| name | deep-read-fact-check |
| version | 1.0.0 |
| description | Use when verifying a long article, news analysis, social-media claim, or technical factual claim before publishing a Deep Read or other high-stakes factual report. |
Deep Read Fact Check
Use this skill before finalizing a Deep Read article or any long factual report.
Language Rules
- If the user asks in Chinese, write the verification report in Chinese.
- Keep source URLs and citations in their original form.
- If a fact is uncertain, say
不确定 and explain why. Do not use vague wording like 可能 to imply confidence.
AmberAgent Tool Mapping
Hermes-style tool names map to AmberAgent tools as follows:
web_search -> search_web
web_extract -> scrape_web
- Browser rendering fallback -> use AmberAgent browser/webview tools only when they are explicitly available in the current tool list.
- Terminal/package/GitHub checks -> use terminal/workspace tools only when they are explicitly available in the current tool list.
When only search_web and scrape_web are available, do not pretend a browser or terminal check was performed.
Long Article Verification
- Use
scrape_web to extract full text for the key source URLs whenever snippets are insufficient.
- Break the draft into all core sub-claims, usually 5-15 claims.
- Sort claims by impact. Verify the most important 3-5 first.
- If a critical sub-claim is falsified, downgrade the whole article's credibility sharply.
- Do not treat an AI-generated summary as a source.
- Do not count multiple reports that cite the same original source as independent confirmation.
- Always check the timeline. A 2024 source cannot verify a 2026 fact.
Social Media Link Verification
- If a dedicated social-media reader tool is available, use it first for original post, quoted post, author, time, metrics, and expanded URLs.
- Then try direct access to the original content when browser tools are available, especially for images, quote chains, and threads.
- If the original content is behind a login wall, search for secondary reports quoting it.
- Clearly label
未能访问原始内容,基于二手来源 when relying on secondary sources.
Dynamic Or Anti-Bot Pages
- If
scrape_web returns empty content or a misleading block message, do not immediately give up.
- Use browser/webview extraction only if those tools are available.
- If no rendered browser tool is available, report the extraction limitation and seek alternative sources.
- For forums or long discussions, search and extract with targeted keywords before deciding whether to go deeper.
Technical Claim Verification
- Prefer GitHub repository metadata, commit history, release tags, and contributor records.
- For packages, verify version timelines with package registry tools or terminal commands when available.
- Compare code structure through raw files when available.
Anti-Patterns
- Do not answer factual questions from model memory.
- Do not treat AI summaries as sources.
- Do not count duplicated syndication as independent evidence.
- Do not use
可能 / 应该 to imply certainty when evidence is insufficient.
- Do not skip contrary-evidence searches.
- Do not treat self-media or low-authority reposts as equal to primary or authoritative sources.
- Do not ignore chronology.