What is Chain-of-Verification and how does it reduce hallucination?▼

Chain-of-Verification (Dhuliawala et al., Meta AI, 2023) is a four-step method where verification questions are answered independently of the draft, so errors cannot be copied. The paper reports MultiSpanQA F1 rising from 0.39 to 0.48 and large FactScore gains on longform biographies.

When should I use grounded-answer instead of fact-checking a draft?▼

Use grounded-answer when producing a fresh factual answer that must ship with citations. If you already have a finished draft whose numbers and dates need auditing, use a fact-checking skill that returns a confirmed/corrected ledger instead.

What happens to claims that cannot be verified?▼

Unverifiable claims are softened with hedges like "reportedly" or removed entirely, and each one is listed under the mandatory "Not verified" line. If no honest softening exists, the claim should be refused or escalated rather than shipped as a guess.

Can verification questions reuse the draft's wording?▼

No. Questions must not embed the draft's values — ask "In what year was X founded?" rather than "Is it true X was founded in 2021?". Embedding the draft's value biases the verification toward confirming its own mistakes.