name: business-consult description: "商业调研证据链：用于行业进入、商业机会、客户问题诊断、市场研究、反证检验和咨询报告；先界定问题，再以来源、诚实标签、限额证伪和同线脚注交付；不适用于法律、财务或投资尽调。Use for auditable business diagnosis, market research, falsification, and consulting reports. 思远AiLab 出品。" metadata: status: active（2026-07-31 三轮定性验收闭环转正） validation_type: qualitative validation_loop_max: 3 author: Siyuan (AI·LAB) release: flagship license: CC-BY-4.0 slug: business-consult version: 1.0.0 displayName: "Business Consult｜可追溯商业调研" summary: "先把商业问题问清，再做可追溯调研、限额证伪和金字塔报告；每条数据结论都能回到来源。" tags: [business-research, market-research, falsification, consulting-report, chinese] homepage: "https://github.com/SiyuanAiLab/business-consult"

Business Consult｜可追溯商业调研

让 AI 做商业调研，每个结论都敢标明身份：这是数据、这是推算、这是没验证的——编不出来的，它写「未知」。

你问「要不要做宠物经济」，多数 AI 直接给你一份行业分析。但你是想开店、做供应链，还是做内容号？问题没定边界，报告越漂亮错得越远。本 Skill 的行为规则相反：先界定问题，再调研；没过完界定门，不开工。

它交付的不是一份报告，是一条可以抽查的证据链：

[Data] 中国养老服务体系基础格局"9073"最早见于 2007 年《上海民政事业发展十一五规划》，国家层面文件从未正式写入该比例。
——来源 S108｜置信度 medium｜aging-economy 示例声明册第 S1-C001 条

30 秒入门

装好以后，直接对它说人话：

1. 「我要不要做宠物殡葬？先别急着调研，帮我把问题界定清楚。」
2. 「调研一下中国银发经济，我要判断是否进入。」
3. 「我这个创业想法，帮我专门找找反证，别只找支持证据。」
4. 「把前面的调研写成一份能拿去汇报的报告。」
5. 「接着上次的调研继续。」（进度自动续跑，不重头再来）

功能 × 触发 × 输出

证据：不看广告，看实物

❌ 普通 AI 调研：「中国银发经济市场规模约 7 万亿。」——数字哪来的？它不说，你也不知道该不该信。

✅ 本 Skill 的同一个结论（真实示例摘录）：

预注册冻结于 2026-08-02，3 条对抗查询全部执行完毕，可信反证为零；裁决说明逐项核对到「国家发改委官网专家谈原句与声明数字一致（约 7 万亿、占 GDP 约 6%）」；第二轮发现的 8-10 万亿口径如实记录为「未构成反证，但未获官网原文核验」。

成品+原料双给：GitHub 仓库 examples/aging-economy/ 是一次完整的中国银发经济真实调研——原料（起点诊断书、研究简报）和成品（107 条登记声明、57 个分级来源、7 条承重声明的逐条裁决、金字塔终稿）全部公开，安装前就可以逐条抽查。

装之前先拿去用：一份 AI 调研能不能信的快速检查表

不管装不装本 Skill，下次任何 AI 给你商业调研，先用这四条量它：

• [ ] 每个结论有没有标身份（数据/推算/假设），还是全都长得一样？
• [ ] 来源能不能当场点开核对，还是只有一句「据统计」？
• [ ] 它找过反证吗，还是只找了支持证据？
• [ ] 它敢说「未决」「未知」吗，还是永远有答案？

四条全过的调研才敢拿去汇报。本 Skill 的全部机制，就是让这四条从「靠自觉」变成「被强制」。

边界与它不适合谁

能力边界：

• 擅长：行业进入判断、商业机会评估、竞争格局、用户痛点、基于公开资料的可辩护研究
• 需要你供给：你的经营模型、决策原则、约束条件（不提供也能跑，它会如实标注 [Assumption] No business context supplied）
• 超出范围：法律、财务、投资尽调；实时数据保证（公开来源覆盖度随时间变化）

它不适合谁：

• 只要一份漂亮报告、不想被追问「你的问题到底是什么」的人——它的第一道工序就是追问你
• 需要尽调级结论做投资决策的人——它是公开资料研究，不是尽职调查
• 想要 AI 拍胸脯说「这个想法能成」的人——它的裁决最高只到「预算内未推翻」，永远不说「是真的」

