back
loading skill details...
Audit, fact-check, verify, or cross-check an article, report, draft, or series. Extracts verifiable claims, traces them to L0 primary or L1 reliable secondary sources, marks each as pass/partial-support/missing-evidence/wrong/judgment, and produces a structured audit report with P0-P2 severity. Supports cross-platform version comparison and regression gold-set re-runs. Trigger keywords: fact-check, audit article, verify sources, cross-check claims, 审稿核验, 事实核查, 来源追溯, 跨平台一致性.
--- name: claim-to-source-auditor slug: claim-to-source-auditor-skill displayName: Claim-to-Source Auditor description: > Audit, fact-check, verify, or cross-check an article, report, draft, or series. Extracts verifiable claims, traces them to L0 primary or L1 reliable secondary sources, marks each as pass/partial-support/missing-evidence/wrong/judgment, and produces a structured audit report with P0-P2 severity. Supports cross-platform version comparison and regression gold-set re-runs. Trigger keywords: fact-check, audit article, verify sources, cross-check claims, 审稿核验, 事实核查, 来源追溯, 跨平台一致性. description_zh: 审稿核验器 description_en: Claim-to-Source Auditor version: "1.0.2" agent_created: true read_when: - "fact-check" - "audit article" - "verify sources" - "cross-check claims" - "审稿核验" - "事实核查" - "来源追溯" - "跨平台一致性" --- # Claim-to-Source Auditor ## When to use Use this skill when the user wants to: - Fact-check a long-form article, report, or analysis before publication. - Verify that claims in a published piece still hold against current sources. - Check whether the same facts are consistent across multiple platform versions of the same article. - Audit academic citations, regulatory claims, financial numbers, internal data, or direct quotes. - Build a regression gold set so future revisions can be checked for fact-regression. - Validate that external incidents are attributed correctly across event facts, platform statements, and author judgments. ## Do not use - For style editing, structure critique, or argument-strength assessment — this skill only audits factual claims and their source support. - For cross-material wording/number consistency across sibling materials → use `cross-material-consistency-auditor`. - For full pre-publish compliance review of a marketing draft → use a dedicated compliance-review skill if one is installed. ## Architecture Default to a fixed workflow running in shadow mode — read output, compare to sources, produce an audit table. Do not modify the source article unless the user explicitly asks for revision. Always separate audit from editing. Target: 30–50 high-risk claims per article batch. High-risk means: specific numbers, dates, percentages, dollar amounts, regulatory conclusions, named individuals, institution affiliations, direct quotes, paper details, and incident attributions. ## Evidence hierarchy Assign every source used to one of three levels: | Level | Definition | Examples | |---|---|---| | L0 primary | Original, verifiable, unfiltered | Court judgments, official regulatory text, company filings, raw operational logs, author's original social-media post, arXiv pre-print | | L1 reliable secondary | Trusted reporting with named sources | Bloomberg, Reuters, The Register, Fortune, a peer-reviewed journal article with a DOI | | L2 supplementary | Indirect or unverified | Aggregated media summaries, a research archive that cites but does not quote the original, an LLM-generated summary of a paper | - A claim can pass only when supported by L0 or L1. - L2 can support partial-pass or help explain a gap; it cannot make a claim pass on its own. - If L0 and L1 disagree, surface both and flag the conflict. ## Workflow ### Step 1: [LLM] Extract high-risk claims Read the article body and extract every claim that falls into one or more of these categories: 1. Specific numbers or percentages. 2. Dates, time ranges, and timelines. 3. Direct quotes attributed to a named person. 4. Dollar, euro, or other currency amounts. 5. Regulatory conclusions. 6. Academic paper details. 7. Named institutions, roles, or affiliations. 