Install any skill in seconds. Free to start, no credit card required.
Get Started Free →当用户需要在银行对公金融场景下,对企业授信申请做贷前初筛、材料完整性检查、准入红旗识别、补件清单整理、访谈问题设计或初筛意见输出时使用本技能。适合输出结构化初筛结论、风险提示、待核验事项和下一步推进建议。
.claude/skills/aifinlab-bank-t121-corporate-finance-creditpre-screen-assistant/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-04 | ✗→✓ | ▲ Improved | 65% | 0% |
| case-10 | ✗→✓ | ▲ Improved | 112% | 0% |
| case-17 | ✗→✓ | ▲ Improved | 203% | 0% |
| case-18 | ✗→✓ | ▲ Improved | 121% | 0% |
| case-19 | ✓→✗ | ▼ Worse | 126% | 0% |
这个 skill 用于银行对公授信场景下的第一轮判断。重点不是替代正式尽调或审批,而是在信息还不完整、时间又比较紧的时候,先回答四个关键问题:
它更适合服务对公客户经理、风险经理、授信审查支持岗和中后台预审人员。输出要能直接支持后续尽调、客户沟通和内部评审,而不是只写一堆泛泛的风险口号。
建议统一用以下四级口径:
可进入下一环节:当前未发现明显硬性阻断项,资料基本完整,可进入正式尽调或审查。有条件推进:总体可继续推进,但存在资料缺口、局部风险或需要管理层说明的事项。审慎推进:存在多项重要疑点,建议补件、补查、访谈或升级审查后再决定是否推进。不建议直接推进:当前已出现明显红旗、证据冲突或还款逻辑难以成立,不建议继续推进。建议尽量覆盖以下信息,缺失时必须显式写明:
详细字段见 input-schema.md。
标准输出至少应包含以下部分:
如果用户需要结构化结果,可直接输出 JSON 风格对象,结构定义见 output-schema.md。
本 skill 已补配套脚本:
示例调用:
bashpython scripts/run_skill.py --input assets/example-input.json --output result.md --format markdown
如果只想看 JSON 结果:
bashpython scripts/run_skill.py --input assets/example-input.json --format json
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 31,114 | 46,581 | +50% | 1 | 1 | 0% | 4,144 | 7,106 | +71% | 0 | 0 | — |
case-02 | pass→pass | 20,651 | 32,006 | +55% | 1 | 1 | 0% | 3,048 | 7,416 | +143% | 0 | 0 | — |
case-03 | fail→fail | 27,015 | 32,609 | +21% | 1 | 1 | 0% | 3,450 | 7,228 | +110% | 0 | 0 | — |
case-04 | fail→pass | 22,055 | 13,553 | -39% | 1 | 1 | 0% | 2,796 | 4,622 | +65% | 0 | 0 | — |
case-05 | pass→pass | 27,850 | 22,387 | -20% | 1 | 1 | 0% | 3,597 | 5,734 | +59% | 0 | 0 | — |
case-06 | pass→pass | 24,234 | 11,148 | -54% | 1 | 1 | 0% | 1,366 | 4,113 | +201% | 0 | 0 | — |
case-07 | pass→pass | 20,552 | 30,495 | +48% | 1 | 1 | 0% | 2,917 | 7,206 | +147% | 0 | 0 | — |
case-08 | pass→pass | 19,263 | 26,530 | +38% | 1 | 1 | 0% | 2,791 | 6,150 | +120% | 0 | 0 | — |
case-09 | pass→pass | 21,272 | 24,976 | +17% | 1 | 1 | 0% | 3,059 | 6,307 | +106% | 0 | 0 | — |
case-10 | fail→pass | 22,224 | 33,911 | +53% | 1 | 1 | 0% | 3,149 | 6,674 | +112% | 0 | 0 | — |
case-11 | pass→pass | 26,705 | 29,126 | +9% | 1 | 1 | 0% | 3,303 | 6,790 | +106% | 0 | 0 | — |
case-12 | pass→pass | 19,383 | 24,708 | +27% | 1 | 1 | 0% | 2,768 | 6,026 | +118% | 0 | 0 | — |
case-13 | pass→pass | 23,656 | 27,947 | +18% | 1 | 1 | 0% | 2,790 | 6,175 | +121% | 0 | 0 | — |
case-14 | fail→fail | 20,734 | 30,458 | +47% | 1 | 1 | 0% | 2,913 | 7,095 | +144% | 0 | 0 | — |
case-15 | pass→pass | 20,330 | 23,663 | +16% | 1 | 1 | 0% | 3,097 | 6,052 | +95% | 0 | 0 | — |
case-16 | pass→pass | 22,259 | 30,087 | +35% | 1 | 1 | 0% | 3,035 | 7,356 | +142% | 0 | 0 | — |
case-17 | fail→pass | 18,835 | 35,650 | +89% | 1 | 1 | 0% | 2,749 | 8,325 | +203% | 0 | 0 | — |
case-18 | fail→pass | 23,563 | 36,644 | +56% | 1 | 1 | 0% | 3,283 | 7,271 | +121% | 0 | 0 | — |
case-19 | pass→fail | 21,107 | 27,981 | +33% | 1 | 1 | 0% | 3,165 | 7,139 | +126% | 0 | 0 | — |
case-20 | pass→pass | 23,581 | 27,784 | +18% | 1 | 1 | 0% | 2,775 | 6,465 | +133% | 0 | 0 | — |
case-21 | pass→pass | 39,281 | 29,644 | -25% | 1 | 1 | 0% | 3,444 | 6,752 | +96% | 0 | 0 | — |
case-22 | pass→fail | 19,609 | 26,791 | +37% | 1 | 1 | 0% | 2,518 | 6,682 | +165% | 0 | 0 | — |
case-23 | pass→pass | 27,844 | 30,596 | +10% | 1 | 1 | 0% | 3,402 | 6,696 | +97% | 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. The headline lift of +9 percentage points is the difference between those two pass rates over the 23 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.