Skill detail

fact-check

Improve factual reliability while creating or reviewing a course or supplied content. Use when the user asks to fact-check, verify accuracy, reduce hallucinations, make a reliable course, or mentions 事实性错误、知识性错误、专业知识准确性、可靠性. During creation, checks the completed pages before delivery; on existing content, returns a short evidence-backed report and lets the user choose what to fix. Not for grammar, style, or layout. Combine with deep-research when current evidence is the course's main subject.

Why we recommend it: Verifies shortlisted high-signal claims in a course or supplied content against materials and web sources, returning an evidence-backed report of corrections.

MatchDirectRecommended for Deep Research
SourceTHU-MAIC/​OpenMAICExternal source
Reported installsNot reportedPopularity signal only

Inspect before use

Read the source instructions and check the requirements before you install this skill.

Saved source preview

SKILL.md

The saved excerpt is a snapshot from when this skill was listed. The external source remains the complete and most current version.

---
name: fact-check
title: "事实核查"
description: "Improve factual reliability while creating or reviewing a course or supplied content. Use when the user asks to fact-check, verify accuracy, reduce hallucinations, make a reliable course, or mentions 事实性错误、知识性错误、专业知识准确性、可靠性. During creation, checks the completed pages before delivery; on existing content, returns a short evidence-backed report and lets the user choose what to fix. Not for grammar, style, or layout. Combine with deep-research when current evidence is the course's main subject."
---

# Fact check

Keep serious factual mistakes and AI hallucinations out of the course without
turning course-making into an exhaustive audit. Focus on the few claims that
materially affect trust.

Choose the mode from the request and current course state; do not ask the user
to choose a mode:

- **Creating:** when there is no course yet or the user asks to build/rebuild
  one, load `stage-design` and run the check only after all pages exist.
- **Reviewing:** when content already exists and the user asks to inspect it,
  report findings first. Do not edit unless fixes were already requested or the
  user approves findings after the report.

## While creating a course

Use the normal `stage-design` workflow; this skill changes factual handling,
not the teaching method, page style, or build sequence.

After all pages exist, call `list_scenes`, then read every completed page with
`read_stage` using `detail:"text"`; follow `nextOffset` until all visible text
and narration have been read. Run a quick final sanity check of exact facts and
cross-page contradictions. Correct obvious errors before delivery because
creating the course already authorizes making its content accurate, subject to
the source-of-truth boundary below. Do not interrupt creation with a separate
audit report or approval gate unless that boundary requires a user decision;
briefly mention only material corrections or remaining uncertainty when
handing off the finished course.

## When reviewing existing content

For a course, call `list_scenes`, then read all visible text and narration with
`read_stage` using `detail:"text"`; follow `nextOffset` until complete. Respect a
narrower scope if the user gave one.

Read once for context and silently shortlist high-signal risks:

- exact numbers, dates, counts, names, and attributed quotations;
- laws, standards, formulas, technical definitions, and classifications;
- “first”, “only”, “always”, “must”, and similar absolute claims;
- causal or professional conclusions stated as settled fact;
- contradictions between pages;
- suspiciously specific claims with no visible support.

Do not verify every claim. Skip correct material, wording preferences, harmless
simplifications, and low-value trivia.

## Verify only the shortlist

Check `list_materials` and relevant `read_material` content first. User sources
may support a claim, but the course being checked cannot prove itself.

Use `web_search` for shortlisted claims that depend on current, exact, disputed,
or specialist knowledge. A normal first pass should need no more than about
6–8 searches. Use `fetch_url` to read the source: a result snippet or the mere
existence of a related source is not evidence.

Do not pause the run to ask the user for sources, permission to use general
knowledge, or permission to continue. Use the tools and materials that are
available. If web search is unavailable, continue with stable knowledge, make
fewer factual commitments, and mark genuinely uncertain claims. Never guess a
URL for `fetch_url`; fetch only a user-provided URL or one returned by
`web_search`.

For a compound statement, isolate the questionable part and verify that exact
part. Prefer primary or official sources. One authoritative source is enough
for an obvious error; add corroboration only for disputed or high-impact
claims. For versioned knowledge such as law, policy, standards, or medicine,
check the relevant date, version, and jurisdict
Read the full source on GitHub (opens external page)
Context

Related work