dragonkid/deep-research

Deep Research (Exa)

Produce thorough, cited research reports by breaking topics into sub-questions and searching with Exa's category-aware tools.

Token Isolation

Always spawn Task agents for search and deep-read steps. Agent returns distilled findings only — main context stays clean.

When dispatching agents, include tool routing and category restrictions in the prompt. Agents don't see this skill — they only know what you tell them. Example:

Research [topic]. For GitHub projects, use `gh api repos/{owner}/{repo}` for
metadata/README — not web search. For community reception, use Exa
web_search_advanced_exa with category "tweet" or includeDomains ["reddit.com"].
Category restrictions: "company" allows NO date/text/domain filters;
"tweet" allows NO includeText/excludeText/includeDomains/excludeDomains.
Return distilled findings with source URLs.

Tool Routing

Not everything should go through Exa. Pick the right tool for each data source:

| Data source | Tool | Why | |-------------|------|-----| | GitHub repo metadata (stars, language, topics) | gh api repos/{owner}/{repo} | Structured data, zero noise, instant | | GitHub README | gh api repos/{owner}/{repo}/readme --jq '.content' \| base64 -d | Full content, no web scraping | | Community discussions (Reddit, HN, Twitter) | Exa web_search_advanced_exa | Category-aware, highlights mode | | News, papers, blogs | Exa web_search_advanced_exa | Category filters, date filters | | Full page content | Exa crawling_exa | Deep reads, maxCharacters control |

When dispatching research agents, include tool routing in the prompt so agents use the right tool for each source.

Workflow

Step 1: Understand the Goal

Ask 1-2 quick clarifying questions:

• "What's your goal — learning, making a decision, or writing something?"
• "Any specific angle or depth you want?"

If the user says "just research it" — skip ahead with reasonable defaults.

Step 1.5: Vault Cross-Reference

Before web research, check the Obsidian vault for existing notes on the topic. This avoids duplicating prior research and surfaces useful context:

Grep pattern="<topic keywords>" path="~/Documents/second-brain" glob="*.md"

If related notes found:

• Read the most relevant 2-3 notes for context
• Note what's already covered vs what's missing — focus web research on gaps
• Include existing note titles in the report's ## Related section later

If nothing found, proceed directly to web research.

Step 1.7: Recency Check

If the topic has strong time-sensitivity signals (product launches, breaking news, market events, "最近", "现在", "刚出的", recent dates, or anything where community reaction in the last 30 days matters more than authoritative analysis), add a note at the end of the report:

> Tip: 这个主题时效性强，可以用 `/last30days <topic>` 获取 Reddit/X/YouTube/Polymarket 的社区实时讨论和 engagement 数据，与本报告互补。

This complements deep-research's authoritative sources (papers, blogs, company data) with community sentiment and engagement signals that Exa cannot capture (Reddit upvotes, X likes, YouTube transcript content, Polymarket odds).

Step 2: Plan Sub-Questions

Break the topic into 3-5 research sub-questions. For each, decide the best Exa category:

| Sub-question type | Tool | Category | Why | |-------------------|------|----------|-----| | Current events, press | web_search_advanced_exa | news | Journalism, announcements | | Academic findings | web_search_advanced_exa | research paper | arXiv, peer-reviewed | | Company/market data | web_search_advanced_exa | company | Homepages, metadata | | Community sentiment | web_search_advanced_exa | tweet | Developer opinions, discussions | | Expert perspectives | web_search_advanced_exa | personal site | Independent analysis, blogs | | General / mixed | web_search_exa | (none) | Broad coverage |

Step 3: Search with Query Variations

For each sub-question, generate 2-3 keyword variations — Exa returns different results for different phrasings. Run variations in parallel, merge and deduplicate.

