Install any skill in seconds. Free to start, no credit card required.
Get Started Free →支払調書の画像を読み取り構造化データを返す。 他のスキルから呼び出されるほか、直接ユーザーが呼び出すことも可能。
.claude/skills/kazukinagata-reading-payment-statement/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-04 | ✗→✓ | ▲ Improved | 61% | 0% |
| case-05 | ✗→✓ | ▲ Improved | 57% | 0% |
| case-07 | ✗→✓ | ▲ Improved | 51% | 0% |
| case-08 | ✗→✓ | ▲ Improved | 10% | 0% |
| case-16 | ✗→✓ | ▲ Improved | -29% | 0% |
支払調書(報酬、料金、契約金及び賞金の支払調書)の画像を読み取り、構造化データとして返すスキル。
ファイルが PDF(.pdf)の場合、画像 OCR の前にテキスト抽出を試みる。
shinkoku pdf extract-text --file-path <path> を実行するshinkoku pdf to-image --file-path <path> --output-dir <dir> で PNG に変換し、以下の画像読み取りフローに進む精度を高めるため、同じ画像を2つの独立したコンテキストで並列に読み取り、結果を照合する。
サブエージェントが使える環境では、2つのサブエージェントを並列で起動し、それぞれ独立に画像を読み取る。 各サブエージェントには以下の「基本ルール」と「出力フォーマット」をプロンプトとして渡し、画像ファイルパスを指定する。
サブエージェントが利用できない環境では、以下の手順で読み取る:
⚠ デュアル検証が利用できないため、必ずユーザーに目視確認を依頼してください。
画像を読み取り、以下の形式で返す:
---PAYMENT_STATEMENT_DATA---
payer_name: 支払者名
category: 区分(報酬/料金/契約金/賞金)
gross_amount: 支払金額(int)
withholding_tax: 源泉徴収税額(int)
---END---複数のファイルパスが指示された場合:
## file1.jpg
---PAYMENT_STATEMENT_DATA---
...
---END---
## file2.jpg
---PAYMENT_STATEMENT_DATA---
...
---END---| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 13,207 | 6,275 | -52% | 1 | 1 | 0% | 240 | 1,270 | +429% | 0 | 0 | — |
case-02 | fail→fail | 10,018 | 8,208 | -18% | 1 | 1 | 0% | 1,198 | 1,374 | +15% | 0 | 0 | — |
case-03 | fail→fail | 6,217 | 18,010 | +190% | 1 | 1 | 0% | 437 | 3,927 | +799% | 0 | 0 | — |
case-04 | fail→pass | 5,007 | 2,426 | -52% | 1 | 1 | 0% | 803 | 1,291 | +61% | 0 | 0 | — |
case-05 | fail→pass | 6,419 | 4,388 | -32% | 1 | 1 | 0% | 1,086 | 1,704 | +57% | 0 | 0 | — |
case-06 | pass→pass | 3,958 | 3,106 | -22% | 1 | 1 | 0% | 664 | 1,400 | +111% | 0 | 0 | — |
case-07 | fail→pass | 6,688 | 3,424 | -49% | 1 | 1 | 0% | 953 | 1,442 | +51% | 0 | 0 | — |
case-08 | fail→pass | 7,566 | 2,743 | -64% | 1 | 1 | 0% | 1,233 | 1,360 | +10% | 0 | 0 | — |
case-09 | pass→pass | 5,118 | 4,345 | -15% | 1 | 1 | 0% | 965 | 1,745 | +81% | 0 | 0 | — |
case-10 | pass→pass | 5,204 | 3,427 | -34% | 1 | 1 | 0% | 853 | 1,522 | +78% | 0 | 0 | — |
case-11 | fail→fail | 6,036 | 1,657 | -73% | 1 | 1 | 0% | 895 | 1,154 | +29% | 0 | 0 | — |
case-12 | pass→pass | 3,086 | 3,202 | +4% | 1 | 1 | 0% | 475 | 1,377 | +190% | 0 | 0 | — |
case-13 | pass→pass | 5,167 | 2,267 | -56% | 1 | 1 | 0% | 932 | 1,327 | +42% | 0 | 0 | — |
case-14 | fail→fail | 6,605 | 2,000 | -70% | 1 | 1 | 0% | 976 | 1,247 | +28% | 0 | 0 | — |
case-15 | pass→pass | 8,252 | 3,768 | -54% | 1 | 1 | 0% | 1,286 | 1,498 | +16% | 0 | 0 | — |
case-16 | fail→pass | 9,475 | 2,289 | -76% | 1 | 1 | 0% | 1,744 | 1,235 | -29% | 0 | 0 | — |
case-17 | pass→pass | 7,679 | 3,234 | -58% | 1 | 1 | 0% | 1,384 | 1,436 | +4% | 0 | 0 | — |
case-18 | fail→pass | 7,158 | 2,138 | -70% | 1 | 1 | 0% | 1,046 | 1,286 | +23% | 0 | 0 | — |
case-19 | fail→pass | 6,646 | 2,352 | -65% | 1 | 1 | 0% | 1,146 | 1,352 | +18% | 0 | 0 | — |
case-20 | fail→pass | 14,226 | 3,937 | -72% | 1 | 1 | 0% | 2,247 | 1,567 | -30% | 0 | 0 | — |
case-21 | fail→fail | 15,936 | 8,267 | -48% | 1 | 1 | 0% | 1,962 | 1,433 | -27% | 0 | 0 | — |
case-22 | fail→fail | 14,564 | 15,368 | +6% | 1 | 1 | 0% | 2,383 | 3,586 | +50% | 0 | 0 | — |
case-23 | fail→fail | 5,432 | 18,274 | +236% | 1 | 1 | 0% | 1,042 | 2,410 | +131% | 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 18 counted toward the lift figure. The other 5 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 +35 percentage points is the difference between those two pass rates over the 18 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.