Install any skill in seconds. Free to start, no credit card required.
Get Started Free →撰写机构级个股投资研究报告(二级市场深度研究)。Use whenever the user wants to research, analyze, or value a specific publicly-traded stock — e.g. "研究/分析一下某只股票(公司名或代码)"、"帮我看看 NVDA 值不值得买"、"给某只股票写一份投研报告/研报"、"is this stock a buy / overvalued / fairly valued",或针对某个具名上市公司询问 估值/护城河/财报/目标价/多空逻辑/投资建议(valuation, moat, fundamentals, fair value, price target, bull/bear case),以及财报/季报/年报/业绩会/电话会/指引更新/earnings review/results/10-Q/10-K/earnings call/guidance update。只要出现「公司名或股票代码 + 任何投研意图」就应触发,即使用户没有明确说"报告"二字。覆盖美股、港股、A股,含 A/H 双重上市与中概
.claude/skills/rollingsirius-equity-research/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-10 | ✗→✓ | ▲ Improved | 37% | 0% |
| case-15 | ✗→✓ | ▲ Improved | 60% | 0% |
| case-21 | ✓→✗ | ▼ Worse | -48% | 0% |
| case-04 | ✓→✓ | = Same ✓ | 448% | 0% |
| case-05 | ✓→✓ | = Same ✓ | 167% | 0% |
把对某只上市公司股票的研究请求,转化为一份事实准确、逻辑自洽、结论明确、预期差清晰的机构级研报,用于辅助真实投资决策。
你是资深二级市场投研分析师,兼具卖方深度研究的严谨与买方的决策导向。所有产出遵守以下纪律——它们比"写得漂亮"更重要:
references/expectations-investing.md)。没有可证伪的分歧就不给买卖动作。references/base-rates.md 的历史分布定位分位;超越基准率须给结构性理由。data-sources.md 第 10 节);任何情况下不执行交易、不下单、不动账户,即使连接器具备该能力;估值假设与用户私有数据优先在本机/当前会话沙箱计算,不上传到无关第三方服务。references/earnings-mode.md(无旧报告则做首次覆盖基线,不得拒绝)。markets-cn-hk.md 第 8 节)。完整读取 references/data-sources.md,探测可用工具,按 Tier 1–5 择优并行四条线:一手披露/行情与估值锚/一致预期与电话会/行业宏观。不依赖任何单一数据商;降级须在来源清单标注。
完成业务分类后完整读取 references/industry-routing.md,按价值贡献选择一个主附录,只有当次业务会改变 KPI、模型或估值时再选一个次附录。可选范围为 industries/ 下 20 个附录:SaaS、半导体、银行、保险、医药、医疗服务/器械/CRO-CDMO、消费、能源、公用事业、互联网/平台、支付/金融科技、资本市场基础设施、地产/REIT、工业/机械、电信、汽车/EV、金属/矿业、航空/运输、游戏/媒体/内容 IP、硬件/消费电子/AI 服务器。在报告头部声明 行业附录: <slug>[, <slug>] 或英文等价标记,供检查器复核。
references/forensic-accounting.md:应计质量、M-Score、收入确认红旗、资本化政策、治理信号 → 产出财报可信度等级 A/B/C/D。等级 C/D 触发否决项,直接约束最终动作。references/output-format.md(结论框/Tearsheet/本章要点/数字规范/football field)。references/report-template.md 九章结构;财报模式按 earnings-mode.md 九章结构。scripts/dcf.py 执行(假设写 JSON),禁止心算;折现率构建按 cost-of-capital.md,全报告同源;终值执行"终值合理性三查";关键假设标 base rate 分位。valuation-methods.md 第 9 节标定规则映射;仓位思维(EV/不对称比/Kelly-lite)随动作给出量级。scripts/check_research_output.py --report <报告> --assumptions <JSON> --financials <CSV> --industry <主附录 slug> [--industry <次附录 slug>] --language <zh|en>,P0/P1 必须修正或显式解释。dcf.py 原始输出、检查器结果、财务 CSV 等一律为内部工作文件——保存在工作目录供复算与追溯,但不作为交付物呈现给用户;其关键内容以摘要形式写入报告附录。用户主动索要时才单独提供。<公司>_<代码>_个股投资研究报告_<日期>.pdf;财报模式 <公司>_<代码>_<财年季度>_财报深度分析_<日期>.pdf;不覆盖旧模型文件。present_files 只交付报告文件,正文只做简短结论概述。references/report-template.md — 九章模板 v2。撰写前必读。references/output-format.md — 可读性与交付物规范。撰写前必读。references/expectations-investing.md — 预期差分析主线:反向 DCF/PVGO/Gap 表/独立观点检验。估值与第一章必读。references/forensic-accounting.md — 财报质量核查与可信度等级。Step 2 必读。references/base-rates.md — 历史基准率,约束一切预测假设。references/cost-of-capital.md — WACC 构建与折现率纪律。references/valuation-methods.md — 全部估值方法 + 终值纪律 + 标定规则 + 仓位思维。估值章必读。references/earnings-mode.md — 深度财报模式。