Academic Deep Research

Academic Deep Research 🔬

When to Use

Trigger this skill when the user wants:

• Deep research, exhaustive analysis, or literature review
• Multi-source verification and evidence hierarchies
• Academic-style reports with citations

Required Stop Points

1. Initial engagement: ask 2–3 clarifying questions and confirm understanding. 2. Research plan: present themes, steps, and deliverables; wait for approval. 3. Final report: deliver full narrative report with citations.

Minimum Requirements

• Two full research cycles per theme.
• Analyze between every tool call.
• Multiple sources per claim; contradictions addressed.
• APA 7th in-text citations and reference list.
• Narrative-only final report (no lists or tables).

References

• reference/protocol.md — research phases, tool order, and analysis rules
• reference/writing-style.md — narrative constraints and phase formatting
• reference/citations-apa.md — citation rules and examples
• reference/report-template.md — required report structure
• reference/error-handling.md — gaps, contradictions, failures
• reference/quality-standards.md — evidence hierarchy and confidence levels
• reference/parallel-research.md — sessions_spawn workflow
• quickref.md — short checklist
• example.md — sample usage

Deep Research Example Workflow

User Request

"I need to understand the current state of AI coding assistants. What's the landscape, who are the players, and what should I know if I'm choosing one for my team?"

---

Phase 1: Initial Engagement

Clarifying Questions: 1. What is the primary question or problem you're trying to solve — are you replacing an existing tool, or is this a first adoption? 2. What depth of analysis do you need — a quick landscape overview, or exhaustive technical and security analysis? 3. Are there specific constraints — budget limits, security requirements, or integration needs with your existing stack?

User Response:

"First adoption, need exhaustive analysis for a 15-person engineering team using VS Code. Security is paramount because we handle healthcare data. Budget is flexible if the value is there."

Reflection:

"So you need a comprehensive, security-focused analysis of AI coding assistants for healthcare software development. The research must prioritize security certifications, data handling practices, and HIPAA compliance considerations alongside technical capabilities. I'll investigate market players, security frameworks, real-world healthcare adoption, and implementation considerations for regulated environments."

---

Phase 2: Research Planning

Major Themes Identified

Research Execution Plan

Theme 1: Market Landscape 1. web_search count=20 "AI coding assistants 2024 2025 market landscape comparison" 2. Analysis — identify players, segments, trends 3. web_fetch on top 5 authoritative sources 4. Synthesis — market structure, competitive dynamics

Theme 2: Security & Compliance 1. web_search count=20 "GitHub Copilot security SOC 2 HIPAA compliance enterprise" 2. Analysis — security frameworks, data handling 3. web_fetch on security whitepapers, compliance docs 4. Synthesis — security landscape, gaps, recommendations

Theme 3: Technical Capabilities 1. web_search count=20 "AI coding assistant benchmarks code completion accuracy 2024" 2. Analysis — feature matrices, performance claims 3. web_fetch on benchmark studies, technical docs 4. Synthesis — capability assessment, differentiators

Theme 4: Healthcare-Specific 1. web_search count=20 "AI coding assistant healthcare HIPAA PHI code security" 2. Analysis — healthcare adoption, regulatory concerns 3. web_fetch on healthcare case studies, compliance guides 4. Synthesis — healthcare readiness, risk factors

Theme 5: Implementation & ROI 1. web_search count=20 "AI coding assistant developer productivity ROI study 2024" 2. Analysis — productivity claims, cost models 3. web_fetch on ROI studies, pricing analysis 4. Synthesis — value proposition, implementation factors

Expected Deliverables

• Comprehensive research paper (20-30 pages)
• APA citations throughout (1-2 per paragraph)
• Security-focused recommendations for healthcare
• Side-by-side comparison matrix (converted to narrative)
• Implementation roadmap

---

Phase 3: Research Cycles (Auto-Execute)

Theme 1: Market Landscape — Cycle 1

Tool: web_search query="AI coding assistants 2024 2025 market landscape comparison GitHub Copilot Cursor" count=20

Results: Found sources on GitHub Copilot, Cursor, Codeium, Amazon CodeWhisperer, Google Code Assist, Tabnine, Replit Ghostwriter.

