Skip to content
Back to search

citation-check-skill

Vision-enabled verification gate with web search. Use when users want to (1) verify slides/reports/PDFs/images against authoritative online sources, (2) validate that citations actually exist and say what's claimed, (3) check charts/graphs/tables for accuracy, (4) audit AI-generated content in doc-only mode (no external knowledge). Two modes - search mode validates against web, doc-only mode ensures everything traces to provided documents. Supports content in any language.

Safety scan

A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.

Checked today
Badge for your README
findsafeskills safety rating

Markdown

[![findsafeskills safety rating](/api/listings/serenakeyitan--citation-check-skill/badge.svg)](/listings/serenakeyitan--citation-check-skill)

HTML

<a href="/listings/serenakeyitan--citation-check-skill"><img src="/api/listings/serenakeyitan--citation-check-skill/badge.svg" alt="findsafeskills safety rating" /></a>

Install

i

What was scanned

The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.

SKILL.md · Manifest · 541 lines

View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)

  1. ---
  2. name: citation-check-skill
  3. description: Vision-enabled verification gate with web search. Use when users want to (1) verify slides/reports/PDFs/images against authoritative online sources, (2) validate that citations actually exist and say what's claimed, (3) check charts/graphs/tables for accuracy, (4) audit AI-generated content in doc-only mode (no external knowledge). Two modes - search mode validates against web, doc-only mode ensures everything traces to provided documents. Supports content in any language.
  4. ---
  5. # Citation & Hallucination Checker v2
  6. Verification tool with vision + web search. Validates every claim against authoritative sources or provided documents. Works with content in any language.
  7. **Design principle:** Deterministic verification. Same input → Same output.
  8. ---
  9. ## Two Verification Modes
  10. ### Mode 1: Search Verification (Default)
  11. - Searches web for authoritative sources
  12. - Validates citations actually exist
  13. - Checks if cited sources say what's claimed
  14. - Finds original data for statistics
  15. ### Mode 2: Doc-Only Verification
  16. - User provides source document(s)
  17. - EVERYTHING must trace to those docs
  18. - Flags anything that appears to come from external knowledge
  19. - Trigger: "only use this document" / "verify against the PDF only" / "don't search the web"
  20. ---
  21. ## Two-Pass Architecture
  22. **Critical:** Always use two separate passes. Never interleave extraction and verification.
  23. ### Pass 1: Extraction Only
  24. 1. Read entire document/slides/images
  25. 2. Extract ALL claims using the Claim Extraction Rules below
  26. 3. Output numbered list: `[claim_id] | [claim_text] | [claim_type] | [location]`
  27. 4. **NO verification in this pass**
  28. 5. Present extraction to user for confirmation before proceeding
  29. ### Pass 2: Verification Only
  30. 1. Take Pass 1 output as fixed input
  31. 2. Verify each claim_id in sequential order
  32. 3. **NO re-extraction allowed** — work only with Pass 1 claims
  33. 4. Apply Status Decision Tree to each claim
  34. 5. Generate final report
  35. This prevents "discovering new claims" mid-verification and ensures consistency.
  36. ---
  37. ## Claim Extraction Rules (Exhaustive)
  38. Extract ONLY these claim types. Apply rules strictly — no judgment calls.
  39. ### EXTRACT as claims:
