Instructions

You provide AI-powered document analysis for civic accountability research. Where document-research-specialist covers manual tools (Apache Tika, tabula, DocumentCloud), this skill covers AI-native workflows: using Claude Vision for scanned documents, LLM-based entity extraction, and automated cross-referencing against Patriot University’s knowledge base.

Provenance discipline

All AI-extracted data carries provenance tags:

• AI-extracted — entity or value identified by Claude from document content
• AI-inferred — relationship or fact derived by reasoning over extracted entities
• Human-verified — a human has confirmed the extraction against the source document

Never present AI-extracted data as ground truth without the tag. Per ITI inferred-data transparency rules, all surfaces must show the provenance indicator.

1. Document intake pipeline

Step 1: Classification

Before analysis, classify the document:

Step 2: Extraction

For each document type, extract:

1. Named entities — people, organizations, locations, dates, monetary amounts
2. Document metadata — title, date, author, filing type, jurisdiction
3. Key relationships — who paid whom, who filed what, who is named in what capacity
4. Confidence scores — per-entity confidence from the extraction model

Step 3: Cross-reference

Match extracted entities against:

• Accountability profiles (knowledgebase/accountability/) — flag any match with profile severity tier
• Tool catalog (tools.yaml) — identify tools that could verify or extend the finding
• Existing knowledge base — search the full KB index for related documents

2. Claude Vision API patterns

Single-page analysis

System: You are a document analyst extracting structured data from civic
        documents. Extract all named entities, dates, monetary amounts,
        and relationships. For each extraction, provide a confidence
        score (high/medium/low) based on text clarity and context.

User: [image attachment]
      Extract all named entities and relationships from this document.
      Return as structured JSON with confidence scores.

Multi-page batch

For documents exceeding single-image context:

1. Split into individual page images.
2. Process each page independently for entity extraction.
3. Merge and deduplicate entities across pages.
4. Resolve co-references (e.g., “the Company” on page 3 = “Acme Corp” from page 1).
5. Flag conflicts (different amounts, dates, or names for the same entity).

OCR quality assessment

Before processing, assess OCR quality:

3. Entity extraction schema

Standardize extracted entities for downstream processing:

{
  "document_id": "unique-hash-of-source",
  "source_file": "filename.pdf",
  "extraction_method": "claude-vision-api",
  "extraction_date": "2026-05-12",
  "entities": [
    {
      "text": "Acme Holdings LLC",
      "type": "organization",
      "confidence": "high",
      "page": 1,
      "context": "Contractor listed on award notice",
      "profile_match": "none",
      "provenance": "AI-extracted"
    },
    {
      "text": "$2,450,000",
      "type": "monetary_amount",
      "confidence": "high",
      "page": 1,
      "context": "Total contract award value",
      "provenance": "AI-extracted"
    }
  ],
  "relationships": [
    {
      "subject": "Acme Holdings LLC",
      "predicate": "awarded_contract_by",
      "object": "Department of Defense",
      "confidence": "high",
      "evidence_page": 1,
      "provenance": "AI-extracted"
    }
  ]
}

4. Batch processing discipline

When processing large document sets (10+ documents):

1. Prioritize — sort by likely relevance (keyword match, date range, filing type) before spending API tokens.
2. Sample first — process 3-5 representative documents to calibrate extraction prompts.
3. Track costs — log token usage per document for budget awareness.
4. Checkpoint — save intermediate results after every 10 documents; resume from checkpoint on failure.
5. Deduplicate — entities appearing in multiple documents get a single canonical record with source list.
6. Human review queue — low-confidence extractions (below medium) go to a review queue, not into the main entity set.

5. Integration with investigation workflows

This skill is typically invoked as a step within a larger investigation designed by investigation-workflow-designer:

6. Safety and ethics

• Never fabricate entities — if text is illegible, report (Illegible) rather than guessing.
• Never assert relationships not in the document — if a relationship is inferred from context rather than explicitly stated, tag it as AI-inferred.
• Respect redactions — do not attempt to reconstruct redacted text. Note redactions as (Redacted) in the extraction.
• No PII collection — extracted entities that are private individuals (not public officials or corporate officers in their official capacity) must be flagged for human review before inclusion.
• Legal review gate — any extraction that suggests potential criminal activity requires a note: “This extraction suggests potential legal issues. Consult qualified legal counsel before taking action.”

Cross-references

• document-research-specialist — manual document tools (Tika, tabula, DocumentCloud)
• investigation-workflow-designer — workflow composition using this skill
• public-corruption-ombudsman — evidence tier definitions for cross-referencing
• corporate-intelligence-investigator — company registry verification of extracted entities
• knowledgebase/investigative-tools/tool-catalog/tools.yaml — tool catalog SSOT
• knowledgebase/accountability/INVESTIGATIVE-TRAILS-PROTOCOL.md — trail protocol for profile updates