1实现原理 · 为什么它能做到
这是纯 prompt 编排型 skill:无任何脚本/可执行代码,靠「Lead Agent + 并行 Subagent」工作流 + 14 份 markdown 知识文件驱动宿主 LLM 产出研究报告
SKILL.md 23.2KB + references/(completeness_review_checklist.md, counter_review_team_guide.md, enterprise_analysis_frameworks.md, enterprise_quality_checklist.md, enterprise_research_methodology.md, formatting_rules.md, quality_gates.md, report_template_v6.md, research_notes_format.md, research_plan_checklist.md, research_report_template.md, source_accessibility_policy.md, source_quality_rubric.md, subagent_prompt.md)——无 scripts/ 目录、无任何脚本文件
上下文隔离是核心机制:子代理各自搜索抓取,只把蒸馏后的笔记写进 research-notes/task-{id}.md,lead agent 只读笔记不读原始搜索结果
Dispatch ──→ Subagent A ──→ writes task-a.md ──┐ ... research-notes/ <──┘ | Context efficiency: Subagents' raw search results stay in their context and are discarded. Lead agent sees only distilled notes (~60-70% context reduction).
笔记文件是唯一溯源通道,每条事实必须能回追到笔记行,构成反幻觉的第一道锁
The research notes are the ONLY communication channel between subagents and the lead agent. Every fact in the final report must be traceable to a line in these notes. No exceptions.
P3 建全局引用注册表并套用定量质量门,不达标的源被显式 Drop 且永不复用
Standard: ≥12 approved sources, ≥5 unique domains, ≥30% official | Lightweight: ≥6 approved sources, ≥3 unique domains, ≥20% official | Max single-source share: ≤25% (Standard), ≤30% (Lightweight) ... **Critical rule:** These [n] are FINAL. P5 may only cite from Approved list. Dropped sources never reappear.
V6.1 的'来源可及性'治理区分 public / semi-public / exclusive-user-provided / private-user-owned,禁止用用户私有数据做自我验证(circular verification),鼓励用户付费源用于第三方研究
When researching the user's own company/assets, if you discover data in user's private accounts (e.g., user's domain registrar showing they own domains), you MUST: 1. Reject it from the registry... 4. Report from external investigator perspective only | When user EXPLICITLY PROVIDES their paid subscriptions or private APIs for third-party research (e.g., "Use my Crunchbase Pro to research competitors"), you SHOULD: 1. Accept it as "exclusive-user-provided" accessibility
P6 强制反方审查:每条主要结论都要做相反解读检查,找不到 ≥3 个问题就重查
1. **Could the conclusion be wrong?** 2. **Which high-impact claims depend on a single source?** 3. **Which claims lack official/academic support?** 4. **Are stale sources used for time-sensitive claims?** 5. **Find ≥3 issues** (re-examine if 0 found)
Enterprise 模式提供六维数据采集 → SWOT/竞争壁垒加权评分/风险矩阵/综合记分卡 → L1/L2/L3 三级质检的完整公司研究流水线
E2: Six-dimension data collection... D1: Company fundamentals ... D6: Internal/proprietary sources ... E3: Structured analysis frameworks: SWOT analysis (evidence-backed, 4 quadrants × 3-5 entries); Competitive barrier quantification (7 dimensions, weighted score); Risk matrix (8 categories, probability × impact); Comprehensive scorecard (6 dimensions, weighted total)
