/analyst-research
A Claude Skill that turns a single hypothesis into a publication-ready research report — three scope modes, one rigorous 8-step workflow.一个把单一假设做成可发表研究报告的 Claude Skill——三档 mode，一套严谨的 8 步工作流。
analyst-research · investment research workflow skill Pick a scope based on your project size: light 5-page decision memo, 0 charts, 10-15 min Single LLM. Pure markdown footnote citations. medium 10-15p topic analysis, 3-8 charts, 30-45 min PDF + Word. Sign-off checkpoint after draft. heavy Flagship report 15k+ words, 20-35+ charts, 2-3 hours Multi-LLM optional. PDF + Word + WeChat + HTML. Which mode fits your project? _analyst-research · 投研工作流 skill 按项目体量挑一档 scope： light 5 页决策 memo，不带图，10-15 分钟 单 LLM，纯 markdown 脚注引用。 medium 10-15 页主题分析，3-8 张图，30-45 分钟 产出 PDF + Word，draft 后留一个 sign-off。 heavy 1.5 万字以上旗舰报告，20-35+ 张图，2-3 小时 可选多 LLM 协作，产出 PDF + Word + 公众号 + HTML。 你的项目适合哪一档？ _
Type /analyst-research in Claude Code, answer the scope prompt, and hand over your hypothesis. The skill loads only the workflow for the mode you picked — no wasted context.在 Claude Code 里敲 /analyst-research，回答 scope 提示，把假设交给它就行。Skill 只会加载你选中那一档的工作流，不浪费 context。
- → 10-15 min, single LLM10-15 分钟，单 LLM
- → 0 charts, pure footnote citations不带图，纯脚注引用
- → BLUF executive summary结论先行的执行摘要
- → 0 hard stops, 12-item grep self-check无硬性停点，12 项 grep 自检
Exec brief · decision support · internal memo执行简报 · 决策支持 · 内部 memo
- → 30-45 min, single LLM30-45 分钟，单 LLM
- → 3-8 custom charts (shared chart_template)3-8 张定制图（共用 chart_template）
- → PDF + Word derivations派生 PDF + Word
- → 1 sign-off checkpoint after draftdraft 后留一个 sign-off 节点
Topic deep-dive · board memo with data · quick analysis主题深挖 · 带数据的决策 memo · 快速分析
- → 2-3 hours, multi-LLM optional2-3 小时，可选多 LLM
- → 20-35+ charts, full visual production stack20-35+ 张图，完整视觉生产栈
- → PDF + Word + WeChat md + HTML publicationPDF + Word + 公众号 md + HTML 刊物
- → 3 sign-off checkpoints (outline / draft / final)3 个 sign-off 节点（outline / draft / 定稿）
Industry deep-dive · macro thesis · policy assessment · flagship publication行业深度 · 宏观主题 · 政策评估 · 旗舰刊物
AI vibe coding,
with discipline.有纪律的
AI vibe coding。
I treat AI like a junior analyst: fast, broad-knowledge, untrained instinct. The skill's job is to design the track it runs on, and to catch its mistakes before they reach the page. The 8 steps below sit behind analyst-research's medium and heavy modes; light mode compresses the same logic into 6. The badges mark where a human decides, and where the quality gates are.我把 AI 当成一个初级分析师：快、知识面广、但直觉没经过训练。Skill 的任务，就是给它铺好跑的轨道，并在错误进入正文之前拦下来。下面这 8 步是 analyst-research 的 medium 和 heavy 两档骨架，light 档把同样的逻辑压成 6 步。标签标出了哪里由人拍板、哪里是质量门。
- 01
Hypothesis lock锁定假设
Human人工A single-sentence research question, fixed before any search begins. It travels with the project file from day one — never back-derived from the draft.一句话写清研究问题，在动手检索之前就定下来。从第一天起跟着项目文件走，绝不事后从草稿里反推。
- 02
Broad-spectrum search广谱检索
Five source categories at once: official statistics, multilateral institutions (IMF / World Bank / IEA), central banks, sell-side banks, primary corporate filings. PDFs downloaded in full, not just the abstract.五类来源一起上：官方统计、多边机构（IMF／世行／IEA）、央行、卖方投行、公司原始披露。PDF 整篇下下来，不只看摘要。
