analyst-research

analyst-research · investment research workflow skill

A field-validated AI-assisted research workflow for investment analysts and policy researchers, packaged as a reusable Claude skill. Built on the methodology that produced the Saudi Vision 2030 deep-dive (35 figures, 15k+ words). Three scope modes; user picks at trigger time.

License: MIT. Copyright © 2026 Ligen ligen.thu@gmail.com. See LICENSE.

Step 0 — pick a mode (REQUIRED before loading references)

When this skill is triggered, before loading any reference file, ask the user to pick a scope. Present these three options verbatim:

This skill has three scope modes. Pick one based on your project size:

  light    4-5 page decision memo, 0 charts, ~15 min budget
           Single LLM session. Pure markdown footnote citations.
           Use for: exec brief, internal memo, quick decision support.

  medium   12-15 page topic analysis, 6-10 charts, ~1 h budget
           Single LLM. PDF + Word derivations. Sign-off checkpoint after draft.
           Use for: topic deep-dive, board memo with data, same-day analysis.

  heavy    Flagship report 30-40 pages / 15k+ words, 25-35+ charts, ~2-3 h budget
           Single or multi-LLM. PDF + Word + WeChat md + HTML publication.
           Runs the full 11-step staged workflow (framing → sourcing →
           analysis → drafting → review), with 3 sign-off checkpoints.
           Use for: industry deep-dive, macro thesis, policy assessment,
                    flagship investor publication.

Which mode fits your project? (reply with light, medium, or heavy)

If the user's trigger message already contains explicit scope hints (page count, chart count, time budget), infer the mode and ask one-line confirmation instead of presenting the full menu:

"Sounds like ~10 pages with a few charts — medium mode. Going with that?"

After the mode is confirmed, ask one short language question: "Report language — English (default), or another language?" (see Language policy below). Then load the reference files for that mode.

Loading order per mode

First always read references/workflow.md (≈40 lines, mode router). It points you to the mode-specific workflow file. (Each reference also has a .zh.md Chinese mirror; the agent reads the English .md.)

light mode

1. Required references/workflow.md — overview + mode router
2. Required references/workflow_light.md — 6-step skeleton, soft stops only, BLUF executive summary, grep self-check
3. Required references/_quarto-light.yml — Quarto template optimized for a 4-5 page memo (no TOC, no number sections, footnote citations, 11pt body)
4. No charts. No bibliography. No HTML/WeChat derivation. Skip report_style_spec.md and chart_template.py.

Then proceed to workflow_light.md §1 hypothesis lock to start the 6-step flow.

medium mode

1. Required references/workflow.md — overview + mode router
2. Required references/workflow_medium.md — 8-step skeleton, single LLM, one sign-off checkpoint after draft
3. Required references/report_style_spec.md — visual spec for the 6-10 charts (chart_template interface contract, color palette, font policy)
4. Required references/_quarto-medium.yml — Quarto template (footnote citations, no .bib, 11pt body, Songti SC CJK)
5. On demand scripts/chart_template.py — chart styling implementation

Then proceed to workflow_medium.md §1 onboarding to scaffold the project and start.

heavy mode

1. Required references/workflow.md — overview + mode router
2. Required references/workflow_heavy.md — full 11-step skeleton, multi-LLM optional, 3 sign-off checkpoints (outline / draft / final)
3. Required references/report_style_spec.md — visual spec including HTML publication derivation
4. On demand scripts/chart_template.py — chart styling implementation
5. On demand scripts/publication-style-template.html — HTML publication template
6. On demand scripts/author.jpg — author photo placeholder for HTML page

Quarto template for heavy mode comes from report_style_spec.md §5.1 (no separate _quarto-heavy.yml file — the spec is the source of truth for heavy mode quarto config).

Then proceed to workflow_heavy.md §1.3 new-project onboarding to scaffold the project (copies the whole analyst-research/ folder into project root) and start.

Mode comparison at a glance

Quick summary below; authoritative source of truth is MODE_REGISTRY.md. Edit that file first when mode parameters change, then propagate here.

