deep-research

star 266

Thorough multi-source research with citations, analysis, and synthesis. TRIGGER: "research", "deep dive", "investigate", "find out about". NOT FOR: github-research, consulting-report use cases.

xg-gh-25 By xg-gh-25 schedule Updated 5/30/2026

name: deep-research description: "Thorough multi-source research with citations, analysis, and synthesis.\n TRIGGER: "research", "deep dive", "investigate", "find out about".\n NOT FOR: github-research, consulting-report
\ use cases." version: 2.0.0 produces_artifact: research tier: always

Deep Research

Conduct systematic, multi-source research that produces well-cited, comprehensive analysis. Never generate content based solely on general knowledge -- the quality of output depends directly on the quality and quantity of research conducted beforehand.

Output Location

Save research documents to:

~/.swarm-ai/SwarmWS/Knowledge/Notes/YYYY-MM-DD-<topic>.md

Once finalized, move to Knowledge/Library/ for long-term reference.

Workflow: 5-Phase Research

Phase 0: Intent Classification & Strategy Planning

Goal: Before any search, classify the research intent and plan the optimal strategy. Output a structured plan that drives all subsequent phases.

Step 1: Classify the research intent. Pick the PRIMARY intent:

Intent Signal Words Example
factual "what is", "how does", "explain" "How does Raft consensus work?"
competitive "vs", "compare", "alternative", "竞品" "SwarmAI vs OpenClaw"
landscape "overview", "landscape", "what's out there", "调研" "AI agent frameworks 2026"
how_to "how to", "implement", "build", "tutorial" "How to implement RAG with Bedrock"
breaking_news "latest", "just happened", "今天", "刚刚" "What did Anthropic announce today?"
trend "trend", "direction", "future", "趋势" "Where is agent memory heading?"
person_org person name, company name, "@handle" "Research Peter Steinberger"
deep_technical "architecture", "internals", "source code" "Claude Code SDK internal architecture"

Step 2: Determine search parameters. Based on intent, set these BEFORE Phase 1:

Parameter factual competitive landscape how_to breaking_news trend person_org deep_technical
search_depth basic advanced advanced basic basic advanced advanced advanced
time_range none month none year day/week year month none
topic general general general general news news general general
source_priority docs→papers→blogs product pages→HN→blogs industry reports→news→blogs GitHub→SO→tutorials news→social→blogs reports→expert blogs→news social→GitHub→blogs→news source code→docs→talks
min_sources 3 5 (both sides) 5 3 3 5 4 3
search_rounds 2-3 3-5 3-5 2-3 2-3 3-5 3-4 2-4

Step 3: Output the plan. Write this to chat before proceeding — it's your contract:

RESEARCH PLAN
━━━━━━━━━━━━
Intent:        <intent>
Query:         <original query>
Time range:    <time constraint>
Source focus:   <top 3 source types>
Search rounds: <N>
Key angles:    <2-4 specific angles to investigate>
Skip:          <what NOT to search — reduces noise>

This plan is reviewable — the user can correct it before you spend time searching. Proceed to Phase 1 only after outputting the plan.

Step 4: Check existing knowledge. Before searching externally, scan what we already have:

  1. Knowledge/Learned/ — glob for related tags/topics. If a card exists from last week, start from there, don't re-research.
  2. Knowledge/Notes/ — prior research on the same or adjacent topic.
  3. MEMORY.md — Key Decisions or Lessons that provide baseline context.

Note what you found in the plan: Prior: <file or "none">. This prevents duplicate research and gives Phase 1 a head start.

Step 5: Bilingual search planning. The user is bilingual (EN/CN). Decide search language:

Condition Action
Topic is a Chinese company, person, or product Search CN first, EN second
Topic is global tech Search EN first, add CN search if relevant China angle
User asked in Chinese Include CN searches
person_org intent + Chinese name Search both 张三 and San Zhang
competitive with CN competitors Both languages, compare coverage gaps

Add Languages: EN or Languages: EN+CN to the plan output.

Fast-path rule: If intent is factual or how_to AND the query is specific enough to search directly (not ambiguous, not multi-faceted), compress Phase 0 to a single inline line and proceed immediately:

Intent: factual | Depth: basic | Sources: docs→papers→blogs | Rounds: 2-3

Don't build a full plan block for "How does Raft consensus work?" — just classify and go. Save the full plan for landscape, competitive, trend, person_org, and deep_technical where strategy actually matters.


Phase 1: Broad Exploration

Goal: Map the topic landscape and identify key dimensions before going deep. Use the search parameters from Phase 0 — don't override them.

Tavily parameter mapping: When using Tavily search, pass Phase 0 parameters directly:

  • search_depth → Tavily search_depth ("basic" or "advanced")
  • time_range → Tavily days (day=1, week=7, month=30, year=365)
  • topic → Tavily topic ("general" or "news")
  1. Identify 3-5 major dimensions/subtopics of the research question
  2. Run 3-5 broad searches to understand the landscape:
    • Use specific, contextual queries (not single keywords)
    • Check the current date before forming temporal queries
    • For same-day information, include month + day + year in searches
  3. Record what you find AND what's missing -- gaps drive Phase 2
Search Strategy Example
Broad landscape "{topic}" overview 2026
Key players "{topic}" companies OR organizations
Recent developments "{topic}" March 2026
Controversy/debate "{topic}" criticism OR challenges
Expert voices "{topic}" expert opinion OR analysis

Output: A mental map of the topic with identified subtopics and gaps.

