verify

Original：🇺🇸 English
Translated

Verify AI claims against source documents using the DeepCitation API

2installs
Sourcedeepcitation/skills
Added on

NPX Install

npx skill4agent add deepcitation/skills verify

Tags

Translated version includes tags in frontmatter
deepcitation-apiai-claim-verificationsource-document-validationworkflow-guidanceartifact-management

SKILL.md Content

View Translation Comparison →

/verify — DeepCitation Verification

Verify claims against source documents using the DeepCitation API, saving JSON artifacts at each step.

Prerequisites

  • DEEPCITATION_API_KEY
    environment variable must be set. If not set, try loading saved credentials:
    bash
    eval "$(npx -y deepcitation env 2>/dev/null)"
    If that also fails, tell the user to run 
    npx -y deepcitation login
    first, then retry.
  • Source files (PDF, DOCX, images, etc.) must be accessible on disk or via URL
  • Accepted file types: PDF, images (JPG, PNG), Office files (DOCX, XLSX, PPTX), CSV, TSV, ODF

Key Rules

  • page_id
    and 
    line_ids
    MUST come from the 
    deepTextPromptPortion
     — see rules/line-ids.md for details.
  • Coverage audit: After generating citations, spawn a subagent to audit the report/chat and confirm all facts, sources, names, dates, and values have deepcitations. The subagent should flag any uncited claims.

Workflow

Step 0: Analyze input

Load rules/analyze-input.md to determine which path to follow (A–F) based on what's in the conversation, arguments, and working directory.

Step 1: Prepare sources

Load rules/prepare-sources.md for CLI usage and what to retain from each prepare response.

Step 2: Build citations

Load rules/build-citations.md for the full citation-building workflow:
  • Path A: Existing 
    <<<CITATION_DATA>>>
    — skip to Step 3
  • Path B: Existing HTML with claims — identify claims, build citation data, generate keys, annotate HTML
  • Path C: Generate new cited response from scratch using the canonical citation format spec

Step 3–5: Verify, inject, validate

Load rules/verify-and-inject.md for verification, HTML injection, and pre-delivery validation.

When building citations

When dealing with 
lineIds
and 
pageNumber
values, load rules/line-ids.md for how sparse line IDs work and how to find the right one.
When annotating HTML with 
data-citation-key
attributes, load rules/annotate-html.md for placement rules, key-map building, and citation drawer triggers.

Output artifacts

All artifacts are saved in 
.deepcitation/
. Use 
{topic}-{timestamp}
naming so re-runs don't clobber each other:
FileContents
prepare-{source}-{timestamp}.json
Upload response with 
attachmentId
and 
deepTextPromptPortion
llm-output-{timestamp}.txt
Full LLM response including 
<<<CITATION_DATA>>>
block
citations-{timestamp}.json
Extracted 
CitationRecord
with human-readable keys
citations-keyed-{timestamp}.json
Re-keyed citations with hashed keys (from 
keygen
)
key-map-{timestamp}.json
Human-readable key → hashed key mapping
annotated-{timestamp}.html
HTML with 
data-citation-key
attributes (before injection)
verify-response-{timestamp}.json
Verification results with statuses and evidence
injected-{timestamp}.html
Injected HTML with CDN runtime (Path B output)

Important rules

  • Product name: Always "DeepCitation" (never "DeepCite")
  • Track attachmentId: Always retain the 
    attachmentId
    from Step 1 — it's the key for verification lookups
  • Strip before display: Use 
    extractVisibleText()
    to remove 
    <<<CITATION_DATA>>>
    before showing text to user
  • CitationRecord is an object: Check emptiness with 
    Object.keys(citations).length === 0
  • API key security: Never log or display 
    DEEPCITATION_API_KEY
  • Verbatim quotes: 
    fullPhrase
    must be copied exactly from the source — do not paraphrase

References

ARGUMENTS: $ARGUMENTS