PDF Statement Parser

PDF Statement Parser

Table of Contents

• Overview
• Input contract
• Workflow
• Document type playbooks
• Output contract
• Confidence scoring
• Guardrails
• Example

Overview

Bank, brokerage, retirement, HSA, and mortgage statements arrive as PDFs in many house-styles. This skill is the prompt-driven extraction methodology: identify the document type, locate the canonical fields by their labels (not position), pull every transaction or holding row, and emit a strict JSON record. All monetary values become integer cents, all dates become ISO 8601 (YYYY-MM-DD).

Use directly with Claude's PDF reading: Read the PDF, then run this skill's workflow on the visible content. No external library needed.

Input contract

The caller provides:

• pdf_path — absolute path to a single statement PDF.
• expected_type (optional) — one of checking_statement | savings_statement | credit_card_statement | brokerage_statement | 401k_statement | hsa_statement | mortgage_statement | tax_form | insurance_doc | unknown. If absent, the skill detects it.
• known_accounts (optional) — array of {id, institution, mask, type} for matching.

Workflow

- [ ] Step 1: Read the PDF (first pass: pages 1-2 for header)
- [ ] Step 2: Detect document type from header keywords
- [ ] Step 3: Extract account identity (institution + mask + type)
- [ ] Step 4: Extract statement period (start/end dates)
- [ ] Step 5: Extract balance markers (opening / closing / available)
- [ ] Step 6: Extract line items (transactions, holdings, or events)
- [ ] Step 7: Normalize to integer cents and ISO dates
- [ ] Step 8: Score confidence on each field; emit JSON

Step 1 — Read the PDF

Use the Read tool with the PDF path. For PDFs > 10 pages, read in chunks: pages 1–5 (header + first transactions), then subsequent ranges as needed. Track total page count.

Step 2 — Detect document type

Look for these header signals (in order of authority):

If two signals tie, prefer the one in the document title or page header.

Step 3 — Extract account identity

• Institution: top-of-page logo text or "Issued by" / "Statement from" label.
• Mask: last 4 digits — search for ****1234, xxxx1234, Account ending in 1234, or the redacted form the institution uses. Always store as ****1234.
• Account type: derived from §2 above plus visible labels ("Checking", "Visa Signature", "Roth IRA").
• Owner names: full names of account holders (used for matching to household.json members downstream).

Match against known_accounts by (institution, mask). If no match, mark account_match: "new" and propose an account_id like acc_<type>_<incrementing>.

Step 4 — Extract statement period

Find labels: "Statement Period", "Cycle Date", "Closing Date", "Period Beginning … Period Ending". Output:

• period_start — first day covered (ISO).
• period_end — last day covered, also the statement closing date (ISO).

If only a closing date is present, set period_end = closing_date and period_start = null with confidence: 0.6 on the period.

Step 5 — Extract balance markers

For cash/credit accounts: opening_balance_cents, closing_balance_cents, available_balance_cents (if present). For credit cards, also statement_balance_cents, minimum_payment_cents, payment_due_date.

For brokerage / 401k / HSA: total_value_cents, cash_cents, invested_cents. Skip opening/closing transaction balances — these are holdings statements.

For mortgages: current_principal_cents, monthly_pi_cents, monthly_escrow_cents, next_payment_due, extra_principal_ytd_cents.

Step 6 — Extract line items

Three modes:

Cash/credit transactions. For each row in the transactions table, capture: date, post_date (if separate column), description_raw, amount. Sign convention: outflows negative, inflows positive. If statement uses two columns ("Withdrawals" and "Deposits"), set sign accordingly. Preserve description_raw exactly as printed — the bookkeeper will normalize the merchant name later.

Brokerage / 401k / HSA holdings. For each holding row: symbol, description, shares (decimal), price (per-share), value, cost_basis (if shown), asset_class (mapped from fund description: see playbook below).

Mortgage activity. Recent payments table: each entry with date, total_paid_cents, principal_cents, interest_cents, escrow_cents, extra_principal_cents.

Tax forms. Form-specific fields per playbook (W-2 boxes, 1099-DIV ordinary/qualified dividends, 1098 mortgage interest paid, 5498-SA HSA contributions).

Step 7 — Normalize

• Money: $1,247.50 → 124750 cents. Negative outflow: -$87.42 → -8742.
• Dates: 01/15/26 → 2026-01-15. Two-digit years assume current century unless context says otherwise.
• Trim whitespace from descriptions; preserve case.

Step 8 — Confidence scoring

Per field, assign a confidence in [0.0, 1.0]:

• 1.0 — extracted directly from a labeled field.
• 0.85–0.95 — extracted from positional context (column header inferred, value clear).
• 0.6–0.8 — inferred from surrounding text or partially OCR'd.
• < 0.6 — flag for human review.