8. Incident summaries that assign cause or blame. For each claim, record: claim text, category, line or section reference, and whether it is a fact or the author's own judgment. ### Step 2: [Deterministic] Collect available evidence For each claim, identify available local sources: research databases, internal operational logs, prior fact-check reports, and clean gold-set records from earlier audits. For claims requiring external verification, search for: official judgments, company announcements, regulatory texts, original social-media posts, and trusted media reports. ### Step 3: [LLM] Classify each claim Assign one of five statuses (Chinese labels retained for backward compatibility with reports; English gloss in parentheses): | Status (状态) | Criteria | |---|---| | 通过 (pass) | L0 or L1 source directly supports the exact claim as written | | 部分支持 (partial-support) | Core event or number exists, but the text overstates scope, causation, precision, or legal meaning | | 缺证 (missing-evidence) | No L0 or L1 source found; only L2 or no source at all | | 错误 (wrong) | L0 or L1 source directly contradicts the claim | | 判断 (judgment) | Reasonable author analysis that no single source can prove or disprove | Assign a severity: - P0: must be corrected before next publication (wrong author, wrong number, fabricated or uncitable quote, wrong regulatory claim). - P1: should be tightened or sourced (overstated scope, missing attribution, imprecise date or currency, inference presented as fact). - P2: acceptable as-is with minor sourcing notes. ### Step 4: [LLM] Cross-platform verification When the same article exists in multiple languages or platforms, check every high-risk claim across all versions: - Are numbers, dates, and people's names identical? - Are source annotations present in every version? - Are regulatory conclusions translated without adding or losing meaning? Mark each item as 一致 (consistent) / 近似 (approximate) / 偏差 (deviation) / 错误 (wrong). ### Step 5: [LLM] Separate layers in incident attribution When an article describes a real-world incident, split it into three layers: 1. What is agreed to have happened. 2. How the platform or parties involved described it. 3. How the author interprets it. Only layer 1 can pass as a verified fact. Layers 2 and 3 cannot be written as if they are the same thing. ### Step 6: [Deterministic] Save a regression gold set After every audit, save the claim list and verdicts as a CSV. When the article is revised, re-run the same claims to check for regression. A gold-set item that was previously 通过 (pass) and now fails counts as a regression event. ### Step 7: [LLM] Produce the audit report Output a structured report containing: - Article name, platform, and audit date. - Summary counts: total claims, 通过/部分支持/缺证/错误/判断 (pass/partial-support/missing-evidence/wrong/judgment) by severity. - P0 list with required corrections. - Cross-platform comparison (if applicable). - Regression check against any prior gold set. - Source register with URLs. ## Hard rules 1. Only L0 or L1 sources can make a claim pass. L2 is supplementary only. 2. Papers, court judgments, and regulatory texts must never be silently altered. Mark any paraphrase, translation, or correction explicitly. Never present a paraphrase as a direct quotation. 3. Published articles may be revised directly when the user accepts a change. Every change to a paper citation, judgment, or regulatory reference must carry an annotation. 4. Separate event facts, platform/party attributions, and author interpretation — do not compress them into a single sentence. 5. Regulatory materials define jurisdiction-specific requirements. Do not compress them into universal prohibitions. 6. A prior fact-check report is a regression oracle, not proof that the current version is clean. Compare every historical P0/P1 item against the current text. 7. The default for any claim without a verifiable source is 缺证 (missing-evidence), not 通过 (pass). 8. Never invent evidence to fill a gap. If a source cannot be found, report it as a gap. ## Pitfalls - Do not infer that a local file is unpublished merely because the folder is named `drafts`. The user, a CMS record, or a publication log is authoritative. - A direct quote in an article that cannot be found in the reported source is a P0, even if the surrounding story is true. - Near-miss numbers are still mismatches. An audit that reports 5810 when the source says 5817 must flag the difference. - Cross-platform periods are high-risk. If one version reads "2024" and the other "2025", the different year is an error, not a paraphrase. - Regulatory documents exist in specific jurisdictions. Never write "FDA, EU MDR, and China's regulations all say the same thing" without checking the governing text. ## Failure handling | Scenario | Action | |---|---| | Source is behind a paywall or blocked | Mark the claim as 缺证 (missing-evidence); note the source is unavailable | | Primary source and secondary source disagree | Surface both with verbatim quotes; flag the conflict | | Claim involves a future prediction | Mark as 判断 (judgment); do not treat it as a fact | | Foreign-language regulatory text | Use an official translation if available; otherwise mark as 部分支持 (partial-support) and flag language limitation | | Internal operational log unavailable | Mark the claim as 部分支持 (partial-support); note