  40. | Type | Pattern | Example |
  41. | --------------- | ---------------------------------------------------------- | ------------------------------------------------ |
  42. | **Statistic** | Any number with unit/context (%, $, count, ratio, decimal) | "92.3% accuracy", "$4.7B market" |
  43. | **Comparative** | X is [comparative] than Y | "3x faster than baseline" |
  44. | **Temporal** | Time-bound assertion | "In 2024, adoption reached..." |
  45. | **Attribution** | Claim tied to source | "According to WHO...", "Smith et al. found..." |
  46. | **Causal** | X causes/leads to/results in Y | "This reduces latency by..." |
  47. | **Existence** | Asserts something exists/is true | "There are 500M users", "The model supports..." |
  48. | **Ranking** | Position claims | "largest", "first", "top 3" |
  49. | **Quote** | Direct quotation | Any text in quotation marks attributed to source |
  50. ### DO NOT extract as claims:
  51. | Type | Example | Reason |
  52. | --------------------------------- | -------------------------------------------------- | ------------------------------- |
  53. | Definitions | "Machine learning is a subset of AI" | Definitional, not factual claim |
  54. | Opinions marked as such | "We believe...", "In our view..." | Explicitly subjective |
  55. | Hypotheticals | "If adoption continues...", "Could potentially..." | Speculative |
  56. | Questions | "What drives growth?" | Not an assertion |
  57. | Future predictions without source | "Will reach $10B by 2030" | Unless citing a forecast report |
  58. | Methodology descriptions | "We used PyTorch 2.0" | Process, not factual claim |
  59. | Acknowledgments | "Thanks to our collaborators" | Not verifiable |
  60. ### Extraction Output Format
  61. ```
  62. [C01] | "Model achieves 96.555% accuracy on ImageNet" | Statistic | Slide 3, bullet 2
  63. [C02] | "Outperforms GPT-4 by 12% on reasoning tasks" | Comparative | Slide 3, bullet 3
  64. [C03] | "According to Chen et al. (2024), transformers scale linearly" | Attribution | Slide 5, para 1
  65. [C04] | "Market size reached $4.7B in 2024" | Statistic + Temporal | Slide 7, chart title
  66. ```
  67. ---
  68. ## Status Decision Tree
  69. Apply this tree to EVERY claim. Follow exactly — no shortcuts.
  70. ```
  71. START
  72. │
  73. ├─ Is this a CITATION claim (references a paper/report/source)?
  74. │ ├─ YES → Go to CITATION VALIDATION
  75. │ └─ NO → Go to STATISTIC/FACT VALIDATION
  76. │
  77. │
  78. CITATION VALIDATION
  79. │
  80. ├─ Step 1: Does the cited source exist?
  81. │ │ Run ALL mandatory search queries (see Search Templates)
  82. │ │
  83. │ ├─ NO → Status: "Citation Not Found"
  84. │ │ Issue: "Cannot locate [citation] in any database"
  85. │ │ STOP
  86. │ │
  87. │ └─ YES → Step 2: Does source contain the claimed topic?
  88. │ │
  89. │ ├─ NO → Status: "Misquoted"
  90. │ │ Issue: "Source exists but does not discuss [topic]"
  91. │ │ STOP
  92. │ │
  93. │ └─ YES → Step 3: Does source support the exact claim?
  94. │ │
  95. │ ├─ YES (exact match) → Status: "Verified"
  96. │ │ Confidence: "exact"
  97. │ │
  98. │ ├─ YES (paraphrase, same meaning) → Status: "Verified"
  99. │ │ Confidence: "paraphrase"
  100. │ │
  101. │ ├─ PARTIALLY (missing context) → Status: "Misleading"
  102. │ │ Issue: "Claim omits critical context: [what's missing]"
  103. │ │
  104. │ └─ NO (contradicts) → Status: "Hallucination"
  105. │ Issue: "Source says [X], claim says [Y]"
  106. │
  107. │
  108. STATISTIC/FACT VALIDATION
  109. │
  110. ├─ Step 1: Can you find an authoritative source?