Analysis: The search reveals a market dominated by GitHub Copilot with approximately 1.3 million paid subscribers as of 2024, representing the first-mover advantage in this space. Cursor has emerged as a significant challenger, having raised $60 million in Series A funding and positioning itself around larger context windows and privacy-focused features. Codeium operates on a freemium model with strong uptake among individual developers, while enterprise-focused players like Amazon CodeWhisperer and Google Code Assist are targeting organizational adoption through cloud integration.

Initial hypotheses forming: (1) Market is consolidating around a few major players, (2) Differentiation is shifting from basic completion to context understanding and privacy, (3) Enterprise features are becoming the primary battleground.

Gaps Identified:

• Limited independent market share data — most figures come from vendor announcements
• Unclear on international market distribution
• Need deeper analysis on enterprise vs. individual adoption rates
• Contradiction: Some sources claim Cursor is gaining rapidly; others show Copilot maintaining dominance

---

Theme 1: Market Landscape — Cycle 2

Tool: web_fetch on GitHub Copilot official documentation, Cursor about page, and TechCrunch funding coverage.

Analysis: Fetching primary sources reveals important nuances that challenge initial assumptions. While GitHub's announced 1.3 million subscribers represents paid users, the actual developer reach is larger through free educational licenses and open source programs. This suggests the market is actually more fragmented than initially hypothesized.

Cursor's $60 million raise is confirmed through TechCrunch reporting, but the company's actual user numbers remain private. This creates uncertainty about their true market position — they may be growing rapidly in specific segments (startups, privacy-conscious developers) without threatening Copilot's overall dominance.

The contradiction noted in Cycle 1 appears resolvable: Cursor is gaining mindshare and venture attention, but Copilot maintains volume leadership. These are different metrics of "success" that can coexist.

New patterns emerging: (1) Market segmentation by use case (individual vs. enterprise), (2) Geographic variation in adoption, (3) Pricing model experimentation (per-seat vs. usage-based).

Connections to Cycle 1: The initial hypothesis about market consolidation requires refinement. Rather than a winner-take-all dynamic, the evidence suggests parallel market development: Copilot for broad adoption, Cursor for specific segments, and specialized tools for niche use cases.

Remaining Uncertainties:

• Actual Cursor user numbers remain undisclosed
• Enterprise adoption rates poorly documented
• International market data largely absent

---

Theme 2: Security & Compliance — Cycle 1

Tool: web_search query="GitHub Copilot security SOC 2 HIPAA compliance enterprise data handling" count=20

Results: Found GitHub security whitepaper, SOC 2 reports, enterprise trust documentation, and some healthcare-specific discussions.

Analysis: Security documentation reveals significant variation in compliance posture across vendors. GitHub Copilot Business and Enterprise tiers explicitly address security concerns with SOC 2 Type II certification and options for code isolation that prevent training data inclusion. However, HIPAA compliance remains ambiguous — GitHub states they will sign Business Associate Agreements but stops short of claiming HIPAA compliance for the AI features themselves.

Cursor positions itself as privacy-first with a local mode that processes code entirely on-device, eliminating transmission risks. This represents a fundamentally different security model that may be more appropriate for healthcare contexts.

Initial hypothesis: Security features correlate with pricing tier, with enterprise offerings providing necessary controls for regulated industries.

Gaps Identified:

• No clear HIPAA compliance claims from any vendor for AI features specifically
• Limited independent security audits published
• Unclear on data retention policies across vendors
• Contradiction: Some sources suggest on-premise options exist; others indicate cloud-only

---

Theme 2: Security & Compliance — Cycle 2

Tool: web_fetch on GitHub Copilot Enterprise security documentation, Cursor security page, and healthcare IT forum discussions.

Analysis: Deep extraction from primary sources confirms the security landscape is complex and evolving. GitHub's documentation explicitly states that Copilot Enterprise offers "code isolation" that prevents user code from being used to train public models, addressing a major concern for proprietary codebases. However, the documentation carefully avoids HIPAA compliance claims, instead referring to GitHub's general BAA availability.

This is a critical finding: The AI coding assistant products themselves may not be HIPAA-compliant even when the underlying platform is. This distinction is crucial for healthcare organizations.

