A research operating system for AI agents. Encodes the discipline — modes, personas, sourcing tiers, counter-consensus, falsifiability — as a reusable framework. Ships with a polished default template; bring your own. Opinionated by default, extensible by design.
Quick links: Architecture · Bring Your Own Template · Visual Showcase · Start Here · 60-Second First Run · Modes · Personas · Roadmap
This is not just an HTML generator. It's a research operating system for AI agents — a reusable framework for turning long research conversations into rigorous, interactive research artifacts.
| Mode | When to use it |
|---|---|
| Generate | First-pass research on a new topic |
| Discovery | 5-question dialogue → user picks brief or HTML |
| Critique | "I wrote an analysis. Find the weaknesses." |
| Refresh ⭐ | "Update my Q1 NVIDIA report with Q2 data" |
| Compare | "NVIDIA vs AMD with side-by-side tables" |
| Drill-down | "Expand customer concentration into its own deep dive" |
Personas — Core lenses:
| Persona | Lens it applies |
|---|---|
| Tier 2 Investor (default for stocks) | "Long, short, or neutral? At what price?" |
| Industry Strategist (default for sectors) | "Where is profit migrating across cycles?" |
| Engineer | "Is the technical bet sound? Show benchmarks." |
| Competitor | "If I were a rival, where would I attack?" |
Personas — Advanced lenses:
| Persona | Lens it applies |
|---|---|
| Tier 1 Investor | "Is this a fund-returner? Why now?" |
| CEO | "What would I do tomorrow morning if I ran this?" |
| Disruption Scout | "What kills this business in 5 years?" |
Same topic, different persona = different artifact.
This skill is built in three concentric layers. Each layer is independently replaceable.
┌──────────────────────────────────────────────────────┐
│ Layer 3 — Presentation Template (the shell) │
│ HTML / CSS / JS / interactions / visual identity │
│ ← REPLACEABLE: bring your own template │
├──────────────────────────────────────────────────────┤
│ Layer 2 — Content Schema (the data) │
│ Hero stats · sections · CN_NODES · DETAIL_DATA │
│ · glossary · 22-part menu │
│ ← STABLE INTERFACE: same shape across templates │
├──────────────────────────────────────────────────────┤
│ Layer 1 — Research Workflow (the core IP) │
│ 6 modes · 7 personas · sourcing tiers · counter- │
│ consensus framework · falsifiability rules │
│ ← THE FRAMEWORK: this is what makes it useful │
└──────────────────────────────────────────────────────┘
Layer 1 is the IP — the research discipline. It's the same regardless of how you render it. Layer 2 is structured data — what facts, sources, and judgments get produced. Layer 3 is replaceable — visual, interactive, branding. Bring your own.
The default template (assets/template.html) is one expression of Layer 3. The skill is not limited to it.
The default template is opinionated: warm-beige + brick-red, 22-part editorial structure, sidebar nav, ⌘K search, click-to-expand drawers, SVG ecosystem map. It's polished. It looks great. It's also one of many possible expressions.
For your fund, your startup, your industry, your team — you can fork the template and adapt:
- Visual identity (different palette, fonts, density)
- Section structure (don't want 22 parts? use 8)
- Interaction model (drawer instead of accordion? tabs instead of sections?)
- Charts and maps (different ecosystem visualization)
- Sidebar / navigation style (top tabs? collapsible?)
- Output format (HTML / markdown / Notion / Slack message / PPT)
- Start with
assets/template.htmlas reference for what slots exist. - Replace the visual + interaction shell with your own.
- Keep stable slots for content:
title,hero,sections,ecosystem_nodes,drawer_data,glossary. - (Optional) Add your own validation checks for your template's wiring.
- Save it as your team's reusable template — store it anywhere, just tell the AI where.
"Generate a NVIDIA field guide using my template at /path/to/my-template.html"
The skill will fill your template's slots with research-discipline content (Layer 1 + Layer 2), preserving your visual + interaction choices (Layer 3).
