Product Deep Research

Type one product name. Get back a 30,000-word research report, with sources, in 25 minutes.

This skill turns a single command like research Linear into a fact-checked, primary-source-driven product teardown. It runs 7 sub-agents in parallel, makes ~200 web requests, and outputs one shareable markdown file.

It also catches your assumptions when they're wrong. Tell it "research Final Round AI, the YC W24 a16z-backed company" — it will quietly look it up, find that's not true, and surface the correction in the report instead of writing fiction.

What you get

One .md file with 7-8 chapters covering everything you'd normally spend a full day digging up:

• Founders & philosophy — who built it, what they believe, with verbatim quotes and URLs
• Timeline & business model — every pricing tier, ARR triangulation, what they refused to build
• Growth engine — PLG flywheel / sales motion / paid acquisition / SEO moat / cross-signal diagnosis
• Use case map — 6-10 user archetypes with real quotes, which are well-served and which aren't
• Ecosystem — plugins, API, marketplace (skipped if the product has none)
• Community & KOLs — Reddit deep dive, top creators, who's promoting vs who's critiquing
• Competitive landscape — 8-10 rivals with switching narratives and growth-strategy comparison
• Risks & ceiling — what could break it, 5-year scenarios

Drop it into Obsidian. Push it to Feishu / Notion. Paste it into Cursor for context. Forward it to your boss.

Why it's different from "just asking Claude"

A plain chat session reads like a Wikipedia summary. This skill enforces six rules that make the output durable:

1. Challenges your brief. Your description of the product is often partly wrong (dates, prices, founders, funding). It verifies against primary sources and flags corrections in a > [!warning] block. No silent fixes.
2. Primary sources beat secondary. Founder blogs > official site > forum posts > podcasts > Hacker News >> "Top 10" listicles.
3. Real quotes only. Every claim about philosophy or pain points has the original English quote plus URL. If a quote can't be found, it says so — never paraphrases as if it were one.
4. Honest about gaps. Reddit blocked? X needs login? It says so in the Sources section instead of fabricating.
5. Length budgets per module. No 200-word "growth" sections. The Growth module is 6-8k characters; lighter modules are 2.5-3.5k. Sub-agents get the number in their prompt.
6. Adapts to product type. An indie tool, a VC SaaS, a big-tech-internal product, and a hardware company all need different research. A Phase 0 classifier picks the right framing before any module runs.

A real run, in numbers

Triggered with research Final Round AI 中文输出:

The report opened with a > [!warning] block correcting the user's stated facts about funding, then walked through a quantitative growth diagnosis (traffic -27% MoM, competitor Cluely already overtook on traffic, ARR triangulation across three methods), and ended with weighted 5-year scenarios.

The entire thing was shareable in one paste.

How to trigger it

The skill activates on phrases like:

research Linear
deep dive on Cursor
帮我调研一下 Notion
competitive analysis on Heptabase
I want to understand how Figma got to scale

Or explicitly: /product-deep-research <product name>

It will ask three things up front (and skip whatever you've already provided):

1. Exact product name
2. Output language (English / Chinese / etc.)
3. Optional: do you want it filtered through your product? If yes, every module ends with implications for yours. If no — the default — the research stays neutral.

Then it shows you the planned module set, you say go, and it runs.

The workflow

Phase 0   Classify the product along 4 dimensions
          (funding · growth motion · extensibility · platform)
              ↓
Phase 1   Capture name + language + optional lens
              ↓
Phase 2   Plan module set (drop modules that don't apply)
              ↓
Phase 3   Run sub-agents
          Batch A:  Module 1 (founders) — establishes anchors
          Batch B:  Modules 2-3, 4, 4b   — in parallel
          Batch C:  Modules 5, 6, 7, 8   — in parallel
              ↓
Phase 4   Generate MOC index + compile to single .md file

Each sub-agent saves its module to [product]/research/NN-module.md. Phase 4 strips the per-module frontmatter, demotes the headings, and concatenates everything into the final deliverable at [product]/[product]-research.md.

If you're an Obsidian user, the intermediate folder stays around as a vault. Otherwise delete it — the single file is the deliverable.

Adapts to whatever you point it at

Phase 0 reads four dimensions and rewrites the rest of the plan accordingly:

So Obsidian gets researched as Indie / PLG / Plugin / Desktop — founder manifesto deep dive, plugin economy chapter, no enterprise sales section. Salesforce gets Public / Sales-led / Marketplace / Web — SEC filings, Gartner quadrants, Dreamforce circuit. Same skill, totally different research.

Install

Drop the skill folder into your Claude Code skills directory:

unzip product-deep-research.skill -d ~/.claude/skills/

Structure:

product-deep-research/
├── SKILL.md                       main workflow (Phase 0-4)
└── references/
    ├── module-prompts.md          8 sub-agent prompt templates
    └── output-format.md           Obsidian markdown conventions

New conversation, then trigger naturally or with the slash command.

Output format

Obsidian-flavored markdown:

• YAML frontmatter per file (title, type, module, date, tags)
• Wikilinks for cross-references: [[01-founders-philosophy]], [[Founder Name]]
• Callouts: > [!info] > [!quote] > [!warning] > [!danger] > [!important] > [!tip]
• Tables for structured comparisons
• Every claim backed by a URL in a per-file Sources section

Renders cleanly in Obsidian. Imports cleanly into Feishu (lark-drive +import --type docx). Pastes cleanly into Notion. The frontmatter is harmless if your reader doesn't parse it.

When not to use it

This skill is heavy. Don't reach for it when:

• You just want a one-line answer ("what does Linear cost?")
• The product is in stealth or under 3 months old (modules will be thin — it'll tell you upfront)
• It's an internal company tool with no public footprint (it'll refuse)
• You need a one-off data point, not a teardown

For those, just ask normally.

What it's good for

• Competitive research — full landscape in a morning instead of a week
• Investment diligence — Module 3's ARR triangulation + Module 8's risk synthesis is most of an IC memo
• Job prep — turn on the lens parameter and every module ends with "implications for your situation"
• Content — the output is rich enough to spin into a blog series, a Twitter thread, or an internal training deck
• PM context — drop the .md into Cursor / Claude as background before designing a competing product

Limits worth knowing

• Reddit blocks scraping — sub-agents use Google site:reddit.com instead; some threads stay invisible
• X requires login — covered partially via search; private follower-only content stays hidden
• Ad libraries (Meta / Google / TikTok) rate-limit — sub-agents check them but sometimes only the meta-guides come back; the report says so
• Self-reported numbers stay labeled as unaudited (e.g., "$10M ARR per founder LinkedIn — confidence medium")
• 2025+ pricing changes frequently — verified at run time, but a report from 3 months ago may be stale

That's it. Run it on a product you actually care about and judge the output.