The one-shot prompt.

GOAL: one self-contained, OFFLINE interactive HTML dashboard that consolidates my bank and
credit-card statement exports (CSV/XLS/XLSX — several accounts, different column layouts) into a
personal-finance intelligence report. Do it end to end, locally. Nothing leaves my machine. Use
subagents where they help. Verify it in a browser before you tell me it's done.

━━ STEP 0 — pick your path ━━
Look in this folder.
• IF you see `public/dashboard.html` and `DATA_CONTRACT.md` (you're in the claude-finance-dashboard
  repo): SCAFFOLD from them — do NOT redesign anything.
    – `public/dashboard.html` is a finished, working, self-contained renderer (Apache ECharts
      inlined). It draws everything from one inlined object: `window.FIN = {…}`.
    – `DATA_CONTRACT.md` documents that object's exact shape, key by key.
    – Your only job: (1) build a local Python pipeline that turns MY statements into that exact
      `window.FIN` JSON, and (2) make my dashboard by copying `public/dashboard.html`, replacing its
      `window.FIN = {…}` block with mine, and stripping the demo branding ("SAMPLE · FICTIONAL DATA",
      "// DEMO", the demo <title>) so it shows my real numbers. Match the contract precisely so every
      chart, table and the audit section just work; drop (don't break) any section I have no data for.
• ELSE (empty folder): build the renderer too, from scratch, per "FULL SPEC" at the bottom.

━━ THE DATA MODEL (do this either way — it's the important part) ━━
Build a small Python pipeline (pandas/openpyxl only if a file needs it):

1) PARSE every file into one normalized ledger: [date, account, account_type, description, outflow,
   inflow, currency]. Treat each file as one account (LibreOffice headless or xlrd for .xls). Where a
   statement prints its own totals (e.g. an AMEX "charges/payments" summary), reconcile your parse to
   them. De-dupe overlapping re-exports by (date, description, amount) with max-count semantics so
   nothing is doubled or lost.

2) CLASSIFY every row:
   • flow_type — exactly one of: spend, income, transfer_internal, card_payment, transfer_person,
     investment, interest, fee, cash, refund, fx.
     THE KEY RULE: money moved between my own accounts, and credit-card bill payments, must NEVER
     count as spending. Only itemized card charges + real debits (mortgage/rent, utilities,
     insurance) are "spend".
   • category (Groceries, Dining, Gas, Travel, Shopping, Housing, Utilities, Insurance,
     Subscriptions, Transport, Health, Charity, …) + a cleaned merchant name (strip processor
     prefixes, store numbers, cities, FX strings). Use a subagent for the long tail of merchants.
   • recurring/subscription detection: a curated catalog of known services (Netflix, Spotify, Apple,
     insurers, telecom, utilities, gym, mortgage…) PLUS a recurrence algorithm (same merchant,
     regular cadence, stable amount). For each subscription compute a robust ~$/month run-rate
     (trailing 12 months, trimming one-off lumps), active/lapsed status, and a monthly timeline.

3) AGGREGATE into the dashboard's data object: monthly / quarterly / yearly rollups; spend-by-category
   over time; income by source; the subscriptions catalog; per-counterparty in/out series; daily
   spend; money-in/out totals; the flow-type summary; and audit-risk metrics (large unexplained
   inbound transfers, income-vs-outflow gap, cash withdrawals, foreign transactions, round-number
   transfers). If scaffolding, emit EXACTLY the `window.FIN` shape in DATA_CONTRACT.md.

4) RENDER: scaffold → swap your data into the copied dashboard. From scratch → inline Apache ECharts
   (vendored locally) + your data into ONE .html that opens offline by double-click.

━━ AUDIT ━━
Include a defensive "if I got audited, what would be flagged?" read: an overall risk score + severity-
ranked flags, each with finding, evidence (the real numbers), exposure, and what to do. Mark it
informational, not tax advice.

━━ PRIVACY ━━
Everything runs locally. Do NOT send my financial data to any external API or service.

━━ WHEN DONE ━━
Open it in a browser, check the console for errors, and screenshot it for me.

━━ FULL SPEC (only if building the renderer from scratch — skip if reusing public/dashboard.html) ━━
One page; every section full-width and collapsible; a dark/light toggle that re-themes the charts;
fully mobile-responsive (stacked panels, 2-col KPIs, scrollable short-label legends, heatmap labels
hidden on narrow screens).
  00 Overview — KPI tiles: money in, identified income, total spending, money out, active subs,
     ~$/mo digital subs, ~$/mo all recurring, span.
  01 Cash flow — diverging monthly chart (income up, spending down, net line); month/quarter/year
     toggle; zoom slider.
  02 Income — monthly stacked by source + an all-time donut. (Person-to-person transfers are tracked
     separately, NOT income.)
  03 Spending — monthly stacked bars by category, a full-width nested treemap (category → merchant,
     click to drill, breadcrumb back), and a category × period heatmap.
  04 Patterns — a GitHub-style daily-spend calendar heatmap (one year + a year selector), a month ×
     year heatmap, quarterly grouped bars.
  05 Subscriptions — group cards, a "how recurring cost has grown" stacked chart, and a sortable/
     filterable table with per-service sparklines + active/lapsed badges.
  06 Transfers — a Sankey (sources → accounts → uses), a top-counterparty diverging bar chart, and a
     click-to-chart per-person in/out timeline + a ledger table.
  07 Accounts — per-account summary cards (in / out / net / # records / date range).
  08 Explorer — a searchable, filterable, paginated table over every transaction (capped height,
     inner scroll, sticky headers).
  09 Audit — the risk assessment above.
  10 Report — a short written read of the key findings, from the data.
DESIGN — quant-terminal, not generic: IBM Plex Mono (headings/labels/figures) + IBM Plex Sans (body),
tabular nums; zero rounded corners; 1px hairline borders; grid-separated tiles; a subtle graph-paper
background; a neon-green accent on near-black for dark mode + a clean white "quant report" light mode;
ECharts with crosshair pointers, dashed gridlines, zoom sliders, scrollable legends; charts must
re-theme correctly when the toggle flips.