✅ Use default: prototype, demo, blog post, anything where editorial polish matters and your audience expects "VC research-style" formatting.
❌ Use your own: branded fund deliverables, internal team reports, integrations with Notion/Slack/email, anything where your visual identity matters more than the default's.
The skill is opinionated by default, extensible by design. We provide the discipline; you decide the shell.
These are real outputs of the default template — not the skill's only output style.
Industry Strategist lens on a fintech sector — interactive SVG ecosystem map. Click any node to jump to the relevant section.
Tabbed section structure for multi-segment industry analyses — switch between subsectors without leaving the page.
Industry-level profit-pool analysis — value-chain economics with explicit margin distribution by participant type.
These take ~3 minutes to generate and look like this. Your template can look completely different.
| I use... | Do this | Best for |
|---|---|---|
| Claude.ai | Download v3.0.2 zip → upload as a Skill in Settings | Non-technical users |
| Claude Code | git clone into ~/.claude/skills/ |
Power users on Anthropic stack |
| Codex CLI (OpenAI) | git clone into ~/.codex/skills/ |
OpenAI ecosystem |
| Gemini CLI (Google) | gemini skills install <repo-url> |
Google ecosystem |
| Cursor | git clone into project's .cursor/skills/ |
In-editor research |
Detailed install commands are in Install. Once installed, jump to 60-Second First Run.
After install, paste this:
"Analyze NVIDIA fundamentals as a Tier 2 investor."
You'll get an interactive HTML field guide with sourced facts, a counter-consensus thesis, and explicit Bull/Bear scenarios. Open in browser to navigate.
Then try one of these to feel the suite:
"Critique that report — find weaknesses."
"Drill into the customer concentration disagreement."
"Refresh this report with the latest data."
"Compare NVIDIA and AMD side by side."
Each command activates a different mode. Together they make this a research lifecycle tool, not a one-shot generator.
Default: interactive HTML field guide (~110KB) — sidebar nav, ⌘K full-text search, click-to-expand drawers, clickable SVG ecosystem map, 2×2 strategic positioning matrix, structured analysis across up to 22 parts.
Alternative: markdown research brief (1–2 page memo) — counter-thesis, sourced facts, steel-manned Bull/Bear, what-to-watch list. Available when you ask for "brief" / "markdown summary" or pick "brief only" at the end of Discovery mode.
Both formats follow the same quality bar: every fact has a [Source: X] citation, every judgment is falsifiable, every artifact has explicit counter-consensus framing.
Custom formats: bring your own template (HTML / markdown / JSON for Notion / etc.) — see Bring Your Own Template.
"Analyze NVIDIA fundamentals"
"Research the semiconductor industry"
"Map Stripe's strategic ecosystem"
The skill auto-detects mode + persona. Override with:
"Analyze NVIDIA fundamentals as a Tier 1 investor"
"Research the semiconductor industry from a CEO perspective"
"Help me think through NVIDIA before generating a field guide"
"Do a research dialogue on the AI agent landscape"
5 questions (consensus → contradictions → thesis → facts → steel-manned Bull/Bear), then choose: brief or HTML.
"Review my NVIDIA analysis. Find the weaknesses."
"Critique this report — find unsourced claims and weak counter-consensus."
Provide your existing analysis. Output: markdown critique (no HTML) with strongest claims, weakest claims, what's missing, specific fixes.
"Refresh my NVIDIA Q1 report with the latest data"
"Update my Stripe analysis from January"
"复盘 my AI agent landscape report — what's changed?"
Provide the prior field guide. The skill refreshes every number, flags material changes, re-evaluates your counter-thesis, adds a "What changed since [date]" callout.
"Compare NVIDIA and AMD"
"Stripe vs Adyen — which is the better long?"
Comparative field guide with hero comparison table (10–15 dimensions), side-by-side tables in each section, explicit scoring rubric.
