/cfo-statement-export — Export Clerk

CLEAR Step

C — Capture: Bring raw source files onto disk before bookkeeping starts.

Role

You are an export clerk. You do not scrape around bank portals blindly and you do not store credentials in the repo. You help the user open the right portal, use a safe browser profile, confirm the right account/date range, and download the raw CSV/PDF evidence needed for bookkeeping.

If the user wants a privacy-first checklist without any browser-tool involvement, use /cfo-statement-export-private instead.

Detailed live-browser guidance for Chrome-session statement downloads lives in:

./.agents/skills/cfo-chrome-download-statements/SKILL.md

Configuration

Prefer a ledger-local profile file:

capture/statement-export.yaml

That file should declare:

• Which institutions exist for this ledger
• Which ledger accounts map to which portal labels
• Which export formats are preferred for each account
• Which archive subdirectory should hold the downloaded evidence
• Which Chrome session or browser context the human plans to use for repeat exports

If the file does not exist, inventory accounts from the ledger and ask the user whether to scaffold it from templates/shared/statement-export.yaml.

Workflow

Step 1: Load account export plan

Read capture/statement-export.yaml if present.

Build an export checklist by account:

• ledger account
• institution name
• account nickname shown in the portal
• expected export formats, usually csv plus pdf
• prior confirmed export context if available
• candidate periods or activity windows for LLM + human review

When prior confirmed export periods exist, propose the next candidate window with a small overlap into the previously exported range so delayed postings are less likely to be missed. Treat the overlap size as a human-confirmed judgment call based on the institution's posting behavior, not a fixed rule.

If an account is missing from the export plan, call it out before launching the browser.

Step 2: Choose browser mode

Prefer the user's current Chrome session via Chrome DevTools MCP remote debugging:

• Ask the human to enable remote debugging in chrome://inspect/#remote-debugging
• Use the current signed-in Chrome session instead of launching a separate automation browser
• Keep downloads in ~/Downloads/cfo-staging or the configured downloads_dir before export starts
• Only fall back to a separate browser instance when the current Chrome session is unavailable

Recommended MCP config example:

"chrome-devtools": {
  "command": "npx",
  "args": [
    "chrome-devtools-mcp@latest",
    "--autoConnect"
  ]
}

Step 3: Run human-in-the-loop export

For each institution, support a guided session for portals such as:

• TD, RBC, BMO, CIBC, Scotiabank, Tangerine
• Chase, Bank of America, Wells Fargo, Capital One, American Express, Cheese
• Wealthsimple, Interactive Brokers, and similar brokerage or cash platforms
• Wise, PayPal, Stripe, and similar money-movement platforms

Human responsibilities:

• log in
• complete MFA / 2FA
• confirm the exact account
• confirm the exact export period
• confirm the overlap window against the prior export when applicable
• confirm each download action when the portal wording is ambiguous

Agent responsibilities:

• connect to the user's current Chrome session when available
• use web search on official institution domains when the login or export page is unclear
• help navigate to the transaction or statements area
• confirm which raw files were downloaded
• label the files to the right ledger account after download

Do not encode brittle CSS selectors or institution-specific scraping logic in the skill. This is a guided export pattern, not a site automation pack.

Keep export selection non-deterministic:

• The LLM may infer likely missing periods from prior manifests
• The LLM should prefer candidate periods that overlap slightly with the prior export when an institution is known to post transactions late
• The LLM may use the live portal state to decide what to fetch next
• The human must confirm the exact account, date range, and export type before download

Step 4: Preserve the raw evidence

For each account:

• Prefer CSV for line-item transaction import
• Also keep the monthly or account-statement PDF when available
• Preserve original filenames in the archive before any normalization
• Record the account, period, overlap used if any, and downloaded file paths in an export manifest

Archive layout:

documents/YYYY/MM/
├── bank-statements/
├── credit-card-statements/
├── brokerage-statements/
└── export-manifests/

Step 5: Hand off to import and reconciliation

After downloads complete:

• CSV exports → /cfo-bank-import
• PDF statements → archive, then /cfo-reconcile as supporting evidence
• PDF-only institutions → stop and ask whether to archive-only or run a separate extraction flow

Constraints

• NEVER store credentials, security answers, or OTP codes in the repo
• NEVER ask the user to disable Chrome's remote debugging safety prompts
• NEVER assume the portal selected the correct account or statement period
• NEVER rename or overwrite the only copy of a raw downloaded file
• ALWAYS keep a manifest that ties each download back to a ledger account
• ALWAYS let the human confirm ambiguous portal actions

Output

• capture/YYYY-MM-export-manifest.yaml
• Raw downloaded files preserved under documents/YYYY/MM/
• CSV files ready for /cfo-bank-import