- 03
Outline + visual plan提纲 + 配图规划
Sign-off人工确认The user signs off on the section structure before any prose is written. How many charts each section gets, and of what type, is planned alongside the argument.章节结构先给用户过目、确认了再动笔。每节配几张图、配什么类型，跟文字论证一起想清楚。
- 04
Data extraction + chart production取数 + 出图
Every chart pairs a CSV with a Python script. chart_template.py enforces palette, fonts and label legibility. The number in the title must match the data on the chart.每张图都对应一组 CSV 加一段 Python 脚本。chart_template.py 统一配色、字体和标签可读性。标题里的数字，必须和图上的数据对得上。
- 05
Draft in one continuous pass一气呵成写初稿
A single LLM writes end-to-end. Citations go in as you write, never bolted on afterward. Filler and self-narration are grep'd out.单 LLM 从头写到尾。边写边标引用，绝不写完再补。废话和自说自话，用 grep 清掉。
- 06
Self-critique, 6 dimensions自我批评，6 个维度
Quality gate质量门Facts, unit consistency, citation accuracy, cross-section coherence, argument flow, language register — three rounds. Anything that fails routes surgically back to step 5.事实、单位一致、引用准确、跨节连贯、论证流畅、语言风格，六个维度查三轮。不过关的，精准打回第 5 步。
- 07
Independent verification独立核验
Quality gate质量门Every load-bearing number is cross-checked through the verifying skill against whitelisted primary sources. Whatever fails gets a ⚠️ flag or gets cut.每个撑着结论的数字，都用 verifying skill 对着白名单里的原始来源交叉核对。核不过的，要么标 ⚠️，要么删掉。
- 08
Final sign-off + derivations定稿 sign-off + 派生
Human人工The user reads the full PDF, marks revisions and signs off. Only then come the derivations: WeChat-paginated markdown, HTML publication, Word.用户读完整篇 PDF、标好修改、签字确认。之后才派生：公众号分页 md、HTML 刊物、Word。
One trigger, three workflows, shared production tooling.一个触发入口，三套工作流，共用一套生产工具。
analyst-research/ ├── SKILL.md # trigger + mode routing ├── references/ │ ├── workflow_light.md # 6 steps │ ├── workflow_medium.md # 8 steps │ ├── workflow_heavy.md # 11 steps │ └── report_style_spec.md # visual production rules └── scripts/ ├── chart_template.py # shared chart styling └── publication-style-template.htmlanalyst-research/ ├── SKILL.md # 触发 + mode 分流 ├── references/ │ ├── workflow_light.md # 6 步 │ ├── workflow_medium.md # 8 步 │ ├── workflow_heavy.md # 11 步 │ └── report_style_spec.md # 视觉生产规范 └── scripts/ ├── chart_template.py # 统一图表样式 └── publication-style-template.html
Each model does the part it's best at.每个模型只做自己最擅长的那一段。
IB-grade English writing, structural skeletons, integration and end-to-end orchestration. The default model at every step.英文 IB 级写作、搭结构骨架、做整合、端到端编排——每一步的默认模型。
Numerical precision, unit-consistency critique, fact-checking and the red lines on language register. The only outside critic in step 6.盯数值精度、查单位一致、核事实、把语言风格的红线——第 6 步唯一的外部批评者。
Chinese-polishing patterns, since distilled into a 6-point style guide that Claude now applies on its own — no longer called at runtime.中文润色的路数，现已提炼成 6 条风格指南，由 Claude 直接套用，运行时不再调用。
See what it produces.看看它能产出什么。
A 15,000-word Saudi Vision 2030 deep-dive, built end-to-end in heavy mode.一份用 heavy mode 端到端跑出来的 1.5 万字沙特愿景 2030 深度研究。
View the example report查看示例报告