Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Cowork-style sweep of a folder of receipts, statements, and expense exports -- categorizes every transaction, matches receipts to statement lines, flags policy violations and anomalies, and outputs a clean expense report plus a findings memo.
.claude/skills/onewave-ai-cowork-expense-audit/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-07 | ✗→✓ | ▲ Improved | 144% | 0% |
| case-04 | ✗→✓ | ▲ Improved | 28% | 0% |
| case-08 | ✗→✓ | ▲ Improved | 45% | 0% |
| case-12 | ✗→✓ | ▲ Improved | 26% | 0% |
| case-17 | ✗→✓ | ▲ Improved | 10% | 0% |
Process an expense folder the way a careful controller does: reconcile receipts against statements, categorize consistently, and separate what is provably documented from what is missing. Input is a directory of receipts (PDF, images), card/bank statements, and any expense exports (.csv/.xlsx). Never invent an amount -- every number in the output traces to a source file.
expense-report.xlsx-ready CSV (or .xlsx if the xlsx skill is available) with the full categorized ledger, plus audit-findings.md: totals by category, reconciliation gaps, flagged items each with severity and the specific source file cited, and a missing-documentation list to chase.As a monthly Cowork scheduled task pointed at a receipts inbox folder: process new files, append to the running ledger, and deliver the month's report and findings without re-litigating prior months.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 12,014 | 5,726 | -52% | 1 | 1 | 0% | 1,207 | 899 | -26% | 0 | 0 | — |
case-02 | fail→fail | 4,005 | 3,314 | -17% | 1 | 1 | 0% | 405 | 1,167 | +188% | 0 | 0 | — |
case-07 | fail→pass | 5,675 | 10,150 | +79% | 1 | 1 | 0% | 953 | 2,323 | +144% | 0 | 0 | — |
case-03 | fail→fail | 17,077 | 3,591 | -79% | 1 | 1 | 0% | 1,829 | 1,270 | -31% | 0 | 0 | — |
case-04 | fail→pass | 11,157 | 5,489 | -51% | 1 | 1 | 0% | 1,250 | 1,605 | +28% | 0 | 0 | — |
case-05 | pass→pass | 4,971 | 6,126 | +23% | 1 | 1 | 0% | 755 | 1,530 | +103% | 0 | 0 | — |
case-06 | pass→pass | 11,446 | 10,554 | -8% | 1 | 1 | 0% | 1,581 | 2,045 | +29% | 0 | 0 | — |
case-08 | fail→pass | 5,991 | 6,553 | +9% | 1 | 1 | 0% | 1,108 | 1,606 | +45% | 0 | 0 | — |
case-09 | pass→pass | 7,095 | 5,138 | -28% | 1 | 1 | 0% | 1,098 | 1,526 | +39% | 0 | 0 | — |
case-10 | fail→fail | 11,589 | 9,537 | -18% | 1 | 1 | 0% | 1,828 | 2,217 | +21% | 0 | 0 | — |
case-11 | pass→pass | 9,547 | 6,319 | -34% | 1 | 1 | 0% | 1,501 | 1,623 | +8% | 0 | 0 | — |
case-12 | fail→pass | 11,572 | 8,768 | -24% | 1 | 1 | 0% | 1,610 | 2,029 | +26% | 0 | 0 | — |
case-13 | pass→pass | 11,578 | 6,550 | -43% | 1 | 1 | 0% | 1,819 | 1,692 | -7% | 0 | 0 | — |
case-14 | pass→pass | 17,485 | 20,070 | +15% | 1 | 1 | 0% | 3,087 | 4,443 | +44% | 0 | 0 | — |
case-15 | pass→pass | 6,172 | 6,836 | +11% | 1 | 1 | 0% | 1,122 | 1,856 | +65% | 0 | 0 | — |
case-16 | fail→fail | 7,681 | 8,717 | +13% | 1 | 1 | 0% | 1,210 | 1,971 | +63% | 0 | 0 | — |
case-17 | fail→pass | 13,972 | 10,436 | -25% | 1 | 1 | 0% | 2,063 | 2,266 | +10% | 0 | 0 | — |
case-18 | pass→pass | 3,346 | 3,288 | -2% | 1 | 1 | 0% | 648 | 1,288 | +99% | 0 | 0 | — |
case-19 | pass→pass | 7,519 | 4,016 | -47% | 1 | 1 | 0% | 970 | 1,300 | +34% | 0 | 0 | — |
case-20 | fail→pass | 9,804 | 14,092 | +44% | 1 | 1 | 0% | 1,654 | 2,997 | +81% | 0 | 0 | — |
case-21 | fail→fail | 12,503 | 14,498 | +16% | 1 | 1 | 0% | 2,151 | 3,029 | +41% | 0 | 0 | — |
case-22 | fail→fail | 34,599 | 23,036 | -33% | 1 | 1 | 0% | 4,985 | 3,894 | -22% | 0 | 0 | — |
case-23 | fail→pass | 9,721 | 11,032 | +13% | 1 | 1 | 0% | 1,344 | 2,254 | +68% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 23 cases were attempted, and 22 counted toward the lift figure. The other 1 produced results that are not comparable between the two arms, so they are excluded from the headline rather than averaged into it. The headline lift of +30 percentage points is the difference between those two pass rates over the 22 comparable cases.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.