收尾把产出导向同 marketplace 的兄弟 skill:fact-checker 复核、ppt-creator 做幻灯片、pdf-creator 导出
A) Verify facts — run /fact-checker on the report (Recommended) B) Create slides — run /daymade-docs:ppt-creator from the findings C) Export as PDF — run /daymade-docs:pdf-creator for formal delivery
2核心能力
3外部依赖
| 类型 | 依赖 |
|---|---|
| cli | claude(Claude Code CLI,headless -p 派子代理) |
| cli | 宿主 agent 内建工具 web_search / web_fetch(技能本身不实现) |
| package | Cowork / DeerFlow / OpenClaw(可选子代理调度宿主,纯说明性,非依赖) |
| api | 无硬编码 API;流程允许用户在会话中提供 Crunchbase Pro / PitchBook / Wind 等付费数据源给模型用于第三方研究 |
| network | 无固定网络端点;全部 URL 仅出现在示例格式块(nationalgeographic.com / cam.ac.uk / wikipedia.org / quora.com / study.com / blogs.microsoft.com / effectivealtruism.org),非代码调用点 |
| package | 同仓库兄弟 skill 转介:fact-checker、daymade-docs:ppt-creator、daymade-docs:pdf-creator(已核实目录存在) |
4风险提醒 风险提醒:黄色 · 留意使用
- 全部质量门是'模型自觉遵守的清单',无程序强制力 — 无代码执行层;LLM 可跳过反方审查或敷衍填 ≥3 个问题,产出质量最终仍取决于宿主模型纪律。
- 抓取内容回流上下文的 prompt injection 面 — web_fetch 全文进入模型(Deep Read Notes),恶意/被黑网页可注入指令;技能无结构性隔离,仅靠格式规约与反幻觉清单缓解。
- 结构冗余与编号漂移增加误执行概率 — 顶部 P0-P7 图与正文阶段错位一阶(图无 dispatch、多 polish);E1/E3-E7 企业内容在 P1/P2 区重复出现。模型按旧图走会跳步。
- 宣称的 Counter-Review Team 并行审查在多数环境不可达 — 4+1 具名 agent 与 ~/.claude/teams 配置不随包提供(? 项),无该机制时静默退化为手工回退;SKILL 却将其标为 'Recommended'。
- 付费/私有数据使用边界依赖模型判断 — 规则允许用用户提供的 Crunchbase Pro 等查第三方,但'第三方 vs 用户自身'的归属判定与披露义务全凭模型执行,误用边界(把自家数据当研究结论)无硬拦截。
5第二遍独立确认
- [ok] 外部 CLI 依赖 claude -p(--allowedTools web_search,web_fetch,write) — references/subagent_prompt.md:91-99 真实存在,含单任务与并行 for 循环写法
- [ok] web_search/web_fetch 为功能前提 — SKILL.md:110-111 P0 能力表,缺失即 Stop / SCAN-only;另有 references/subagent_prompt.md:23-24 指令
- [ok] Cowork / DeerFlow / OpenClaw 派发途径 — references/subagent_prompt.md:104-115 存在说明与 task() 示例;但仅为宿主选项说明,无实现代码
- [ok] 跨 skill 转介 /fact-checker、/daymade-docs:ppt-creator、/daymade-docs:pdf-creator — SKILL.md:540-542 存在;fact-checker/、daymade-docs/ppt-creator/、daymade-docs/pdf-creator/ 目录在同仓库均已核实存在
- [ok] 无脚本/无可执行代码 — deep-research/ 仅 SKILL.md + .security-scan-passed + references/ 14 个 .md,无 scripts/、无 .py/.sh/.mjs
- [ok] 无凭证读取/无密钥硬编码 — 全目录正则扫 api[_-]?key|Bearer|sk-|password|Authorization|secret 无真实命中(仅 task-a.md 中 'sk-a' 子串误报);.security-scan-passed 亦由 gitleaks 扫描背书
- [ok] 无硬编码网络端点/无 HTTP 调用代码 — 全部 URL 位于示例数据块(research_notes_format.md:32-35、122-131 等),无 curl/wget/requests/axios/fetch
- [ok] SKILL P3 质量门数值与 quality_gates.md Gate2 一致 — 两处均为 ≥12/≥6 源、≥5/≥3 域名、≥30%/≥20% official、单源 ≤25%/≤30%,无矛盾
6结论
cddc8d3acbe77673…63a65c1fcc