1# Deep Research Methodology: 8-Phase Pipeline
2 
3## Overview
4 
5This document contains the detailed methodology for conducting deep research. The 8 phases represent a comprehensive approach to gathering, verifying, and synthesizing information from multiple sources.
6 
7---
8 
9## Phase 1: SCOPE - Research Framing
10 
11**Objective:** Define research boundaries and success criteria
12 
13**Activities:**
141. Decompose the question into core components
152. Identify stakeholder perspectives
163. Define scope boundaries (what's in/out)
174. Establish success criteria
185. List key assumptions to validate
19 
20**Ultrathink Application:** Use extended reasoning to explore multiple framings of the question before committing to scope.
21 
22**Output:** Structured scope document with research boundaries
23 
24---
25 
26## Phase 2: PLAN - Strategy Formulation
27 
28**Objective:** Create an intelligent research roadmap
29 
30**Activities:**
311. Identify primary and secondary sources
322. Map knowledge dependencies (what must be understood first)
333. Create search query strategy with variants
344. Plan triangulation approach
355. Estimate time/effort per phase
366. Define quality gates
37 
38**Graph-of-Thoughts:** Branch into multiple potential research paths, then converge on optimal strategy.
39 
40**Output:** Research plan with prioritized investigation paths
41 
42---
43 
44## Phase 3: RETRIEVE - Parallel Information Gathering
45 
46**Objective:** Systematically collect information from multiple sources using parallel execution for maximum speed
47 
48**CRITICAL: Execute ALL searches in parallel using a single message with multiple tool calls**
49 
50### Query Decomposition Strategy
51 
52Before launching searches, decompose the research question into 5-10 independent search angles:
53 
541. **Core topic (semantic search)** - Meaning-based exploration of main concept
552. **Technical details (keyword search)** - Specific terms, APIs, implementations
563. **Recent developments (date-filtered)** - What's new in last 12-18 months (use current date from Step 0)
574. **Academic sources (domain-specific)** - Papers, research, formal analysis
585. **Alternative perspectives (comparison)** - Competing approaches, criticisms
596. **Statistical/data sources** - Quantitative evidence, metrics, benchmarks
607. **Industry analysis** - Commercial applications, market trends
618. **Critical analysis/limitations** - Known problems, failure modes, edge cases
62 
63### Parallel Execution Protocol
64 
65**Step 0: Get the current date**
66 
67Before ANY searches, retrieve today's date using Bash: `date +%Y-%m-%d`
68Use the returned year for all date-filtered queries and recency checks. Do NOT assume a year from training data.
69 
70**Step 1: Launch ALL searches concurrently (single message)**
71 
72**CRITICAL: Use correct tool and parameters to avoid errors**
73 
74**Primary: search-cli (multi-provider, always use first)**
75- Unified CLI aggregating Brave, Serper, Exa, Jina, and Firecrawl
76- Auto-detects best provider per query type (academic, news, general, people)
77- JSON output for structured processing: `search "query" --json`
78- Modes: general, news, academic, scholar, patents, people, images, extract, scrape
79- Example: `search "quantum computing 2025" -m academic --json -c 15`
80- For page content extraction: `search "URL" -m extract --json`
81- For scraping: `search "URL" -m scrape --json`
82- Run via Bash tool: `search "query" --json -c 10`
83 
84**Fallback: WebSearch (if search-cli fails or is unavailable)**
85- Built-in Claude web search, no setup required
86- Parameters: `query` (required), optional `allowed_domains`, `blocked_domains`
87- Use when: search-cli returns errors, rate-limited, or for domain-restricted queries
88 
89**Optional: Exa MCP (if configured, for semantic/neural search)**
90- Tool name: `mcp__Exa__exa_search`
91- Use for semantic exploration alongside search-cli keyword results
92 
93 
94**NEVER mix parameter styles** - this causes "Invalid tool parameters" errors.
95 
96**Step 2: Spawn parallel deep-dive agents**
97 
98Use Task tool with general-purpose agents (3-5 agents) for:
99- Academic paper analysis (PDFs, detailed extraction)
100- Documentation deep dives (technical specs, API docs)
101- Repository analysis (code examples, implementations)
102- Specialized domain research (requires multi-step investigation)
103 
104**Sub-agent output format:** Require all sub-agents to return structured evidence, not free text:
105```json
106{"claim": "specific claim text", "evidence_quote": "exact quote from source", "source_url": "https://...", "source_title": "...", "confidence": 0.85}
107```
108This prevents synthesis fatigue when merging results from 3-5 agents.