Example for sub-question "What are the main AI applications in healthcare?":

• Variation 1: "AI applications healthcare clinical deployment 2025"
• Variation 2: "machine learning medical diagnosis treatment real-world"
• Variation 3: "artificial intelligence hospital workflow automation"

Search parameters:

• Use highlights mode for search results (10x fewer tokens than full text)
• Set numResults: 10-15 per variation
• For news: add startPublishedDate for recency (last 12 months)
• For research papers: add includeDomains: ["arxiv.org"] if needed
• Aim for 15-30 unique sources total across all sub-questions

Step 4: Category Filter Restrictions

These restrictions cause 400 errors if violated:

• company: NO date filters, NO text filters, NO includeDomains/excludeDomains
• people: NO date filters, NO text filters, NO excludeDomains. includeDomains only accepts LinkedIn
• tweet: NO includeText, excludeText, includeDomains, excludeDomains, moderation
• financial report: NO excludeText
• research paper, personal site, news: all filters work
• Universal: includeText/excludeText only support single-item arrays

When a category restricts filters you need, drop the category and use web_search_exa with keywords in the query instead.

Step 5: Deep-Read Key Sources

For the 3-5 most promising URLs from search results, fetch full content:

crawling_exa(url: "<url>", maxCharacters: 10000)

Use text mode (not highlights) for deep reads — you need the full context to synthesize accurately.

Step 6: Synthesize Report

Structure the report with inline citations:

# [Topic]: Research Report
*Generated: [date] | Sources: [N] | Confidence: [High/Medium/Low]*

## Executive Summary
[3-5 sentence overview of key findings]

## 1. [First Major Theme]
[Findings with inline citations]
- Key point ([Source Name](url))
- Supporting data ([Source Name](url))

## 2. [Second Major Theme]
...

## Key Takeaways
- [Actionable insight 1] ([Source Name](url))
- [Actionable insight 2] ([Source Name](url))

Every takeaway must cite at least one supporting source so the reader can verify and judge independently.

## Sources
1. [Title](url) — [one-line summary]
2. ...

## Methodology
Searched [N] queries across [categories used]. Analyzed [M] sources.
Sub-questions investigated: [list]

Step 7: Deliver

• Short topics: post full report in chat
• Long reports: post executive summary + key takeaways, save full report to a file

Parallel Execution

For broad topics, parallelize sub-questions across Task agents:

Launch 3 research agents:
1. Agent 1: Sub-questions 1-2 (news + research paper categories)
2. Agent 2: Sub-questions 3-4 (company + general categories)
3. Agent 3: Sub-question 5 + cross-cutting themes (tweet + personal site)

Each agent searches, reads sources, and returns findings. Main session synthesizes into the final report.

Quality Rules

1. Every claim needs a source. No unsourced assertions.
2. Cross-reference: if only one source says it, flag as unverified.
3. Recency matters: prefer sources from the last 12 months.
4. Acknowledge gaps: if you couldn't find good info, say so.
5. No hallucination: if you don't know, say "insufficient data found."
6. Separate fact from inference: label estimates and opinions clearly.
7. Conclusions need citations: Key Takeaways and any concluding analysis must include inline source references, not just the body sections. The reader should be able to trace every conclusion back to its evidence without scrolling to the Sources list.

Company Research Shortcut

When the research target is a company or market landscape, use this pre-planned sub-question template instead of designing from scratch:

1. Discovery — category: "company", find companies in the space
2. Deep dive — no category, includeDomains for TechCrunch/Crunchbase/Bloomberg
3. News — category: "news", startPublishedDate for recent coverage
4. Social — category: "tweet", developer/user sentiment
5. People — category: "people", key executives and team

Dynamic tuning for company lists:

• User says "a few" → numResults: 10-20
• User says "comprehensive" → numResults: 50-100
• User specifies a number → match it

Output format for company lists: use structured table (name, description, funding, headcount) instead of prose. For single-company deep dives, use the standard report template.

Business-oriented research: When the goal is a business decision (investor diligence, market sizing, competitive positioning, vendor evaluation), read references/business-research-modes.md first to get the analysis checklist for that mode. Use the checklist to plan sub-questions, then execute searches with the tool routing and category rules from this skill. The checklist defines what to collect; this skill defines how to search.

Deep Search (Optional)

For complex topics that benefit from structured output, use Exa's deep search:

web_search_advanced_exa(
  query: "compare frontier AI model releases 2025-2026",
  type: "deep",
  numResults: 10
)

Deep search runs multiple query variations internally and returns synthesized results. Use type: "deep" when a single sub-question is complex enough to warrant multi-step reasoning. Don't use for simple factual lookups — auto is faster and cheaper.