Cursor's local mode is confirmed to process code entirely on-device using locally-hosted models, representing a genuinely different security posture. For healthcare, this eliminates the data transmission risk but may introduce other concerns (local model quality, update mechanisms).

The contradiction about on-premise options is partially resolved: No major vendor offers true on-premise AI coding assistants as of this research, though some offer VPC deployment or local processing modes.

Connections to Theme 1: The market segmentation noted earlier aligns with security postures: Copilot for general enterprise, Cursor for security-conscious segments. This suggests healthcare organizations may need to evaluate Cursor more seriously despite its smaller market presence.

---

Theme 3: Technical Capabilities — Cycle 1

[Similar two-cycle process with analysis between each tool call...]

---

Theme 4: Healthcare-Specific Considerations — Cycle 1

[Similar two-cycle process...]

---

Theme 5: Implementation & ROI — Cycle 1

[Similar two-cycle process...]

---

Cross-Theme Knowledge Integration

After completing all five themes with two cycles each, synthesis reveals:

Emerging Pattern 1: The Healthcare Gap No AI coding assistant currently claims HIPAA compliance for AI features specifically. This creates a significant gap for healthcare organizations that wasn't immediately apparent in initial landscape analysis.

Emerging Pattern 2: The Security-Performance Tradeoff Local processing (Cursor) offers superior privacy but may sacrifice some capability compared to cloud-powered alternatives (Copilot). This tradeoff is central to the healthcare decision.

Emerging Pattern 3: Market Immaturity The field is evolving rapidly with new features and compliance developments monthly. Any recommendation must account for this volatility.

Contradictions Resolved:

• Market dominance vs. challenger growth: Different metrics, both valid
• Cloud vs. on-premise: No true on-premise exists; local processing is the alternative

Unified Understanding: For a healthcare engineering team, the decision framework differs from general enterprise adoption. Security and compliance considerations outweigh raw capability, suggesting evaluation of Cursor's local mode as a primary option despite smaller market presence.

---

Phase 4: Final Report

[Presented as cohesive research paper with narrative sections, proper APA citations, no bullet points, 6-8+ paragraphs per major section...]

---

Key Distinctions from Standard Research

Deep Research Quick Reference

Invocation

• /research or mention "deep research" / "exhaustive analysis"

Four Phases

Stop Points (Only Three)

1. ✅ After clarifying questions (Phase 1) 2. ✅ After research plan presentation (Phase 2) 3. ✅ Final report delivery (Phase 4)

Tool Usage Sequence (Per Theme)

1. START: web_search for landscape (count=20) 2. ANALYZE: Synthesize findings, identify patterns and gaps 3. DIVE: web_fetch for depth on key sources 4. PROCESS: Synthesize new findings, challenge assumptions 5. REPEAT: Second cycle targeting identified gaps

Required Analysis After Every Tool Use

• Connect new findings to previous results
• Show evolution of understanding
• Highlight pattern changes
• Address contradictions
• Build coherent narrative

Research Standards

• Every conclusion cites multiple sources
• All contradictions addressed
• Uncertainties acknowledged
• Limitations discussed
• Gaps identified

Writing Style (Final Report)

• Flowing narrative — paragraphs only, no lists
• Academic but accessible
• Evidence integrated naturally in prose
• Progressive logical development
• Smooth transitions between concepts

Prohibited in Final Report

• Bullet points or numbered lists
• Tables (convert to prose)
• Isolated data without context
• Section headers without narrative

Citation Standards (APA 7th)

• Density: 1-2 citations per paragraph
• Format: (Author, Year) in-text
• References: Full APA with hanging indent
• All claims cited — no exceptions

Confidence Annotations

• [HIGH] — Multiple high-quality sources agree
• [MEDIUM] — Limited or mixed evidence
• [LOW] — Single source, needs verification
• [SPECULATIVE] — Emerging area

Report Sections (Narrative Format)

1. Executive Summary — 2-3 paragraphs 2. Knowledge Development — evolution of understanding (6-8+ paragraphs) 3. Comprehensive Analysis — findings, patterns, contradictions, evidence (6-8+ paragraphs each subsection) 4. Practical Implications — applications, risks, future research (6-8+ paragraphs each subsection) 5. References — APA format, alphabetical 6. Appendices — optional

Critical Reminders

• Stop only at three major points
• Always analyze between tool usage
• Show clear thinking progression
• Connect findings explicitly
• Build coherent narrative throughout
• No shortcuts or rushed analysis

Academic Deep Research 🔬

Transparent, rigorous, self-contained research — not a black-box API wrapper.