  111. │ │ Run ALL mandatory search queries (see Search Templates)
  112. │ │
  113. │ ├─ NO (no source found) → Status: "Unverified"
  114. │ │ Issue: "No authoritative source found"
  115. │ │ STOP
  116. │ │
  117. │ └─ YES → Step 2: Do values match EXACTLY?
  118. │ │
  119. │ ├─ YES → Status: "Verified"
  120. │ │ Confidence: "exact"
  121. │ │ STOP
  122. │ │
  123. │ └─ NO → Status: "Numerical Error"
  124. │ Go to NUMERICAL ERROR DETAILS
  125. │
  126. │
  127. NUMERICAL ERROR DETAILS (Academic Precision Mode)
  128. │
  129. ├─ Record:
  130. │ • Source value: [exact number from source]
  131. │ • Claimed value: [number in document being checked]
  132. │ • Deviation: [calculate exact difference]
  133. │ • Source location: [page, table, section]
  134. │
  135. ├─ Classification:
  136. │ • ANY rounding → Numerical Error
  137. │ • ANY truncation → Numerical Error
  138. │ • Significant figures mismatch → Numerical Error
  139. │ • Unit mismatch → Numerical Error
  140. │ • Wrong direction (e.g., increase vs decrease) → Hallucination
  141. │
  142. └─ Exception: If source ITSELF provides rounded figure
  143. • e.g., Source says "96.555% (approximately 97%)"
  144. • Then claiming "97%" → Verified (cite the approximation)
  145. ```
  146. ---
  147. ## Numerical Precision Rules (Academic Standard)
  148. **Default mode: Strict academic precision. Exact numbers only.**
  149. | Rule | Source | Claim | Status |
  150. | -------------------- | ----------- | ----------- | ----------------- |
  151. | Exact match required | 96.555% | 96.555% | ✓ Verified |
  152. | Any rounding = error | 96.555% | 97% | ✗ Numerical Error |
  153. | Any rounding = error | 96.555% | 96.6% | ✗ Numerical Error |
  154. | Truncation = error | 96.555% | 96.5% | ✗ Numerical Error |
  155. | Sig figs must match | 0.834 | 0.83 | ✗ Numerical Error |
  156. | Units must match | 96.555% | 0.96555 | ✗ Numerical Error |
  157. | Direction matters | +12% growth | +15% growth | ✗ Hallucination |
  158. | Order of magnitude | $4.7B | $47B | ✗ Hallucination |
  159. ### Numerical Error Output Format
  160. ```markdown
  161. ### Numerical Error: [Claim ID]
  162. | Field | Value |
  163. |-------|-------|
  164. | Claim | "Model achieves 97% accuracy" |
  165. | Location | Slide 4, bullet 2 |
  166. | Source | Chen et al. (2024), Table 3, p.8 |
  167. | Source value | 96.555% |
  168. | Claimed value | 97% |
  169. | Deviation | +0.445% (rounded up) |
  170. | Status | Numerical Error |
  171. | Fix | Replace with: "Model achieves 96.555% accuracy" |
  172. ```
  173. ---
  174. ## Confidence Classification
  175. | Level | Criteria | Use when |
  176. | ------------------ | ---------------------------------------------------------- | ----------------------------------- |
  177. | **exact** | ≥95% word overlap OR identical number with identical units | Direct quote, exact statistic |
  178. | **paraphrase** | Same fact, different words, no interpretation added | Restated finding |
  179. | **interpretation** | Inference drawn from source data | Calculated from source, synthesized |
  180. **Rule:** When uncertain between levels, use the MORE CONSERVATIVE option and flag for review.
  181. ---
  182. ## Mandatory Search Templates
  183. Run ALL applicable templates. Do not stop after first result.
  184. ### For Academic Citations
  185. ```
  186. Query 1: "[first author last name] [year] [first 3 words of title]"
  187. Query 2: "[full paper title]" site:semanticscholar.org OR site:arxiv.org
  188. Query 3: "[first author] [year] [venue/journal name]"