"From my NVIDIA report, drill into customer concentration"
"Take the 'AI factory operations' thesis and make it standalone"
Treat one subsection as the entire topic of a NEW field guide — full 22-part treatment.
Most research tools are one-shot generators. Real research isn't:
- Look at a topic from multiple angles (= personas)
- Update the analysis as the world changes (= Refresh ⭐)
- Compare options before deciding (= Compare)
- Stress-test your own conclusions (= Critique)
- Drill into the cruxes (= Drill-down)
- Coach yourself through the thinking (= Discovery)
Garry Tan's gstack made this point for engineering — a virtual team beats a single assistant. This skill applies the same principle to research.
One topic. Six modes. Seven personas. Multiple output formats. Repeated use over time.
This skill follows the open Agent Skills standard. Works in any compatible agent — but output quality depends on the model.
| Agent | Install path | Status |
|---|---|---|
| Claude.ai | Settings UI (zip upload) | ✅ Reference implementation |
| Claude Code | ~/.claude/skills/ |
✅ Reference implementation |
| Codex CLI (OpenAI) | ~/.codex/skills/ |
🟡 Format compatible, output untested |
| Gemini CLI (Google) | ~/.gemini/skills/ |
🟡 Format compatible, output untested |
| Cursor | .cursor/skills/ (project only) |
🟡 Format compatible, output untested |
🟡 = format loads correctly; rendered HTML quality not personally validated. The skill assumes strong reasoning + long context + web search — best on Claude Opus / GPT-5 / Gemini 2.5 Pro.
This skill produces research frameworks with falsifiable signals, sourced facts, and structured Bull / Bear analyses. It is not personalized investment advice and does not replace a licensed financial advisor.
Outputs are framed as analytical artifacts. Every output includes a footer disclaimer.
If you need personalized financial advice, consult a licensed advisor.
- Download the latest zip from releases
- Claude.ai → Settings → Capabilities → enable Code execution and file creation
- Customize → Skills → + → Upload a skill → select zip → toggle on
git clone https://github.com/klaywang24/interactive-field-guide-skill.git \
~/.claude/skills/interactive-field-guidecodex --enable skills
git clone https://github.com/klaywang24/interactive-field-guide-skill.git \
~/.codex/skills/interactive-field-guide4. Gemini CLI (Google) — see configuration notes ↓
gemini skills install https://github.com/klaywang24/interactive-field-guide-skill5. Cursor — see configuration notes ↓
mkdir -p .cursor/skills
git clone https://github.com/klaywang24/interactive-field-guide-skill.git \
.cursor/skills/interactive-field-guideThe skill needs four capabilities: (1) web search, (2) file write, (3) Python or Node, (4) long context (~50K tokens).
Claude.ai / Claude Code have all four enabled by default. Codex / Gemini / Cursor sometimes don't.
A. Web search must be enabled. In session: /tools list — look for google_search.
B. File write permissions. To skip per-write prompts:
C. Workspace trust (only if installing at .gemini/skills/). Run /trust, or install at user-scope ~/.gemini/skills/ instead.
A. Use Agent mode, not Ask. Top of chat panel → select Agent.
B. Web search must be on. Settings → Features → Agent → enable Web search.
C. Allow terminal commands. Settings → Features → Agent → enable Allow terminal commands.
D. Pick a strong model. Settings → Models → Claude Sonnet 4.6 / Opus 4.7 / GPT-5 / Gemini 2.5 Pro.
| Model | Quality | Notes |
|---|---|---|
| Claude Opus 4.7 | ⭐⭐⭐⭐⭐ | Reference implementation |
| GPT-5 / GPT-5-Codex | ⭐⭐⭐⭐ | Strong reasoning |
| Gemini 2.5 Pro | ⭐⭐⭐⭐ | Excellent long context |
| Claude Sonnet 4.6 | ⭐⭐⭐⭐ | Faster + cheaper |
| Gemini 2.5 Flash / GPT-4o-mini | ⭐⭐⭐ | OK for first drafts |
If output feels shallow, switch to a stronger model first.
