emaduneme Verification & Fact-Checking Skill
What this skill does
Provides journalists with a structured verification pipeline:
- Decompose complex claims into individually verifiable sub-claims
- Search for prior fact-checks on a claim
- Find primary sources for statistics and data
- Check quote attribution and accuracy
- Flag red flags in images, videos, or documents
- Build a documented verification trail
Core principle
Verification is documentation. For every fact you publish, you need to be able to answer:
- Who said or produced it?
- Where can I point a reader to see it themselves?
- What would change my mind?
Workflow
Step 1 — Decompose the claim
Break any complex claim into the smallest individually verifiable units.
Example claim: "The city's crime rate dropped 40% after the new police program launched in 2022."
Sub-claims:
- There is a police program that launched in 2022. (Verifiable: check city records)
- Crime rate dropped 40%. (Verifiable: check official crime statistics)
- The drop is causally linked to the program. (Harder — requires analysis)
Run decomposition automatically:
python scripts/claim_checker.py --claim "The city's crime rate dropped 40% after the new program" --mode decompose
Step 2 — Search for existing fact-checks
Use web_search:
"[claim keywords]" fact check OR debunked OR verified
"[claim keywords]" site:snopes.com OR site:politifact.com OR site:factcheck.org
Or run the fact-check search tool:
python scripts/claim_checker.py --claim "Claim text here" --mode factcheck-search
This queries:
- Google Fact Check API (aggregates Snopes, PolitiFact, FactCheck.org, AP Fact Check, Reuters Fact Check)
- Google News for recent debunks
Step 3 — Find primary sources
For statistics: Go to the original producing organization, not the news article that reported them.
Common primary source categories:
| Claim type | Go here first |
|---|---|
| Crime statistics | FBI UCR, local PD annual reports |
| Health/medical | CDC, NIH, peer-reviewed journals (PubMed) |
| Economic data | BLS, Census Bureau, Federal Reserve |
| Environmental | EPA, NOAA, peer-reviewed science |
| Corporate claims | SEC filings, company annual reports |
| Government claims | Official agency websites, FOIA records |
| Historical events | Library of Congress, National Archives |
web_search pattern: "[statistic]" site:cdc.gov or "[claim]" site:bls.gov filetype:pdf
Step 4 — Verify the quote
For attributed quotes:
- Do you have a recording? Use that — do not rely on notes alone.
- Does the quote appear in a prior published source? Find the original, not a republication.
- Search:
"[exact quote phrase]"— Google will surface the original publication if it's widely reported. - For social media quotes: archive the post before it can be deleted.
python scripts/claim_checker.py --mode verify-quote --quote "Exact quote text" --source "Name"
Step 5 — Image/video verification
Quick checks (no tools needed):
- Reverse image search: drag into Google Images or TinEye
- Check image metadata: use ExifTool (
exiftool image.jpg) - Look for inconsistencies: shadows, edges, lighting, duplicated textures
For deeper investigation:
- InVID/WeVerify browser extension: video frame analysis
- FotoForensics.com: error level analysis (free, upload required)
- AI-generated image detection: Hive Moderation, AI or Not
python scripts/verify_media.py --image photo.jpg --check-metadata --check-exif
Step 6 — Build the verification trail
For every verified claim, document:
CLAIM: "[Exact claim text]"
STATUS: [Verified / Unverified / False / Partially true / Needs more evidence]
PRIMARY SOURCE: [URL or document name]
ACCESSED: [Date]
NOTES: [What you found, any caveats or context]
VERIFIED BY: [Reporter name]
Save the full trail:
python scripts/verification_log.py --add --claim "Crime rate dropped 40%" \
--status "Partially verified" \
--source "https://fbi.gov/crime-data/2022" \
--notes "40% figure applies only to violent crime, not overall crime rate"
Step 7 — Pre-publication checklist
Before sending the story to an editor:
- Every statistic has a named primary source
- Every quote is from a recording or contemporaneous document
- Every claim with a named person has been given opportunity to respond
- Contested or contested claims are labeled as such in the story
- Images and video have been reverse-searched
- Verification trail is saved and accessible
Python Scripts
scripts/claim_checker.py
Three modes:
decompose: Breaks a claim into sub-claimsfactcheck-search: Queries Google Fact Check Tools API + Google Newsverify-quote: Searches for the original source of a quote
python scripts/claim_checker.py --claim "TEXT" --mode decompose
python scripts/claim_checker.py --claim "TEXT" --mode factcheck-search
python scripts/claim_checker.py --quote "QUOTE TEXT" --source "Person Name" --mode verify-quote
scripts/verification_log.py
Maintains a JSON log of verification decisions for a story. Produces a clean verification trail that can be shared with editors.
python scripts/verification_log.py --add --claim "TEXT" --status "Verified" --source "URL"
python scripts/verification_log.py --list --story "Story name"
python scripts/verification_log.py --export --story "Story name" --output trail.txt
scripts/verify_media.py
Checks image/video file metadata (EXIF) for signs of manipulation or misdating.
Requires: pip install Pillow
python scripts/verify_media.py --image photo.jpg
python scripts/verify_media.py --image photo.jpg --check-exif --verbose
Notes for the AI
- Never present a web search result as verification — it is a lead, not proof
- If a claim cannot be verified with available tools, say so explicitly and suggest what would be needed
- Distinguish between: (a) "we found no evidence this is true" and (b) "we found evidence this is false" — these are different conclusions
- When using the fact-check API, cross-reference at least 2 independent fact-checkers before calling something debunked
- For contested empirical claims (policy effectiveness, causality), note the range of expert opinion, not just one side
- Always flag when a primary source has a potential conflict of interest