verify-citations – Writing & Content skill for AI agents

Citation Verification

Verify that citations in a document actually resolve and support the claims they're attached to. Uses the Stipple API (free anonymous tier, no signup) for citation resolution, arithmetic recomputation, and unsupported-claim detection.

When to use

• Before submitting or publishing a research report, tender response, or whitepaper
• Reviewing an LLM-generated document (LLM citations are plausibly-formatted and frequently wrong)
• Due diligence on third-party reports
• Academic reference checking

Instructions

1. Get the document. Ask the user for a URL to the report, or a local file path (PDF, DOCX, Markdown). If the user pastes text directly, skip to step 3 with text input.

2. Run verification. POST the document to Stipple's citation verification endpoint:

   curl -X POST https://www.stipple.sh/v1/verify-references \
     -F "[email protected]" \
     -H "Authorization: Bearer $STIPPLE_API_KEY"
   

The anonymous free tier works without the Authorization header. Deep mode (?deep=true) costs more credits but cross-checks citations against live web sources.

3. Interpret the response. The result includes:

• verification_coverage — percentage of claims verified (e.g. "78%")
• citations[] — per-citation status: resolved and matching, resolved but mismatched, or unresolvable, with the issue explained
• arithmetic[] — recomputed figures vs stated figures (flags decimal shifts, wrong sums)
• unsupported_claims[] — claims with no citation at all

4. Report honestly. Present results as *verification coverage*, not a truth verdict:

• "21/27 citations resolve and match"
• "[x] FY24+FY25 revenue stated $4.2m, actual $3.7m"
• "[!] 'industry-leading accuracy' — no source in document"
• Unverified ≠ false. The goal is telling the user *which* claims are backed and which aren't.

5. Offer remediation. For failed citations, suggest: fixing the decimal shift, finding the correct source, or removing the unsupported claim.

Output format

Verification coverage: 78%

Citations: 21/27 resolve and match
  [+] "ABS unemployment 4.1% April 2026" — matches abs.gov.au
  [-] "AI adoption grew 340% in 2025" — source states 34%, decimal shifted

Arithmetic: 12/13 recompute correctly
  [x] FY24 + FY25 revenue — stated $4.2m, actual $3.7m

Unsupported claims: 2
  [!] "industry-leading accuracy" — no source in document

Limitations and Safety

• Uploading a report or manuscript sends its contents to a hosted third party; deep

mode also retrieves external sources. Obtain approval before transmission, remove

confidential or personal material, and verify current retention and deletion terms.

• Resolution and textual support do not establish that a source is authoritative or

that a claim is true. Inspect consequential citations in the primary source and

preserve page, edition, date, and access context.

• Treat unresolved or unsupported claims as review items, not proof of fabrication,

and keep a human reviewer responsible for publication or academic decisions.

Notes

• Works on PDF, DOCX, MD, TXT. For pasted text, POST JSON: {"text": "..."}
• Deep verification (deep=true) is slower and costs more credits but resolves citations against live sources
• Pairs well with verify-document (is the source doc itself authentic?) run first
• Anonymous free tier: shared weekly allowance. Get a free key at https://www.stipple.sh for your own metering