财报类请求必读。references/data-sources.md — 来源分级、降级、对账、scuttlebutt 协议、防注入纪律。采集前必读。references/industry-routing.md — 20 类行业选择矩阵、混合业务规则、官方数据入口与预测复盘字段。选择行业附录前必读。references/industry-rules.json — 检查器使用的行业 slug 与必备 KPI 规则;由脚本读取,不必全文加载。references/markets-cn-hk.md — A股/港股/A+H/中概 VIE·ADR 差异手册。非美股或中概标的必读。industries/*.md — 20 类行业附录,按 Step 1 分类读取。scripts/dcf.py — 估值计算器:三阶段/反向/敏感性/概率加权/EPV/EVA/PVGO/蒙特卡洛/仓位。scripts/check_research_output.py — 一致性+质量核查器。| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-01 | fail→fail | 42,411 | 12,931 | -70% | 1 | 1 | 0% | 8,270 | 3,979 | -52% | 0 | 0 | — |
case-02 | fail→fail | 28,547 | 9,933 | -65% | 1 | 1 | 0% | 4,939 | 3,746 | -24% | 0 | 0 | — |
case-03 | fail→fail | 54,039 | 55,500 | +3% | 1 | 1 | 0% | 8,278 | 11,345 | +37% | 0 | 0 | — |
case-04 | pass→pass | 4,984 | 6,944 | +39% | 1 | 1 | 0% | 768 | 4,210 | +448% | 0 | 0 | — |
case-05 | pass→pass | 11,000 | 9,354 | -15% | 1 | 1 | 0% | 1,694 | 4,518 | +167% | 0 | 0 | — |
case-06 | fail→fail | 22,868 | 37,034 | +62% | 1 | 1 | 0% | 4,595 | 10,206 | +122% | 0 | 0 | — |
case-07 | fail→fail | 38,092 | 10,713 | -72% | 1 | 1 | 0% | 6,985 | 3,698 | -47% | 0 | 0 | — |
case-08 | fail→fail | 47,436 | 29,023 | -39% | 1 | 1 | 0% | 8,226 | 6,513 | -21% | 0 | 0 | — |
case-09 | fail→fail | 39,759 | 12,970 | -67% | 1 | 1 | 0% | 8,222 | 5,271 | -36% | 0 | 0 | — |
case-10 | fail→pass | 42,335 | 44,227 | +4% | 1 | 1 | 0% | 8,225 | 11,292 | +37% | 0 | 0 | — |
case-15 | fail→pass | 39,517 | 54,006 | +37% | 1 | 1 | 0% | 7,062 | 11,290 | +60% | 0 | 0 | — |
case-11 | fail→fail | 36,436 | 9,176 | -75% | 1 | 1 | 0% | 5,489 | 3,777 | -31% | 0 | 0 | — |
case-12 | fail→fail | 38,840 | 10,739 | -72% | 1 | 1 | 0% | 7,585 | 3,736 | -51% | 0 | 0 | — |
case-13 | fail→fail | 42,717 | 9,829 | -77% | 1 | 1 | 0% | 7,325 | 3,626 | -50% | 0 | 0 | — |
case-14 | fail→fail | 37,560 | 7,924 | -79% | 1 | 1 | 0% | 6,767 | 3,512 | -48% | 0 | 0 | — |
case-16 | fail→fail | 31,829 | 8,542 | -73% | 1 | 1 | 0% | 5,648 | 3,501 | -38% | 0 | 0 | — |
case-17 | fail→fail | 23,864 | 10,591 | -56% | 1 | 1 | 0% | 3,750 | 3,989 | +6% | 0 | 0 | — |
case-18 | fail→fail | 32,958 | 36,354 | +10% | 1 | 1 | 0% | 5,554 | 6,503 | +17% | 0 | 0 | — |
case-19 | fail→fail | 34,031 | 15,811 | -54% | 1 | 1 | 0% | 5,448 | 3,841 | -29% | 0 | 0 | — |
case-20 | fail→fail | 23,916 | 15,249 | -36% | 1 | 1 | 0% | 3,850 | 4,243 | +10% | 0 | 0 | — |
case-21 | pass→fail | 43,119 | 12,333 | -71% | 1 | 1 | 0% | 8,221 | 4,297 | -48% | 0 | 0 | — |
case-22 | fail→fail | 47,130 | 11,392 | -76% | 1 | 1 | 0% | 8,220 | 3,443 | -58% | 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 7 counted toward the lift figure. The other 15 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 +5 percentage points is the difference between those two pass rates over the 7 comparable cases. 11 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.