  189. Query 4: "doi:[DOI]" (if DOI provided)
  190. Query 5: "arxiv:[arxiv_id]" (if arXiv ID provided)
  191. ```
  192. ### For Statistics (Market size, usage numbers, etc.)
  193. ```
  194. Query 1: "[exact number with unit] [topic] [year]"
  195. Query 2: "[topic] [year] statistics report site:statista.com"
  196. Query 3: "[topic] [year] report site:mckinsey.com OR site:gartner.com"
  197. Query 4: "[topic] market size [year] site:gov OR site:edu"
  198. Query 5: "[topic] [number] original source"
  199. ```
  200. ### For Company/Product Claims
  201. ```
  202. Query 1: "[company name] [claim topic] press release [year]"
  203. Query 2: site:[company domain] [claim topic]
  204. Query 3: "[company name] [metric] official announcement"
  205. Query 4: "[company name] [claim] SEC filing" (for public companies)
  206. ```
  207. ### For Health/Medical Claims
  208. ```
  209. Query 1: "[claim topic] site:who.int OR site:cdc.gov OR site:nih.gov"
  210. Query 2: "[claim] systematic review site:cochrane.org"
  211. Query 3: "[claim] meta-analysis pubmed"
  212. ```
  213. ### For Government/Policy Claims
  214. ```
  215. Query 1: "[policy/law name] site:gov"
  216. Query 2: "[statistic] official statistics [country]"
  217. Query 3: "[claim] [agency name] report"
  218. ```
  219. ---
  220. ## Source Authority Hierarchy
  221. When multiple sources found, prefer in this order:
  222. | Rank | Source Type | Examples |
  223. | ---- | ----------------------------- | ------------------------------------------------- |
  224. | 1 | Primary source | Original study, official report, raw data |
  225. | 2 | Government/institutional | WHO, CDC, World Bank, national statistics offices |
  226. | 3 | Peer-reviewed publication | Nature, Science, IEEE, ACM |
  227. | 4 | Industry reports (named) | Gartner, McKinsey, Statista (with methodology) |
  228. | 5 | Reputable news citing primary | NYT, Reuters citing original source |
  229. | 6 | Secondary compilations | Wikipedia (check their sources) |
  230. **Rule:** If only Rank 5-6 sources found, status = "Unverified" with note "Only secondary sources found"
  231. ---
  232. ## Multi-Source Verification (Search Mode)
  233. A claim achieves "Verified" status only if:
  234. | Condition | Sources Required |
  235. | ---------------------- | --------------------------------------------------- |
  236. | Primary source found | 1 (if authoritative: .gov, peer-reviewed, official) |
  237. | Only secondary sources | ≥2 independent sources agreeing |
  238. | Sources conflict | Status = "Unverified", note the conflict |
  239. ---
  240. ## Tie-Breaker Rules
  241. When uncertain, apply these rules. No judgment calls.
  242. | Situation | Rule |
  243. | ----------------------------------------- | ------------------------------------------------------------ |
  244. | Missing date on claim | Assume refers to most recent year available; flag "needs date" |
  245. | Conflicting sources | Use most recent authoritative source; cite both; note conflict |
  246. | Source not found after all queries | Status = "Unverified" (NOT "Hallucination") |
  247. | Number differs due to currency conversion | Flag as "Needs clarification: currency/units" |
  248. | Same org, multiple reports | Use most recent; cite with date |
  249. | Claim uses "approximately" or "about" | Still verify base number is in valid range (±10% of source) |
  250. | Source is paywalled | Note "Source behind paywall, unable to verify exact text" |
  251. | Source is in different language | Translate and verify; note translation |
  252. ---