109 
110**Evidence persistence (v3.0):** After each retrieval batch, persist evidence immediately:
111```bash
112# Register the source first (returns stable source_id)
113python scripts/citation_manager.py register-source --json '{"raw_url": "...", "title": "..."}' --dir [folder]
114 
115# Then persist each evidence span from that source
116python scripts/evidence_store.py add --json '{"source_id": "...", "quote": "exact text", "evidence_type": "direct_quote", "locator": "page 5"}' --dir [folder]
117```
118Evidence must not live only in model context — it must be persisted to `evidence.jsonl` before synthesis begins. This ensures continuation agents and claim-support verification can access the full evidence trail.
119 
120**Example parallel execution (using search-cli via Bash):**
121```
122[Single message with multiple Bash tool calls]
123- Bash: search "quantum computing 2026 state of the art" --json -c 10
124- Bash: search "quantum computing limitations challenges" --json -c 10
125- Bash: search "quantum computing commercial applications 2026" -m news --json -c 10
126- Bash: search "quantum computing vs classical comparison" --json -c 10
127- Bash: search "quantum error correction research" -m academic --json -c 10
128- Task(subagent_type="general-purpose", description="Analyze quantum computing papers", prompt="Deep dive into quantum computing academic papers from [CURRENT_YEAR], extract key findings and methodologies")
129- Task(subagent_type="general-purpose", description="Industry analysis", prompt="Analyze quantum computing industry reports and market data, identify commercial applications")
130- Task(subagent_type="general-purpose", description="Technical challenges", prompt="Extract technical limitations and challenges from quantum computing research")
131```
132 
133**Example parallel execution (using Exa MCP - if available):**
134```
135[Single message with multiple tool calls]
136- mcp__Exa__exa_search(query="quantum computing state of the art", type="neural", num_results=10, start_published_date="[use current year from Step 0]")
137- mcp__Exa__exa_search(query="quantum computing limitations", type="keyword", num_results=10)
138- mcp__Exa__exa_search(query="quantum computing commercial", type="auto", num_results=10, start_published_date="[use current year from Step 0]")
139- mcp__Exa__exa_search(query="quantum error correction", type="neural", num_results=10, include_domains=["arxiv.org"])
140- Task(subagent_type="general-purpose", description="Academic analysis", prompt="Analyze quantum computing academic papers")
141```
142 
143**Step 3: Collect and organize results**
144 
145As results arrive:
1461. Extract key passages with source metadata (title, URL, date, credibility)
1472. Track information gaps that emerge
1483. Follow promising tangents with additional targeted searches
1494. Maintain source diversity (mix academic, industry, news, technical docs)
1505. Monitor for quality threshold (see FFS pattern below)
151 
152### First Finish Search (FFS) Pattern
153 
154**Adaptive completion based on quality threshold:**
155 
156**Quality gate:** Proceed to Phase 4 when FIRST threshold reached:
157- **Quick mode:** 10+ sources with avg credibility >60/100 OR 2 minutes elapsed
158- **Standard mode:** 15+ sources with avg credibility >60/100 OR 5 minutes elapsed
159- **Deep mode:** 25+ sources with avg credibility >70/100 OR 10 minutes elapsed
160- **UltraDeep mode:** 30+ sources with avg credibility >75/100 OR 15 minutes elapsed
161 
162**Continue background searches:**
163- If threshold reached early, continue remaining parallel searches in background
164- Additional sources used in Phase 5 (SYNTHESIZE) for depth and diversity
165- Allows fast progression without sacrificing thoroughness
166 
167### Quality Standards
168 
169**Source diversity requirements:**
170- Minimum 3 source types (academic, industry, news, technical docs)
171- Temporal diversity (mix of recent 12-18 months + foundational older sources)
172- Perspective diversity (proponents + critics + neutral analysis)
173- Geographic diversity (not just US sources)
174 
175**Credibility tracking:**
176- Score each source 0-100 using source_evaluator.py
177- Flag low-credibility sources (<40) for additional verification
178- Prioritize high-credibility sources (>80) for core claims
179 
180**Techniques:**
181- Use search-cli for all searches (primary tool, multi-provider)
182- Fall back to WebSearch if search-cli fails or is rate-limited
183- Use WebFetch for deep dives into specific sources (secondary)
184- Use Exa search (via WebSearch with type="neural") for semantic exploration
185- Use Grep/Read for local documentation
186- Execute code for computational analysis (when needed)
187- Use Task tool to spawn parallel retrieval agents (3-5 agents)
188 
189**Output:** Organized information repository with source tracking, credibility scores, and coverage map
