Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Use as the reference manual for the corporate travel and expense policy covering per-diem limits, approval thresholds, and reimbursable categories. Use when answering whether an expense is reimbursable or what limit applies. Do NOT use to submit, approve, or modify expense reports.
.claude/skills/bonaniibm-expense-policy/SKILL.md| Model | Eval pass | Runs |
|---|---|---|
| gemini-3.6-flashlowest | 88% | 17 |
| gemini-3.1-pro-preview | 100% | 1 |
| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-02 | ✗→✓ | ▲ Improved | -76% | 0% |
| case-03 | ✗→✓ | ▲ Improved | -66% | 0% |
| case-09 | ✗→✓ | ▲ Improved | -73% | 0% |
| case-01 | ✗→✓ | ▲ Improved | -57% | 0% |
| case-04 | ✗→✓ | ▲ Improved | -45% | 0% |
This skill documents the reimbursement rules the model cannot infer. It answers eligibility and limit questions; it does not take actions.
| Category | Limit | Approval above limit | |---|---|---| | Meals (per diem) | 60 / day | Manager | | Hotel (per night) | 220 | Manager | | Airfare | Economy only | Director for upgrades |
Identify the category, apply the limit, and state whether manager or director approval is required above it.
"Is a 240 hotel night reimbursable?" — Hotel limit is 220, so the night is reimbursable up to 220 and the overage requires manager approval.
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-02 | fail→pass | 9,891 | 1,579 | -84% | 1 | 1 | 0% | 1,688 | 411 | -76% | 0 | 0 | — |
case-03 | fail→pass | 7,475 | 1,756 | -77% | 1 | 1 | 0% | 1,225 | 422 | -66% | 0 | 0 | — |
case-09 | fail→pass | 10,375 | 1,954 | -81% | 1 | 1 | 0% | 1,801 | 486 | -73% | 0 | 0 | — |
case-01 | fail→pass | 6,696 | 1,729 | -74% | 1 | 1 | 0% | 1,040 | 447 | -57% | 0 | 0 | — |
case-04 | fail→pass | 7,374 | 2,616 | -65% | 1 | 1 | 0% | 1,203 | 663 | -45% | 0 | 0 | — |
case-05 | fail→pass | 7,835 | 2,101 | -73% | 1 | 1 | 0% | 1,361 | 480 | -65% | 0 | 0 | — |
case-06 | fail→pass | 8,848 | 1,718 | -81% | 1 | 1 | 0% | 1,463 | 456 | -69% | 0 | 0 | — |
case-07 | fail→pass | 8,683 | 1,927 | -78% | 1 | 1 | 0% | 1,390 | 493 | -65% | 0 | 0 | — |
case-08 | fail→pass | 7,956 | 2,243 | -72% | 1 | 1 | 0% | 1,301 | 527 | -59% | 0 | 0 | — |
case-10 | fail→pass | 8,693 | 1,708 | -80% | 1 | 1 | 0% | 1,481 | 444 | -70% | 0 | 0 | — |
case-11 | fail→pass | 9,920 | 2,051 | -79% | 1 | 1 | 0% | 1,577 | 439 | -72% | 0 | 0 | — |
case-12 | pass→pass | 9,574 | 1,731 | -82% | 1 | 1 | 0% | 1,493 | 422 | -72% | 0 | 0 | — |
case-13 | fail→pass | 9,023 | 2,372 | -74% | 1 | 1 | 0% | 1,436 | 477 | -67% | 0 | 0 | — |
case-14 | fail→pass | 9,039 | 2,240 | -75% | 1 | 1 | 0% | 1,499 | 512 | -66% | 0 | 0 | — |
case-15 | pass→pass | 12,896 | 1,996 | -85% | 1 | 1 | 0% | 2,241 | 466 | -79% | 0 | 0 | — |
case-16 | fail→pass | 10,001 | 2,362 | -76% | 1 | 1 | 0% | 1,588 | 580 | -63% | 0 | 0 | — |
case-17 | fail→pass | 8,547 | 2,098 | -75% | 1 | 1 | 0% | 1,389 | 460 | -67% | 0 | 0 | — |
case-18 | fail→pass | 9,573 | 1,231 | -87% | 1 | 1 | 0% | 1,582 | 333 | -79% | 0 | 0 | — |
case-19 | fail→pass | 8,347 | 1,314 | -84% | 1 | 1 | 0% | 1,416 | 336 | -76% | 0 | 0 | — |
case-20 | fail→pass | 5,029 | 2,397 | -52% | 1 | 1 | 0% | 873 | 636 | -27% | 0 | 0 | — |
case-21 | pass→pass | 9,835 | 2,146 | -78% | 1 | 1 | 0% | 1,573 | 502 | -68% | 0 | 0 | — |
case-22 | fail→pass | 6,019 | 1,554 | -74% | 1 | 1 | 0% | 955 | 375 | -61% | 0 | 0 | — |
case-23 | fail→pass | 4,277 | 1,594 | -63% | 1 | 1 | 0% | 710 | 456 | -36% | 0 | 0 | — |
case-24 | pass→pass | 4,752 | 1,560 | -67% | 1 | 1 | 0% | 647 | 392 | -39% | 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. 24 cases were attempted. The headline lift of +83 percentage points is the difference between those two pass rates over the 24 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.