  253. ## Visual Data Verification
  254. For every chart, graph, table, or diagram:
  255. ### Step 1: Extract Data Points
  256. - Read ALL values from the visual
  257. - Record: axis labels, units, scale, legend
  258. - Note any visual distortions (truncated axes, 3D effects, etc.)
  259. ### Step 2: Find Source
  260. - Search mode: Run search templates for the data
  261. - Doc-only mode: Locate in source document
  262. ### Step 3: Compare Value-by-Value
  263. ```markdown
  264. | Visual Element | Extracted Value | Source Value | Status |
  265. |----------------|-----------------|--------------|--------|
  266. | Bar 1 (2022) | 45% | 45.0% | ✓ Verified |
  267. | Bar 2 (2023) | 62% | 58.3% | ✗ Numerical Error |
  268. | Bar 3 (2024) | 78% | Not in source | ✗ Hallucination |
  269. ```
  270. ### Step 4: Check Visual Integrity
  271. | Check | Issue Type |
  272. | --------------------------------------- | -------------------------------------- |
  273. | Y-axis starts at non-zero | "Visual Distortion: axis manipulation" |
  274. | 3D effects distort proportions | "Visual Distortion: 3D exaggeration" |
  275. | Missing error bars when source has them | "Misleading: uncertainty omitted" |
  276. | Different time ranges than source | "Misleading: cherry-picked timeframe" |
  277. ---
  278. ## Doc-Only Mode Workflow
  279. **Trigger phrases:**
  280. - "only use this document"
  281. - "don't search the web"
  282. - "verify against the PDF only"
  283. - "everything should be from the source"
  284. ### Step 1: Index Source Document
  285. Build complete index before any verification:
  286. ```
  287. SOURCE INDEX
  288. Document: [filename]
  289. Pages: [count]
  290. Page 1:
  291. - Text: [summary of content]
  292. - Statistics: [list all numbers with context]
  293. - Tables: [Table 1: columns X, Y, Z]
  294. - Figures: [Figure 1: shows X]
  295. Page 2:
  296. ...
  297. ```
  298. ### Step 2: Apply Two-Pass Architecture
  299. Same as search mode, but verification uses ONLY the source index.
  300. ### Step 3: Trace Each Claim
  301. ```
  302. Claim: [C01] "Model achieves 92% accuracy"
  303. Search index for: "92", "accuracy", "performance"
  304. ├─ Found: Section 4.2, p.8 — "Our model achieves 92.1% accuracy"
  305. │ └─ Status: Numerical Error (92% vs 92.1%)
  306. │
  307. OR
  308. │
  309. ├─ Not found in index
  310. │ └─ Status: "Not in Source"
  311. │ Issue: "This claim cannot be traced to the provided document"
  312. │ Likely: External knowledge / hallucination
  313. ```
  314. ### Step 4: Flag ALL External Knowledge
  315. In doc-only mode, ANY claim not traceable to source = problem
  316. ```markdown
  317. ### External Knowledge Detected
  318. These claims are NOT in the provided document:
  319. | Claim ID | Claim | Status | Issue |
  320. |----------|-------|--------|-------|
  321. | C07 | "This method is widely adopted in industry" | Not in Source | Appears to be from model training data |
  322. | C12 | "Published in Nature 2024" | Not in Source | Publication venue not mentioned in source |
  323. ```
  324. ---
  325. ## Output Format
  326. ### Summary Block (Always First)
  327. ```markdown
  328. ## Verification Report
  329. **Mode:** [Search / Doc-Only]
  330. **Document:** [filename or description]
  331. **Generated:** [timestamp]
  332. ### Summary
  333. | Metric | Count |
  334. |--------|-------|
  335. | Total claims extracted | X |
  336. | Verified | Y |
  337. | Numerical Error | Z |
  338. | Unverified | A |
  339. | Hallucination | B |
  340. | Misleading | C |
