Research

Research

Structured research with multi-source synthesis. The key insight: planning your sources before executing dramatically improves output quality.

This skill scales from "quick answer with built-in web search" to "deep multi-tool investigation with browser automation, dedicated search APIs, and AI-assisted analysis" — depending on what tools you have connected.

Connectors

The research skill works at every level — more tools unlocked = deeper research capability.

Connector	Required	What it unlocks
None (built-in)	—	WebSearch, WebFetch, sub-agents — enough for most research
Web scraper	Optional	Scrape specific pages for structured data (pricing tables, feature lists, changelogs)
Exa search	Optional	Semantic search — finds conceptually related content, not just keyword matches
Browser automation	Optional	Drive NotebookLM, access paywalled content via logged-in sessions, interact with web apps

Start with built-in tools. Add connectors as you hit walls — "I can't access this page" or "I need deeper analysis on this PDF."

Configuration

Load env.md for the research output path. No skill-specific config needed.

Research outputs are saved to the path defined in your env.md under Research output. Default: ~/research/

File naming: YYYY-MM-DD-[subject-slug]-[type].md

Tool Reference

Built-in (always available)

Tool	Best for	Notes
WebSearch	General queries, recent news, product info	Good breadth, limited depth per result
WebFetch	Reading a specific URL	Converts HTML to markdown, summarizes long pages
Sub-agents	Parallel research tasks	Launch multiple agents to research different facets simultaneously

With Web Scraper (Firecrawl)

Tool	Best for	Notes
Firecrawl scrape	Extracting structured data from JS-heavy pages	Great for pricing tables, feature matrices, changelogs
Firecrawl extract	Pulling specific fields from a page via schema	Define a JSON schema → get structured data back
Firecrawl search	Searching a specific site or domain	More targeted than general web search

With Exa Search

Tool	Best for	Notes
Exa web search	Finding conceptually related content	Semantic search — finds content by meaning, not just keywords
Exa get code context	Finding code examples and technical implementations	Better than general search for "how does X implement Y"

With Browser Automation (Chrome MCP)

Tool	Best for	Notes
NotebookLM	Deep analysis of long documents, PDFs, research papers	Navigate to notebooklm.google.com, upload sources, use "Deep Research" mode
Logged-in content	Paywalled articles, internal dashboards, authenticated APIs	Uses your active Chrome session
Interactive research	Filling search forms, navigating multi-page results, downloading reports	Full browser control for pages that don't expose clean APIs

Workflow

Step 1: Define Scope

Use AskUserQuestion — one question at a time:

Q1: "What are we researching? Give me the subject and the central question you want answered."

Wait. Then:

Q2: "What type of research is this — product comparison, technical deep-dive, or knowledge synthesis? And where will the output go (doc, Slack, email)?"

The output channel matters — a Slack message needs a different format than a research doc.

Step 2: Plan Sources (Most Important Step)

Specify exactly which sources you'll use. This is the step that separates good research from noise.

Work with the user to build a research plan:

Source Checklist:

• Official sources (documentation, help centers, announcements)
• User sentiment (X, Reddit, G2 reviews, HackerNews)
• Technical details (DevRel content, API docs, GitHub repos/issues)
• Market/competitive (analyst reports, industry press, BuiltWith)
• Expert analysis (conference talks, interviews, newsletters)

Recency Requirements:

• Most recent only (software features change fast)
• Historical trends (market analysis)
• Recency not critical (general research)

Tool Selection — match tools to sources:

Source type	Built-in	With scraper	With Exa	With browser
Official docs	WebFetch	Firecrawl (JS pages)	—	—
User sentiment	WebSearch	—	Exa (semantic)	Chrome → X/Reddit
Pricing/features	WebFetch	Firecrawl extract	—	Chrome (dynamic pricing)
Technical deep-dive	WebSearch + WebFetch	Firecrawl (GitHub)	Exa code context	—
Long documents/PDFs	WebFetch (limited)	—	—	Chrome → NotebookLM
Paywalled content	—	—	—	Chrome (logged-in)
Real-time discussion	WebSearch	—	Exa	Chrome → X/Grok

Save the plan before executing.

See references/source-mapping.md for source selection guidance by research type.

Step 3: Execute (Parallel, Save As You Go)

Run everything you can in parallel. The key is to start slow tasks first.

