expense-filer

Expense Filer Skill

Nobody's judgment is improved by hand-typing receipts. This skill does the clerk work — extraction, categorization, policy screening, currency normalization — and stops at the line that matters: submission happens only after a human reads the report. Numbers must reconcile; anything ambiguous is flagged, never guessed.

What This Skill Produces

• The itemized report — merchant, date, amount, currency, category, project/cost-center, per receipt
• Policy flags — over-limit items, missing-itemization risks, personal-expense suspicion, duplicates, each with the fix
• The reconciliation line — receipts total vs report total, to the cent
• The filing plan — what gets submitted where, gated on approval

Required Inputs

Ask for these if not provided:

• The receipts — images, PDFs, forwarded emails, or a folder the agent can read
• The policy — per-diem limits, category caps, itemization rules (or "use conservative defaults and flag everything near an edge")
• Context — trip/project the expenses attach to, cost center, currency of the report
• The expense system — Concur/Expensify/Ramp/a spreadsheet — and whether draft-only or submit is desired

Framework

1. Extract with provenance — every line links to its source receipt; a number without a receipt is a flag, not a line.
2. Categorize by the policy's taxonomy, not intuition; unknown category → flag with a best-guess clearly marked.
3. Screen before math: duplicates (same merchant+amount±1 day), split-transaction patterns, weekend/personal-suspect items, alcohol-on-meal-receipt rules — flag, never silently drop or include.
4. Normalize currency at the receipt-date rate, shown per line.
5. Reconcile or refuse: if extracted totals and report totals disagree, the report doesn't ship.

Output Format

Expense Report: [trip/project] — [period]

| # | Date | Merchant | Amount | Curr | → [report curr] | Category | Receipt | Flags |

|---|---|---|---|---|---|---|---|---|

Total: [n] items · [amount] · reconciles ✓

Flags needing you: [each with the question to answer]

Quality Checks

• [ ] Every line traces to a receipt; every receipt appears exactly once
• [ ] Totals reconcile to the cent, conversion shown per line
• [ ] Every policy edge is flagged with the specific rule it grazes
• [ ] Nothing ambiguous was guessed — flags ask, lines assert

Anti-Patterns

• [ ] Do not round away discrepancies — a report that's $3 off is wrong, not close
• [ ] Do not categorize creatively to fit under caps — flag the overage; gaming policy is the user's career, not your cleverness
• [ ] Do not drop suspect items silently — surfacing them is the service
• [ ] Do not submit anything — filing is gated below, always

Execution

For agents with file/OCR access and expense-system access (API or UI). Without tools, the report itself is the deliverable. Rules per [SKILLSPEC.md §5](../../SKILLSPEC.md).

Preconditions

• The report above produced, flags resolved by the user, and the final version explicitly approved by a human.
• Expense-system access authenticated; the target report/trip container named.
• The filing plan (create entries, attach receipts, save as draft vs submit) displayed and confirmed — submit requires its own explicit yes.

Allowed actions

• Create the expense entries exactly as in the approved report; attach the corresponding receipt files.
• Save the report as a draft in the system.
• Submit only if the user's approval explicitly included the word-level go for submission.
• Nothing else: no editing existing reports, no touching payment methods, no approving on behalf of anyone.

Verification

• Re-read the created report from the system: line count, per-line amounts, and total match the approved version; every entry shows its attachment.
• Report the system's own total back next to the approved total.

Rollback

• Draft reports: delete the draft. Submitted reports: recall/withdraw if the system allows; otherwise notify the user immediately with the exact state.
• Stop and ask a human if: the system rejects an entry, an attachment fails, or any created total drifts from the approved one.