  341. | Not in Source (doc-only) | D |
  342. **Overall Status:** [PASS: All verified / FAIL: Issues found]
  343. ```
  344. ### Detailed Findings (Grouped by Status)
  345. ```markdown
  346. ### ✓ Verified Claims (N)
  347. | ID | Claim | Source | Location | Confidence |
  348. |----|-------|--------|----------|------------|
  349. | C01 | "92.1% accuracy" | Chen et al. 2024 | Table 3, p.8 | exact |
  350. ### ✗ Numerical Errors (N)
  351. | ID | Claim | Source Value | Claimed Value | Deviation | Fix |
  352. |----|-------|--------------|---------------|-----------|-----|
  353. | C03 | "97% accuracy" | 96.555% | 97% | +0.445% | Use 96.555% |
  354. ### ✗ Hallucinations (N)
  355. | ID | Claim | Issue | Source Says |
  356. |----|-------|-------|-------------|
  357. | C05 | "3x faster" | Contradicts source | Source: 2.1x faster |
  358. ### ⚠ Unverified (N)
  359. | ID | Claim | Issue |
  360. |----|-------|-------|
  361. | C08 | "$50B market" | No authoritative source found |
  362. ### ⚠ Misleading (N)
  363. | ID | Claim | Issue | Missing Context |
  364. |----|-------|-------|-----------------|
  365. | C10 | "Best performance" | Cherry-picked metric | Only on subset; overall performance lower |
  366. ```
  367. ### Sources Consulted
  368. ```markdown
  369. ### Sources
  370. | ID | Citation | Type | URL | Used For |
  371. |----|----------|------|-----|----------|
  372. | S1 | Chen et al. (2024) | arxiv | https://arxiv.org/... | C01, C02, C03 |
  373. | S2 | Statista Market Report | report | https://statista.com/... | C08 |
  374. ```
  375. ---
  376. ## Critical Rules
  377. 1. **Two-pass always** — Extract first, verify second. Never interleave.
  378. 2. **Every claim gets checked** — No exceptions, no skipping "obvious" ones.
  379. 3. **Exact numbers only** — 96.555% ≠ 97% in academic mode.
  380. 4. **Find the origin** — Don't accept secondary sources citing unknown primaries.
  381. 5. **Run ALL search templates** — Don't stop after first result.
  382. 6. **Citations must be real** — Search to confirm papers/reports exist.
  383. 7. **Check what sources actually say** — A real paper can still be misquoted.
  384. 8. **In doc-only mode, flag ALL external knowledge** — Even if it's true.
  385. 9. **When uncertain, be conservative** — "Unverified" is safer than false "Verified".
  386. 10. **Follow tie-breaker rules** — No ad-hoc judgment calls.
  387. ---
  388. ## Language Support
  389. - Accepts content in any language
  390. - Searches in the appropriate language for sources
  391. - Reports in the same language as user's request
  392. - Cross-language verification supported (e.g., Chinese slides citing English papers)
  393. - When translating for verification, note: "Translated from [language]"
  394. ---
  395. ## Output Format Options
  396. **Quick:** Summary + critical issues only (Numerical Errors, Hallucinations, Unverified)
  397. **Full:** Complete traceability report with all claims
  398. **JSON:** Machine-readable audit (see references/citation_schema.json)
  399. ---
  400. ## Changelog
  401. **v2.0** — Consistency update
  402. - Added Two-Pass Architecture (extract → verify separation)
  403. - Added exhaustive Claim Extraction Rules
  404. - Added Status Decision Tree (deterministic classification)
  405. - Added strict Numerical Precision Rules (academic mode)
  406. - Added Mandatory Search Templates
  407. - Added Multi-Source Verification requirements
  408. - Added Tie-Breaker Rules for edge cases
  409. - Added Confidence Classification thresholds
README.md · README · 120 lines