190 
191---
192 
193## Phase 4: TRIANGULATE - Cross-Reference Verification
194 
195**Objective:** Validate information across multiple independent sources
196 
197**Activities:**
1981. Identify claims requiring verification
1992. Cross-reference facts across 3+ sources
2003. Flag contradictions or uncertainties
2014. Assess source credibility
2025. Note consensus vs. debate areas
2036. Document verification status per claim
204 
205**Quality Standards:**
206- Core claims must have 3+ independent sources
207- Flag any single-source information
208- Note recency of information
209- Identify potential biases
210 
211**Output:** Verified fact base with confidence levels
212 
213---
214 
215## Phase 4.5: OUTLINE REFINEMENT - Dynamic Evolution (WebWeaver 2025)
216 
217**Objective:** Adapt research direction based on evidence discovered
218 
219**Problem Solved:** Prevents "locked-in" research when evidence points to different conclusions or uncovers more important angles than initially planned.
220 
221**When to Execute:**
222- **Standard/Deep/UltraDeep modes only** (Quick mode skips this)
223- After Phase 4 (TRIANGULATE) completes
224- Before Phase 5 (SYNTHESIZE)
225 
226**Activities:**
227 
2281. **Review Initial Scope vs. Actual Findings**
229   - Compare Phase 1 scope with Phase 3-4 discoveries
230   - Identify unexpected patterns or contradictions
231   - Note underexplored angles that emerged as critical
232   - Flag overexplored areas that proved less important
233 
2342. **Evaluate Outline Adaptation Need**
235 
236   **Signals for adaptation (ANY triggers refinement):**
237   - Major findings contradict initial assumptions
238   - Evidence reveals more important angle than originally scoped
239   - Critical subtopic emerged that wasn't in original plan
240   - Original research question was too broad/narrow based on evidence
241   - Sources consistently discuss aspects not in initial outline
242 
243   **Signals to keep current outline:**
244   - Evidence aligns with initial scope
245   - All key angles adequately covered
246   - No major gaps or surprises
247 
2483. **Refine Outline (if needed)**
249 
250   **Update structure to reflect evidence:**
251   - Add sections for unexpected but important findings
252   - Demote/remove sections with insufficient evidence
253   - Reorder sections based on evidence strength and importance
254   - Adjust scope boundaries based on what's actually discoverable
255 
256   **Example adaptation:**
257   ```
258   Original outline:
259   1. Introduction
260   2. Technical Architecture
261   3. Performance Benchmarks
262   4. Conclusion
263 
264   Refined after Phase 4 (evidence revealed security as critical):
265   1. Introduction
266   2. Technical Architecture
267   3. **Security Vulnerabilities (NEW - major finding)**
268   4. Performance Benchmarks (demoted - less critical than expected)
269   5. **Real-World Failure Modes (NEW - pattern emerged)**
270   6. Synthesis & Recommendations
271   ```
272 
2734. **Targeted Gap Filling (if major gaps found)**
274 
275   If outline refinement reveals critical knowledge gaps:
276   - Launch 2-3 targeted searches for newly identified angles
277   - Quick retrieval only (don't restart full Phase 3)
278   - Time-box to 2-5 minutes
279   - Update triangulation for new evidence only
280 
2815. **Document Adaptation Rationale**
282 
283   Record in methodology appendix:
284   - What changed in outline
285   - Why it changed (evidence-driven reasons)
286   - What additional research was conducted (if any)
287 
288**Quality Standards:**
289- Adaptation must be evidence-driven (cite specific sources that prompted change)
290- No more than 50% outline restructuring (if more needed, scope was severely mis scoped)
291- Retain original research question core (don't drift into different topic entirely)
292- New sections must have supporting evidence already gathered
293 
294**Output:** Refined outline that accurately reflects evidence landscape, ready for synthesis
295 
296**Anti-Pattern Warning:**
297- ❌ DON'T adapt outline based on speculation or "what would be interesting"
298- ❌ DON'T add sections without supporting evidence already in hand
299- ❌ DON'T completely abandon original research question
300- ✅ DO adapt when evidence clearly indicates better structure
301- ✅ DO document rationale for changes
302- ✅ DO stay within original topic scope
303 
304---
305 
306## Phase 5: SYNTHESIZE - Deep Analysis
307 
308**Objective:** Connect insights and generate novel understanding
309 
310**Activities:**
3111. Identify patterns across sources
3122. Map relationships between concepts
3133. Generate insights beyond source material
3144. Create conceptual frameworks
3155. Build argument structures
3166. Develop evidence hierarchies
317 
318**Ultrathink Integration:** Use extended reasoning to explore non-obvious connections and second-order implications.