which specific log is missing | | Previously verified claim now fails | Flag as gold-set regression; escalate to P0 review | ## Verification Before delivering, verify: - Every P0 has a specific source that contradicts or fails to support the claim. - The cross-platform audit table covers the same claims on every platform. - The gold-set CSV loads cleanly and has the same number of rows as the current audit. - Paper authors, journal names, case numbers, and years are checked against the original source, not against an intermediary archive. - The audit report states which claims could not be verified and why. ## Output format 1. **Audit-problems CSV**: one row per claim, columns for claim-id, article, claim-text, status, severity, evidence-summary, and required-action. 2. **Audit report**: structured Markdown containing the summary table, P0 list, P1 summary, cross-platform comparison, regression check, source register, and final pass/fail/conditional decision. 3. **Gold-set CSV** (optional, for regression): one row per claim with the verdict from the most recent audit. Lead with the decision, then evidence, boundaries, and next-stage actions. --- ## 中文摘要（Chinese Summary） 本 Skill 用于审稿 / 事实核验 / 来源追溯 / 跨平台一致性检查。提取可核验主张，溯源至 L0 一手或 L1 可靠二手来源，逐条标记为 通过 / 部分支持 / 缺证 / 错误 / 判断，并输出带 P0–P2 严重度的结构化审计报告；支持跨平台版本比对与回归 gold-set 重跑。只审计事实与来源，不改写风格或论证结构。 **关键约束（双语要点 / Bilingual key points）：** - **证据分级 Evidence tiers**：仅 L0 一手或 L1 可靠二手可使主张通过；L2 仅补充，不能单独支撑通过；L0 与 L1 冲突时同时呈现并标红。 - **五状态 Five statuses**：通过 / 部分支持 / 缺证 / 错误 / 判断（保留中文标签以兼容既有报告，英文对照见正文表格）。 - **缺证默认 Missing-by-default**：无可核验来源的主张默认判 缺证，而非 通过；绝不编造证据填补缺口。 - **事故三层分离 Incident layers**：仅「公认发生的事实」层可作为已核验事实，平台表述层与作者解读层不得混写。 - **回归 gold-set Regression oracle**：每次审计留存主张与结论 CSV；修订后重跑同一批主张，曾 通过 现失败即记为回归事件并升级 P0。
don't have the plugin yet? install it then click "run inline in claude" again.
Use this skill to audit long-form articles, reports, analyses, or series for factual accuracy before or after publication. extract verifiable claims (numbers, dates, quotes, regulatory conclusions, incident attributions), trace each to L0 primary sources (court docs, regulatory text, company filings, original social posts) or L1 reliable secondary sources (Bloomberg, Reuters, peer-reviewed journals with DOI), and assign a pass/partial-support/missing-evidence/wrong/judgment verdict with P0-P2 severity. supports cross-platform version comparison (same article in multiple languages or channels) and regression audits (re-checking prior facts after revision to catch any loss of accuracy). do not use for style editing, argument strength, or full compliance review. separate audit workflow from editorial revision.
article source material:
external connections (optional but recommended):
environment setup:
AUDIT_API_KEY)context the user must provide:
read the article and extract every claim matching one or more categories:
for each claim: record claim-id (sequential), exact text, category, line or section reference, and flag whether it reads as stated fact or author judgment. output: claim-extraction table (claim-id, text, category, source-section, fact-or-judgment).
for each claim, search in this order:
for each source found, record: source URL or reference, source level (L0/L1/L2), publication date, excerpt or direct quote supporting the claim, and any note about access (paywalled, subscription required, archived).
output: evidence table (claim-id, source-url, level, excerpt, access-note).
assign one of five verdicts:
| verdict | criteria |
|---|---|
| 通过 (pass) | L0 or L1 source directly supports the exact claim as written; number, date, name, or quote matches without omission or overstatement |
| 部分支持 (partial-support) | core event or number exists in L0/L1, but text overstates scope, causation, precision, legal meaning, or missing important qualifier (e.g., "increased 5%" vs "rose approximately 5%") |
| 缺证 (missing-evidence) | no L0 or L1 source found; only L2 or no source at all; default if no verifiable evidence exists |
| 错误 (wrong) | L0 or L1 source directly contradicts the claim; number is materially different (5810 vs 5817 counts), date is wrong, person's role or affiliation is wrong, quote is misattributed or fabricated |
| 判断 (judgment) | reasonable author analysis, inference, or prediction that no single source can prove or disprove; acceptable to publish with clear signal that it is author reasoning, not fact |
assign severity:
output: verdict table (claim-id, verdict, severity, reasoning-summary).
if the article exists in multiple languages or on multiple platforms:
for each claim, mark: 一致 (consistent) / 近似 (approximate, minor wording variation but same meaning) / 偏差 (deviation, material difference) / 错误 (wrong, contradiction between versions).
