Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Build, rebuild, audit, and compare QQ mailbox invoice ground truth datasets for this repository using the existing truth-building scripts and evidence artifacts. Use when the task is to generate a QQ mailbox ground truth set, validate QQ batch output against truth, review QQ mailbox invoice evidence, or investigate QQ-specific invoice extraction pitfalls in this project.
.claude/skills/ethanyoq-qq-email-ground-truth/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-04 | ✗→✓ | ▲ Improved | -26% | 0% |
| case-07 | ✗→✓ | ▲ Improved | -14% | 0% |
| case-08 | ✗→✓ | ▲ Improved | -17% | 0% |
| case-09 | ✗→✓ | ▲ Improved | -50% | 0% |
| case-10 | ✗→✓ | ▲ Improved | -49% | 0% |
Use this project-level skill for QQ mailbox ground truth work in this repository.
Read references/workflow.md and references/pitfalls.md first. Read references/case-study-qq-20260201-20260311.md and references/project-artifacts.md when you need a validated example, current artifact paths, or prior evidence.
build_truth_dataset.py and audit_email_truth.py. Do not create a parallel truth-building flow unless the user explicitly asks for one.references/workflow.md for the build and validation sequence.references/pitfalls.md as the default debug checklist when counts or fields look wrong.references/case-study-qq-20260201-20260311.md only as a worked example and evidence sample.references/project-artifacts.md to find the current canonical manifests, reports, and diagnostics.truth_manifest.json for machine comparison.ground_truth_report.md for human review.pending_review_count = 0 before calling a dataset final.| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-06 | pass→pass | 8,536 | 3,473 | -59% | 1 | 1 | 0% | 1,333 | 1,050 | -21% | 0 | 0 | — |
case-01 | fail→fail | 13,575 | 4,184 | -69% | 1 | 1 | 0% | 2,833 | 669 | -76% | 0 | 0 | — |
case-02 | fail→fail | 26,010 | 4,296 | -83% | 1 | 1 | 0% | 6,194 | 674 | -89% | 0 | 0 | — |
case-03 | fail→fail | 13,863 | 4,270 | -69% | 1 | 1 | 0% | 2,820 | 691 | -75% | 0 | 0 | — |
case-04 | fail→pass | 17,921 | 10,876 | -39% | 1 | 1 | 0% | 3,002 | 2,210 | -26% | 0 | 0 | — |
case-05 | pass→pass | 10,961 | 5,340 | -51% | 1 | 1 | 0% | 1,797 | 1,379 | -23% | 0 | 0 | — |
case-07 | fail→pass | 11,245 | 6,566 | -42% | 1 | 1 | 0% | 1,813 | 1,565 | -14% | 0 | 0 | — |
case-08 | fail→pass | 14,850 | 8,938 | -40% | 1 | 1 | 0% | 2,358 | 1,961 | -17% | 0 | 0 | — |
case-09 | fail→pass | 12,311 | 3,756 | -69% | 1 | 1 | 0% | 2,098 | 1,050 | -50% | 0 | 0 | — |
case-10 | fail→pass | 9,438 | 2,396 | -75% | 1 | 1 | 0% | 1,583 | 802 | -49% | 0 | 0 | — |
case-11 | fail→pass | 7,410 | 1,909 | -74% | 1 | 1 | 0% | 1,206 | 667 | -45% | 0 | 0 | — |
case-12 | pass→pass | 8,179 | 3,839 | -53% | 1 | 1 | 0% | 1,440 | 1,208 | -16% | 0 | 0 | — |
case-13 | pass→pass | 6,665 | 3,981 | -40% | 1 | 1 | 0% | 1,216 | 1,202 | -1% | 0 | 0 | — |
case-14 | pass→pass | 11,962 | 4,998 | -58% | 1 | 1 | 0% | 1,980 | 1,291 | -35% | 0 | 0 | — |
case-15 | fail→pass | 16,712 | 1,731 | -90% | 1 | 1 | 0% | 962 | 726 | -25% | 0 | 0 | — |
case-16 | pass→pass | 9,701 | 3,844 | -60% | 1 | 1 | 0% | 1,544 | 1,125 | -27% | 0 | 0 | — |
case-17 | fail→pass | 7,737 | 2,082 | -73% | 1 | 1 | 0% | 1,358 | 803 | -41% | 0 | 0 | — |
case-18 | fail→pass | 6,451 | 2,255 | -65% | 1 | 1 | 0% | 1,110 | 826 | -26% | 0 | 0 | — |
case-19 | pass→pass | 8,311 | 4,322 | -48% | 1 | 1 | 0% | 1,310 | 1,260 | -4% | 0 | 0 | — |
case-20 | pass→pass | 10,448 | 10,653 | +2% | 1 | 1 | 0% | 2,102 | 2,599 | +24% | 0 | 0 | — |
case-21 | pass→pass | 13,805 | 11,406 | -17% | 1 | 1 | 0% | 2,606 | 2,503 | -4% | 0 | 0 | — |
case-22 | pass→pass | 17,388 | 17,015 | -2% | 1 | 1 | 0% | 3,956 | 4,112 | +4% | 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. 22 cases were attempted, and 18 counted toward the lift figure. The other 4 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 +41 percentage points is the difference between those two pass rates over the 18 comparable cases. 2 cases got worse with the skill loaded, and they are included in that figure.
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.