319 
320**Output:** Synthesized understanding with insight generation
321 
322---
323 
324## Phase 6: CRITIQUE - Quality Assurance
325 
326**Objective:** Rigorously evaluate research quality
327 
328**Activities:**
3291. Review for logical consistency
3302. Check citation completeness
3313. Identify gaps or weaknesses
3324. Assess balance and objectivity
3335. Verify claims against sources
3346. Test alternative interpretations
335 
336**Red Team Questions:**
337- What's missing?
338- What could be wrong?
339- What alternative explanations exist?
340- What biases might be present?
341- What counterfactuals should be considered?
342 
343**Persona-Based Critique (Deep/UltraDeep only):**
344Simulate 2-3 specific critic personas relevant to the topic:
345- "Skeptical Practitioner" — Would someone doing this daily trust these findings?
346- "Adversarial Reviewer" — What would a peer reviewer reject?
347- "Implementation Engineer" — Can these recommendations actually be executed?
348 
349**Critical Gap Loop-Back:**
350If critique identifies a critical knowledge gap (not just a writing issue), return to Phase 3 with targeted "delta-queries" before proceeding to Phase 7. Time-box to 3-5 minutes. This prevents publishing reports with known blind spots.
351 
352**Output:** Critique report with improvement recommendations
353 
354---
355 
356## Phase 7: REFINE - Iterative Improvement
357 
358**Objective:** Address gaps and strengthen weak areas
359 
360**Activities:**
3611. Conduct additional research for gaps
3622. Strengthen weak arguments
3633. Add missing perspectives
3644. Resolve contradictions
3655. Enhance clarity
3666. Verify revised content
367 
368**Output:** Strengthened research with addressed deficiencies
369 
370---
371 
372## Phase 8: PACKAGE - Report Generation
373 
374**Objective:** Deliver professional, actionable research
375 
376**Activities:**
3771. Structure report with clear hierarchy
3782. Write executive summary
3793. Develop detailed sections
3804. Create visualizations (tables, diagrams)
3815. Compile full bibliography
3826. Add methodology appendix
383 
384**Output:** Complete research report ready for use
385 
386---
387 
388## Advanced Features
389 
390### Graph-of-Thoughts Reasoning
391 
392Rather than linear thinking, branch into multiple reasoning paths:
393- Explore alternative framings in parallel
394- Pursue tangential leads that might be relevant
395- Merge insights from different branches
396- Backtrack and revise as new information emerges
397 
398### Parallel Agent Deployment
399 
400Use Task tool to spawn sub-agents for:
401- Parallel source retrieval
402- Independent verification paths
403- Competing hypothesis evaluation
404- Specialized domain analysis
405 
406### Adaptive Depth Control
407 
408Automatically adjust research depth based on:
409- Information complexity
410- Source availability
411- Time constraints
412- Confidence levels
413 
414### Citation Intelligence
415 
416Smart citation management:
417- Track provenance of every claim
418- Link to original sources
419- Assess source credibility
420- Handle conflicting sources
421- Generate proper bibliographies
422