Dimension	light	medium	heavy
Output length	4-5 pages	12-15 pages	30-40 pages / 15k+ words
Charts	0	6-10	25-35+
Runtime budget	~15 min	~1 h	~2-3 h
LLM model	single	single	single or multi-LLM
Workflow steps	6	8	11
Hard stops	0	1 (sign-off after draft)	3 (outline / draft / final)
Derived outputs	PDF + Word	PDF + Word	PDF + Word + WeChat md + HTML
Citation	Markdown footnote	Markdown footnote	BibTeX + APA
Project scaffolding	minimal	minimal	full (10 numbered dirs)
Chart template	n/a	shared (chart_template.py)	shared (chart_template.py)

For per-mode file dependencies, mode-upgrade trajectory, and retrospective section pointers, see MODE_REGISTRY.md.

When NOT to use this skill

• Single-piece news commentary → use market-research-skills:topic-brief
• Slide deck / PPT → use deckster-slide-generator
• One-shot Q&A (no written report) → just answer directly, no skill
• Pure literary or marketing copy → not this skill's domain
• Tool/script-only project with no report output → not this skill's domain

This skill is built for synthesis-and-analysis of an existing body of research and data (it does not build original models, per workflow_heavy.md §2.2). It works best on topics with a deep existing literature (IMF / World Bank / IEA / BIS, investment-bank and consulting research, academic papers, regulatory disclosure). Bleeding-edge, news-driven topics with thin institutional coverage are a poor fit.

How to upgrade or downgrade scope mid-project

If you start a project in light and find it needs more depth, re-trigger the skill in medium — the workflow's first step is identical (hypothesis lock), so the early work transfers. Going medium → heavy is the same. Downgrading is harder (you've already invested in scaffolding); cut deliverables rather than re-run.

Project scaffold lives outside the skill

The skill stays read-only in ~/.claude/plugins/.../analyst-research/. Per-project artifacts live in the user's project directory.

For light mode, the skill creates only _quarto.yml (copied from _quarto-light.yml) and a single working file in the project. No subfolders.

For medium mode, the skill creates _quarto.yml (from _quarto-medium.yml), 5_scripts/_path.py (sys.path injector to load chart_template), and minimal numbered output directories.

For heavy mode, the skill copies the entire analyst-research/ folder into the project root as a local working copy, then creates the full 10-numbered-directory scaffold per workflow_heavy.md §11. This lets project-level overrides (palette, fonts, domain conventions) live in the local copy without polluting the skill.

Skill evolution

After each project closes, follow the retrospective section of the mode-specific workflow file (workflow_heavy.md §9, workflow_medium.md §8, workflow_light.md §6) to decide which project learnings get promoted back to the skill itself. The skill is git-versioned; each promotion bumps minor version. Major architectural changes bump major version.

Language policy

This skill is bilingual (English + Chinese) following the marketplace convention. Every authored doc exists in two files: an English .md (e.g. SKILL.md, references/workflow_heavy.md) and a Chinese mirror .zh.md (e.g. SKILL.zh.md, references/workflow_heavy.zh.md). English is the single source of truth; the .zh.md is a synchronized translation, not an independent version. Edit protocol, no exceptions: always write or change the English .md first, then propagate the same change to the .zh.md translation in the same change-set. Never edit only the Chinese, and never let the two drift; when wording conflicts, the English .md wins.

1. Conversation replies follow the user's chat language. English chat → reply in English; Chinese chat → reply in Chinese; likewise for any other language. This is runtime behaviour, not a stored artifact.
2. Report language defaults to English. At Step 0 / onboarding, after the mode is picked, ask one short question: "Report language — English (default), or another language?" If the user does not specify, write the draft in English; if they name another language, write in that. Lock the choice in the project CLAUDE.md. This supersedes any older "draft follows hypothesis language" rule.
3. Language-conditioned grep redlines. For an English report the Chinese colon-ratio redline is skipped and the unescaped-$ redline is mandatory (escape dollar amounts as \$). For a Chinese report the reverse applies.

The Chinese mirror of this file is SKILL.zh.md. 中文版见 SKILL.zh.md。