Phase 2: Deep Dive

Goal: Conduct targeted research on each important subtopic identified in Phase 1.

For each subtopic:

  1. Run 2-3 targeted searches with specific, refined queries
  2. Use WebFetch to read full content from the most authoritative sources
  3. Extract: facts, data points, quotes, methodology details
  4. Note source quality (official docs > peer-reviewed > news > blogs > social)

Source Priority Ranking:

Tier Source Type Trust Level
1 Official documentation, primary sources, peer-reviewed papers High
2 Established news outlets, recognized industry analysts Medium-High
3 Technical blogs from known experts, conference talks Medium
4 Community discussions (HN, Reddit, StackOverflow) Medium-Low
5 Social media, anonymous posts, undated content Low

Phase 3: Diversity & Validation

Goal: Ensure comprehensive coverage and cross-validate key claims.

For each major finding, verify you can confidently answer:

Dimension Question
Facts What are the established, verifiable facts?
Examples What real-world cases illustrate the points?
Expert views What do recognized authorities say?
Trends What direction is this heading?
Comparisons How does this relate to alternatives/competitors?
Challenges What are the known problems, criticisms, or risks?

Validation rules:

  • Key claims need 2+ independent sources
  • If only one source exists, note it explicitly
  • Conflicting information gets both sides presented with assessment
  • Never suppress inconvenient findings

Phase 4: Synthesis Check

Goal: Verify research is comprehensive BEFORE writing anything.

Ask yourself these questions. If any answer is "no", go back to Phase 2 or 3:

  • Can I state the key facts without hedging?
  • Do I have real-world examples (not hypothetical)?
  • Can I cite at least one expert perspective?
  • Do I know the current trends and trajectory?
  • Have I identified the main challenges/criticisms?
  • Could I explain the relevance to the user's context?

If you cannot check all boxes, do NOT proceed to writing. Return to earlier phases and fill the gaps.


Writing the Research Document

Only after Phase 4 is complete:

Document Structure

# {Research Topic}

**Date:** YYYY-MM-DD
**Research depth:** {number of sources consulted}

## Summary
{3-5 sentences covering the key findings -- self-contained for skimmers}

## Key Findings

### {Finding 1}
{Analysis with inline citations}

### {Finding 2}
{Analysis with inline citations}

...

## Open Questions
{What remains unclear or under-researched}

## Sources
1. {Title} -- {URL} -- {accessed date}
2. ...

Writing Rules

  • Lead with insights, not background
  • Every factual claim needs a citation
  • Distinguish clearly between facts, analysis, and speculation
  • Use tables for comparisons, not prose
  • Keep paragraphs focused -- one idea each
  • Flag uncertainty explicitly ("Source X claims... but this is unverified")

Intent-Specific Search Strategies

Use these as Phase 1 templates based on the Phase 0 intent:

Intent Phase 1 Searches Phase 2 Focus
factual Definition + mechanism + edge cases Official docs, verify claims across sources
competitive Each product separately + "X vs Y" + user reviews Feature matrix, pricing, real user experiences
landscape Category overview + key players + recent entrants + market reports Each player's differentiator, adoption signals
how_to Tutorial search + GitHub repos + SO questions Implementation details, common pitfalls, working examples
breaking_news News search (topic=news, time=day) + social reactions + expert commentary Primary source, timeline of events, impact analysis
trend Historical trajectory + current state + expert predictions Data-backed claims, not opinion-only; adoption metrics
person_org Recent activity + projects + talks/writing + community mentions Cross-platform presence, contribution patterns
deep_technical Source code + architecture docs + design decisions + conference talks Implementation details, trade-offs, performance data

Search Technique Reference

Technique When to Use Example
Quoted phrase Finding exact matches "exact phrase here"
Site-scoped Targeting specific sources site:arxiv.org "{topic}"
Temporal Recent information "{topic}" 2026 or "{topic}" March 2026
Exclusion Filtering noise "{topic}" -pinterest -youtube
Comparison Evaluating alternatives "{topic}" vs "{alternative}"
Technical depth Implementation details "{topic}" implementation OR architecture OR design

Troubleshooting

Problem Solution
Too many surface-level results Add technical terms, use site: for authoritative sources
Topic is too broad Narrow to a specific angle before Phase 2
Conflicting information Note both sides, assess source credibility, present honestly
Very recent topic (<1 week) Include exact dates in searches, check social media (Tier 5), note limited sourcing
User wants speed over depth Do Phase 1 + abbreviated Phase 2, skip Phase 3, note reduced confidence

Verification

Before marking this task complete, show evidence for each:

  • Sources cited with URLs — every factual claim links to a specific source, with access date noted
  • Multi-source validation — key findings backed by 2+ independent sources; single-source claims flagged
  • Findings synthesized — research document follows the template (Summary, Key Findings, Open Questions, Sources)
  • Output saved to Knowledge/ — research file exists at Knowledge/Notes/YYYY-MM-DD-<topic>.md (or Knowledge/Library/ if finalized)
  • Phase 4 checklist passed — all six synthesis-check questions answered affirmatively before writing
Install via CLI
npx skills add https://github.com/xg-gh-25/SwarmAI --skill deep-research
Repository Details
star Stars 266
call_split Forks 32
navigation Branch main
article Path SKILL.md
More from Creator