The overall record gets confidence = min(per-field confidences for required fields).

Document type playbooks

checking_statement / savings_statement

Required: institution, mask, period_start, period_end, opening_balance_cents, closing_balance_cents, transactions[]. Reconcile: opening + Σ(transactions.amount) = closing within ±$1.

credit_card_statement

Required: institution, mask, period_start, period_end, opening_balance_cents, closing_balance_cents, statement_balance_cents, minimum_payment_cents, payment_due_date, transactions[]. Note: credit-card outflows from the cardholder's perspective are purchases (positive on the card balance) — but for the unified household ledger, treat purchases as negative (money leaving the household) and payments to the card as positive (or as internal transfers). Document the sign convention you used.

brokerage_statement

Required: institution, mask, account_type, period_end, total_value_cents, holdings[]. Optional: cash_cents, transactions[] (buys/sells/dividends). Asset-class mapping:

If unmappable, leave asset_class: null with confidence 0.5.

401k_statement

Required: institution, owner, period_end, total_balance_cents, ytd_contribution_cents (if shown), employer_match_ytd_cents (if shown), holdings[]. Watch for vesting columns — record vested_cents separately if shown.

hsa_statement

Required: institution, period_end, cash_cents, invested_cents, total_cents, ytd_contribution_cents, ytd_distribution_cents (if shown). Note coverage type if printed: self_only or family.

mortgage_statement

Required: institution, mask, period_end, current_principal_cents, monthly_pi_cents, monthly_escrow_cents, next_payment_due, recent_payments[]. Capture the rate APR if printed.

tax_form

Required: form_type (W-2, 1099-DIV, 1099-INT, 1099-B, 1098, 5498-SA, 1095), tax_year, issuer, recipient, key_boxes (form-specific). Defer interpretation to the tax-compliance agent — this skill just captures the boxes.

Output contract

Emit a single JSON object:

{
  "document_type": "checking_statement",
  "source_path": "inbox/2026-01-15-batch/checking-jan.pdf",
  "institution": "Bank Name",
  "account": {
    "mask": "****1234",
    "type": "checking",
    "owners": ["Jane Doe", "John Doe"],
    "match": "acc_chk_001",
    "match_confidence": 0.95
  },
  "period": {
    "start": "2025-12-15",
    "end": "2026-01-14"
  },
  "balances": {
    "opening_cents": 1124300,
    "closing_cents": 1247500,
    "available_cents": 1247500
  },
  "transactions": [
    {
      "date": "2025-12-16",
      "post_date": "2025-12-17",
      "description_raw": "TRADER JOE'S #123 PORTLAND OR",
      "amount_cents": -8742,
      "confidence": 0.97
    }
  ],
  "reconciliation": {
    "expected_closing_cents": 1247500,
    "computed_closing_cents": 1247500,
    "diff_cents": 0,
    "ok": true
  },
  "extracted_at": "2026-01-20T14:00:00Z",
  "overall_confidence": 0.94,
  "warnings": []
}

For brokerage/401k/HSA replace transactions with holdings and balance fields per playbook.

Confidence scoring

Reduce overall confidence and add a warnings[] entry whenever:

• Reconciliation diff > $1 → warning: "reconcile_failed: diff $X.XX" and overall_confidence: max(0.5, current * 0.7).
• Any required field missing → warning: "missing_required: <field>" and overall_confidence ≤ 0.5.
• OCR-quality page (illegible characters) → warning: "low_ocr_quality_page_<n>" and apply the per-field confidence penalty.
• Unknown asset class on > 20% of holdings → warning: "unknown_asset_class_high".

Guardrails

• Never invent values. If a field is not visible in the PDF, set null and add a warning. Do not interpolate.
• Never normalize merchant names here. That is the bookkeeper's job; preserve description_raw byte-for-byte.
• Never categorize transactions here. Categorization is downstream (transaction-categorizer).
• Always reconcile for cash/credit statements before declaring ok: true.
• Currency: assume USD unless the statement says otherwise. If non-USD, capture the currency code; do not auto-convert.
• Account mask only. Never extract or store the full account number.

Example

Given a 4-page Chase checking statement for the cycle 2025-12-15 to 2026-01-14 with opening balance $11,243.00, three deposits totaling $4,832.00, twenty-seven debits totaling $3,600.00, and closing balance $12,475.00, the parser emits:

• document_type: "checking_statement"
• institution: "Chase"
• account.mask: "****1234", account.type: "checking"
• period.start: "2025-12-15", period.end: "2026-01-14"
• balances.opening_cents: 1124300, balances.closing_cents: 1247500
• transactions: [...] — 30 entries, each with description_raw, amount_cents (signed), and per-field confidence ≥ 0.92
• reconciliation.ok: true, diff_cents: 0
• overall_confidence: 0.95, warnings: []