当前版本说明： 中文平台数据采集组件（zh-data）在路线图上但未建成，本包不创建空壳；CC BY 4.0，署名即可商用；单人 best-effort 维护，版本号外露，不承诺固定发布日期。

证明链导航

本 Skill 是思远AiLab工作方式的一个可交互样例：判断归人、证据归链、边界外露。完整源码与 aging-economy 示例包见 GitHub 仓库（frontmatter homepage 字段）；同系列 Skill 陆续上架中。

——思远AiLab 出品

Route only. Do not perform a child's analysis inside this Skill. This SkillHub package has one public entry and four bundled child components. When dispatching, read the named children/<child>/SKILL.md and resolve that child's scripts and references from its own directory.

Phase 0: Resolve the project

1. Choose one explicit project root. Keep all child outputs inside it.
2. Read references/contract.md.
3. Inspect {project-root}/PROGRESS.md.
4. Resume only when its skill field matches the selected child.
5. If it belongs to another active child, do not overwrite it. Finish or hand off through the progress script.
6. If no progress file exists, initialize it:

python3 scripts/progress.py init \
  --project-root "/absolute/project-root" \
  --project "project-slug" \
  --skill "diagnose" \
  --mode "live"

Route

For a pending route, state that the route is reserved but unavailable in this build. Do not simulate an undeveloped child.

Dispatch

For diagnose:

1. Initialize or resume PROGRESS.md only with owner diagnose.
2. Read and follow children/diagnose/SKILL.md; it asks exactly one question, persists the session, and stops for the reply.
3. Keep ownership with diagnose until the public-shell validator accepts completion:

python3 children/diagnose/scripts/validate_diagnose.py completed "/absolute/project-root"
python3 children/diagnose/scripts/validate_diagnose.py research-probe "/absolute/project-root"

1. Hand off only when the probe emits handoff: accepted and skip_duplicate_intake: true.

For business-consult-research:

1. Resume only when PROGRESS.md is owned by research.
2. Probe {project-root}/diagnosis.md and {project-root}/diagnose/session.json as one pair.
3. If either exists, run the public diagnose research-probe; stop on a missing pair, in-progress state, invalid state, or any payload other than handoff: accepted.
4. Pass only the accepted payload, then read and follow children/business-consult-research/SKILL.md and skip duplicate intake fields.
5. If neither diagnose output exists, read and follow children/business-consult-research/SKILL.md in standalone mode.

For falsify:

1. Resume only when PROGRESS.md is owned by falsify.
2. Run python3 children/falsify/scripts/validate_falsify.py probe "/absolute/project-root".
3. Read and follow children/falsify/SKILL.md only after the fixed research input set passes.
4. Run the validate command before handing off to report.

For report:

1. Resume only when PROGRESS.md is owned by report.
2. Run python3 children/report/scripts/validate_report.py probe "/absolute/project-root" without fixture overrides.
3. Read and follow children/report/SKILL.md only after both verdict inputs and every required research input pass.

Progress discipline

• Run progress.py resume before every child dispatch.
• Mark a stage complete immediately after its required output passes validation.
• Do not mark completion based on prose alone.
• Hand off only after the current skill is completed.
• Preserve notes and timestamps.

python3 scripts/progress.py resume --project-root "/absolute/project-root" --skill research
python3 scripts/progress.py mark --project-root "/absolute/project-root" --skill research --stage s1
python3 scripts/progress.py handoff --project-root "/absolute/project-root" --from-skill research --to-skill falsify

Stop conditions

• Stop on a skill ownership mismatch.
• Stop when a required output exists but fails its schema or claim-label check.
• Stop before any route marked pending.
• Never write outside the selected project root during a run.

这个 Skill 质量相当不错，做得最出色的地方是：每个结论都强制标注来源和可信度等级，不会瞎编数据告诉你「一定能成」，找不到反证也会老老实实说「未决」。调研流程有四个严格步骤相互制约，重要结论还要专门「证伪」检验，比普通 AI 调研可信得多。不足之处是中文平台的数据采集组件还没做好，对中文市场调研有一定影响；另外调研流程较严谨完整，速度可能偏慢，不太适合紧急决策场景。适合需要可靠、可解释的商业判断依据的用户。