Execution order:

Minute 0:    Start NotebookLM Deep Research (if using — takes 3-5 min)
Minute 0:    Launch sub-agents for independent research threads
Minute 0-2:  Run web searches (WebSearch, Exa)
Minute 2-5:  Scrape specific pages (Firecrawl, WebFetch)
Minute 5:    Pull NotebookLM results via Chrome
Minute 5+:   Compile and cross-reference

Using NotebookLM (if browser automation available):

1. Navigate to notebooklm.google.com via Chrome MCP
2. Create a new notebook
3. Upload source materials (URLs, PDFs, documents)
4. Use "Deep Research" mode with a specific question
5. Wait for results (3-5 minutes), then pull the findings

Using Exa (if available):

Exa web search: "[conceptual query]" — finds semantically similar content
Exa get code context: "[how does X implement Y]" — finds technical implementations

Using Firecrawl (if available):

{
  "url": "https://example.com/pricing",
  "formats": ["json"],
  "jsonOptions": {
    "prompt": "Extract all pricing tiers with features and prices",
    "schema": { ... }
  }
}

Save findings from each source in a structured format:

# [Source Name] Research: [Subject]
## Date: [Date]
## Tool used: [WebSearch / Firecrawl / Exa / NotebookLM / etc.]

## Key Findings
- Finding 1 (Source: [citation with URL])
- Finding 2 (Source: [citation with URL])

## Raw Data
[Summary of what the source returned]

## Confidence Level
[HIGH — official source / MEDIUM — credible analysis / LOW — anecdotal]

## Questions for Follow-up
- [What couldn't this source answer?]

Step 4: Synthesize

1. Read all saved findings
2. Select template based on research type AND output channel (see references/output-templates.md)
3. Ask for personal examples — "Do you have a personal use case that would make this concrete?"
4. Create "At a Glance" table — 3-5 key dimensions as analysis spine (skip for quick takes)
5. Draft synthesis answering the key question
6. Iterate with the user — first draft is never the final version

Research Type	Output Channel	Template
Any	Slack/email	Quick Take (short, scannable)
Product comparison	Doc	Comparison (recommendation + At a Glance table)
Technical deep-dive	Doc	Deep-dive (architecture, tradeoffs)
Knowledge synthesis	Doc	Synthesis (themes, conflicting views, gaps)

Step 5: Red-Team (Required)

Before finalizing, challenge the analysis:

• Asymmetric treatment — Did you cite negatives for one side but not the other?
• Weak sample sizes — Is "39 posts" enough to generalize? Note sample sizes.
• Claims that don't hold — What scenarios break your recommendation?
• Missing perspectives — What would a skeptic say?
• Stale data — Is any source older than 6 months? Flag it.

Feed findings back. Expect 2-3 rounds of iteration.

Verification Checklist:

• Is this fact in one of our saved reports?
• Is the source credible? (Official docs > blog posts > tweets)
• Is the data recent? (Flag anything >6 months old)
• Do multiple sources agree? (Note conflicts explicitly)
• Are sample sizes noted for sentiment data?
• Are confidence levels assigned? (HIGH/MEDIUM/LOW per finding)

Error Handling

Scenario	Response
Source returns no results	Note the gap, try alternative query or different tool, move on
Conflicting sources	Document both views with confidence levels — don't silently pick one
Page requires login	If Chrome MCP available, try logged-in session. Otherwise, note as inaccessible.
Firecrawl/scraper times out	Retry with longer wait, fall back to WebFetch, note if data is missing
User wants quick answer	Use "Quick Take" template — skip Steps 2-3, go straight to synthesis
Too many sources, losing focus	Return to the key question from Step 1. Cut sources that don't directly answer it.

Customization Notes

How to adapt this skill to your own setup.

• Source mapping — Edit references/source-mapping.md to match your research domains. A marketing team might add "Social listening tools" and "Ad library analysis."
• Output templates — Edit references/output-templates.md to add templates for your common formats (Notion, Confluence, Google Docs, etc.)
• Tool selection — The tool reference table above shows what each connector unlocks. You don't need all of them — add tools as you find limits with the built-in ones.
• NotebookLM workflow — If you use NotebookLM regularly, consider creating a separate NotebookLM-specific reference doc with your preferred prompts and source upload patterns.
• Save location — Research outputs go where you set in env.md. Consider version-controlling them if your research builds on itself over time.