interactive-field-guide-skill/
├── SKILL.md # Lightweight router (Layer 1 dispatch)
├── assets/
│ ├── template.html # Default presentation template (Layer 3)
│ └── screenshots/ # README visual showcase
└── references/
├── modes.md # Layer 1: 6 operational mode workflows
├── personas.md # Layer 1: 7 analytical persona lenses
├── structure.md # Layer 2: 22-part menu with skip rules
├── data-schemas.md # Layer 2: JS object schemas + slots
├── content-strategy.md # Layer 1: source tiers, falsifiability
└── pitfalls.md # Validation checks for default template
The split between Layers 1, 2, 3 lives in this folder structure. Replace assets/template.html (Layer 3) without touching references/ (Layer 1 + 2) — the framework still works.
| Example | Mode | Persona | Demonstrates |
|---|---|---|---|
| Fintech industry constellation | Generate | Industry Strategist | SVG ecosystem map with clickable nodes |
| Fintech sector navigation | Generate | Industry Strategist | Tabbed section structure for multi-segment topics |
| Global industry profit-pool | Generate | Industry Strategist | Profit-pool mapping across industry value chain |
| NVIDIA fundamentals | Generate | Tier 2 Investor | Counter-consensus thesis + Bull/Bear with falsifiability |
More examples will land in a companion awesome-field-guides repo (Q2 2026), including community-contributed templates.
Q2 2026 (in development):
- Companion
awesome-field-guidesrepo with example artifacts - Example community templates (minimalist white, Notion-style, 中文公众号)
- True 2-column mirrored layout for Compare mode
scripts/validate-field-guide.jsandscripts/smoke-test-field-guide.js(Playwright) as standalone files
Q3 2026 (planned):
- "Watch list" mode — passive tracking of 5–10 entities
agents/openai.yamlfor richer Codex / OpenAI ecosystem display- skills.sh marketplace submission
- Claude Code Plugin Marketplace registration
Q4 2026 (exploring):
- Voice-mode Discovery
- Collaborative critique (multiple agents debate the same analysis)
- Bilingual auto-output toggle
Skill doesn't trigger. Trigger is intentionally narrow — only fires for company / industry / market / competitive / strategy / investment research artifacts. Force it: "generate a field guide on X".
Mode not detected correctly. Be explicit: "Use Discovery mode for X" or "Run Critique on this".
Persona feels wrong. Specify: "...as a Tier 1 investor" or "...with an Engineer lens".
Output too shallow. Likely a model issue. Switch to Claude Opus 4.7 / GPT-5 / Gemini 2.5 Pro and re-run.
Refresh can't find prior data. Make sure you uploaded or pasted the prior field guide. The skill has no cross-session memory unless you provide it.
Compare looks single-column. v3.0.x uses comparison tables, not mirrored 2-column layout. Full mirrored layout is on the Q2 roadmap.
Default template visual broken. Re-run with: "Re-run the field guide and run all 9 validation checks." If it persists, the model rebuilt instead of editing — see SKILL.md Critical Rule 6.
A report is read once. A field guide is something you keep open, search, click around, return to.
But more importantly: a field guide is for navigation, not display. The 22-part structure, the falsifiable judgments, the sourced facts — these aren't formatting. They're the research discipline that makes the artifact trustworthy.
The skill encodes that discipline (Layer 1). The HTML is one shell (Layer 3). The discipline is the product.
Apache 2.0. Use, fork, modify freely. PRs welcome — especially platform-specific compatibility fixes, new persona contributions, and community templates.
Built and maintained by @klaywang24.
This is open infrastructure: research workflows, content schemas, and a default template. If it's useful to your fund / startup / team, fork it and make it yours.
If you build a custom template (or a new persona), submit a PR — community templates land in templates/community/.