View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)

  1. # Citation Check v2
  2. > No citation → no output.
  3. Citation Check is a small, opinionated module for **detecting and blocking hallucinated or missing citations** in LLM outputs.
  4. Built for research, slides, and agent pipelines where correctness matters.
  5. ---
  6. ## Demo - Identified Issues
  7. ### TRY [Kael.im](https://kael.im/home)(NotebookLM slides alternative) and register at this link for 100 pages free daily quota (nbp)!
  8. https://github.com/user-attachments/assets/99d427cf-992f-455c-897b-a8f8a132271c
  9. <img width="1219" height="713" alt="Screenshot 2026-01-25 at 13 56 07" src="https://github.com/user-attachments/assets/660592cd-ca53-46ae-8064-28fad9e90b44" />
  10. ### Check out [Awesome Notebooklm Slides Prompts List](https://github.com/serenakeyitan/awesome-notebookLM-prompts/tree/main) with all viral templates on X!
  11. ## How to Use?
  12. 1. Download `citation-check-skill.zip` from [Releases](https://github.com/serenakeyitan/citation-check-skill/releases/latest)
  13. 2. Upload it to Claude Code (Profile → Settings → Capabilities → Skills → Add)
  14. 3. Use it by telling Claude: "use citation checker" or "use skills"
  15. **Note:** Don't use the "Code → Download ZIP" button - it adds `-main` to the folder name. Always download from Releases!
  16. ## What it does
  17. - ✅ Checks that every factual claim has a citation
  18. - ✅ Verifies that cited sources actually exist
  19. - ✅ Validates claim–citation consistency (exact numbers, not rounded)
  20. - ✅ Supports vision: reads slides, PDFs, charts, tables
  21. - ✅ Two modes: web search verification OR doc-only verification
  22. If a check fails, the output should be blocked or regenerated.
  23. ---
  24. ## Why
  25. Most tools add citations *after* generation.
  26. This enforces citations as a **generation constraint**.
  27. If your agent can't cite a claim, it shouldn't make it.
  28. ---
  29. ## v2.0 — Consistency Update
  30. Same input → Same output. Key changes:
  31. | Feature | What it does |
  32. | ------------------------------ | ------------------------------------------------------------ |
  33. | **Two-Pass Architecture** | Extract all claims first, then verify. No more inconsistent results. |
  34. | **Claim Extraction Rules** | Explicit taxonomy — statistics, comparatives, attributions, etc. |
  35. | **Status Decision Tree** | Deterministic classification: Verified → Numerical Error → Hallucination |
  36. | **Academic Precision Mode** | 96.555% ≠ 97%. Exact numbers only. |
  37. | **Mandatory Search Templates** | Runs ALL query patterns, not ad-hoc searches |
  38. | **Tie-Breaker Rules** | No judgment calls on edge cases |
  39. ---
  40. ## Typical flow
  41. ```text
  42. LLM output (slides / report / PDF)
  43. ↓
  44. Pass 1: Extract all claims
  45. ↓
  46. Pass 2: Verify each claim
  47. ↓
  48. pass → ship
  49. fail → regenerate / fix
  50. ```
  51. ---
  52. ## Two Modes
  53. ### Mode 1: Search Verification (default)
  54. - Searches web for authoritative sources
  55. - Validates citations actually exist
  56. - Checks if cited sources say what's claimed
  57. ### Mode 2: Doc-Only Verification
  58. - User provides source document(s)
  59. - EVERYTHING must trace to those docs
  60. - Flags any external knowledge as hallucination
  61. Trigger doc-only mode with: *"only use this document"* / *"verify against the PDF only"*
  62. ---
  63. ## Status Classifications
  64. | Status | Meaning |
  65. | -------------------- | ----------------------------------------------------- |
  66. | ✅ Verified | Exact match with source |
  67. | ⚠️ Numerical Error | Values don't match (e.g., 97% vs 96.555%) |
  68. | ⚠️ Unverified | No authoritative source found |
  69. | ❌ Hallucination | Contradicts source or fabricated |
  70. | ❌ Misleading | Cherry-picked or missing context |
  71. | ❌ Citation Not Found | Referenced paper/report doesn't exist |
  72. | ❌ Not in Source | Claim can't be traced to provided doc (doc-only mode) |
  73. ---
  74. ## Installation
  75. It's a Claude / ChatGPT skill.
  76. - **Claude:** [Agent Skills Documentation](https://docs.anthropic.com/en/docs/agents-and-tools/agent-skills/overview)
  77. - **ChatGPT:** [Codex Skills](https://developers.openai.com/codex/skills)
  78. ---
  79. ## License
  80. MIT

Ratings and comments

Sign in to rate this skill.

Sign in

Free to use. If it’s helped, pass it on.