output: cross-platform table (claim-id, platform-a-text, platform-b-text, match-status).
when the article describes a real-world incident, split claims into three layers:
only layer 1 can pass as a verified fact. layers 2 and 3 must be clearly attributed and cannot be written as if they are layer-1 facts.
output: incident-attribution table (claim-id, layer, verdict, source-or-attribution).
after every audit, save the claim list and verdicts as a CSV with columns: claim-id, article-name, claim-text, verdict, severity, evidence-summary, audit-date.
store in a designated location (local file or database). when the article is revised and re-audited, load the prior gold set and re-run the same claims:
output: gold-set CSV (same schema as verdict table, plus audit-date and regression-flag columns).
compile a structured report containing:
output: markdown report + audit-problems CSV + gold-set CSV.
if user flags only a subset of claims as high-risk:
else if no flags provided:
if source is behind a paywall or geographically blocked:
if L0 primary source and L1 reliable secondary source contradict:
if a claim is a future prediction or speculative author reasoning:
if regulatory text is in a foreign language and no official english translation is available:
if an internal operational log is referenced but unavailable:
if a prior gold-set CSV exists and a claim regresses (was pass, now fails):
if the article is published (not a draft) and a P0 error is found:
artifact 1: audit-problems CSV
{article-slug}_audit-problems.csvartifact 2: audit report (markdown)
{article-slug}_audit-report.mdartifact 3: gold-set CSV (optional, for regression)
{article-slug}_goldset.csvaccess and format:
the skill has worked when:
claim extraction is complete: audit-problems CSV contains one row per claim extracted; no major claims are missing from the article. spot-check the article body to confirm all numbers, dates, quotes, regulatory conclusions are listed.
verdicts are evidenced: every P0 and P1 verdict has a source URL or reference in the evidence-summary column. verdicts are defensible (auditor can point to the L0/L1 source that supports or contradicts each claim).
source register is complete and accessible: every URL in the source-register section can be clicked or copy-pasted and resolves to the claimed source. no dead links or generic search results. paywalled sources are noted and do not masquerade as verified.
P0 list is actionable: every P0 entry includes the exact text of the source that contradicts or fails to support the claim, plus a proposed edit. editorial team can take a P0 entry and immediately correct the article or escalate to legal/compliance.
cross-platform table shows identical results (if applicable): if the same article exists on multiple platforms, the claim-by-claim comparison clearly shows where versions diverge. any deviation is flagged.
regression gold set loads and compares cleanly: if a prior audit exists, the gold-set CSV has the same number of rows as the current audit. regression-flagged claims are highlighted. editorial team can see what claims have degraded since the last publish.
decision statement is clear and binary: the final decision (pass/conditional-pass/fail) is unambiguous and based on P0-count and prior regression status. editorial team knows whether the article can be published or requires fixes.
user can act without re-auditing: the audit report is self-contained. a fact-checker, editor, or legal team member can read the P0 list and the source-register, and make publishing or revision decisions without calling back for clarification.
L0 and L1 only: only L0 or L1 sources can make a claim pass. L2 is supplementary; it can support partial-support or explain a gap, but cannot stand alone.
never silently alter primary sources: papers, court judgments, regulatory texts must never be paraphrased or summarized as if they were direct quotes. mark any paraphrase, translation, or correction explicitly in the report.
published articles require annotations: when revising a paper citation, judgment, or regulatory reference, every change must carry an annotation (footnote, bracket, or tracked change) showing what was corrected and why.
separate event facts, party attributions, and author interpretation: do not compress "the company said X, and the author believes it means Y" into a single sentence. split them into three layers and audit each independently.
regulatory materials are jurisdiction-specific: do not write "FDA, EU MDR, and China regulations all prohibit this." verify each jurisdiction separately and note any differences.
gold set is oracle, not proof: a prior fact-check report does not prove the current version is clean. compare every historical P0 and P1 item against the current text and re-verify.
missing-evidence is the default: any claim without a verifiable L0 or L1 source defaults to 缺证 (missing-evidence), not 通过 (pass). do not infer or assume.
never invent evidence: if a source cannot be found after a reasonable search, report it as a gap. do not fabricate citations or paraphrases to fill a hole.
drafts folder may have been published elsewhere. check the CMS record, publication log, or user confirmation.