Why This Skill Exists

Most "deep research" tools are wrappers around external APIs. You send a query, get a report, and have no idea what happened in between.

This skill is different:

• ✅ Full methodology visible — Every step documented, reproducible
• ✅ No external dependencies — Runs entirely on OpenClaw native tools
• ✅ User control — 3 explicit checkpoints for approval
• ✅ Academic rigor — APA citations, evidence hierarchy, confidence levels
• ✅ Works offline — No API keys, no cloud services

Comparison with Cloud-Based Research Tools

Core Features

Mandated Research Cycles

Every theme gets minimum 2 full research cycles: 1. Broad landscape search → Analysis → Gap identification 2. Targeted deep dive → Challenge assumptions → Synthesis

No shortcuts. No single-pass summaries.

Evidence Standards

• Every conclusion cites multiple sources
• Contradictions must be addressed — not hidden
• Confidence annotations: [HIGH], [MEDIUM], [LOW], [SPECULATIVE]
• Evidence hierarchy: Meta-analyses > RCTs > Observational > Expert opinion

Academic Output

• Flowing narrative prose (no bullet point dumps)
• APA 7th edition citations (1-2 per paragraph)
• Proper paragraph structure: claim → evidence → analysis → transition
• Executive summary, methodology, findings, limitations, references

User Control

Three mandatory stop points: 1. Initial Engagement — Clarify scope before research 2. Research Planning — Approve themes and approach 3. Final Report — Review completed analysis

Quick Start

/research "Comprehensive analysis of [your topic]"

Or just ask for "deep research on..." or "exhaustive analysis of..."

Research Protocol

Phase 1: Clarification

Agent asks 2-3 essential questions, confirms understanding, waits for you.

Phase 2: Planning

Agent presents:

• Major themes identified (3-5)
• Research execution plan (table format)
• Expected deliverables

You approve before execution begins.

Phase 3: Execution (Auto)

For each theme, two full cycles:

• web_search (count=20) for landscape
• Analysis and gap identification
• web_fetch on primary sources
• Synthesis and assumption challenging
• Repeat for depth

Required: Explicit analysis between every tool call showing evolution of understanding.

Phase 4: Report

Academic narrative with:

• Executive Summary
• Knowledge Development
• Comprehensive Analysis
• Practical Implications
• APA References

File Structure

deep-research/
├── SKILL.md       # Full methodology (500+ lines)
├── README.md      # This file
├── quickref.md    # One-page cheat sheet
├── example.md     # Complete workflow example
└── LICENSE        # Apache 2.0

When to Use This

• Literature reviews requiring academic rigor
• Competitive intelligence with source verification
• Complex topics needing multi-source synthesis
• Any research where you need to show your work
• When you don't trust black-box AI summaries

License

Apache 2.0 — See LICENSE

---

Built for researchers who care about methodology, not just outputs.

APA 7th Citations

In-Text Examples

Recent research shows significant effects (Johnson et al., 2023).
Multiple meta-analyses confirm this pattern (Smith, 2020; Garcia et al., 2022).

Reference Format

Garcia, J., Martinez, A., & Lee, S. (2022). Article title. Journal Name, 15(3), 245–267. https://doi.org/10.xxxx/jes.2022.15.3.245

Citation Rules

• 1–2 citations per paragraph.
• Use “et al.” for 3+ authors in-text.
• Every in-text citation must appear in references.
• Alphabetize references by first author surname.
• If missing data: use (Source Name, n.d.) with URL.

Error Handling

Insufficient Search Results

1. Broaden query terms. 2. Search adjacent concepts. 3. Document the gap. 4. Lower confidence appropriately.

Unresolved Contradictions

1. Present both claims. 2. Analyze methodological differences. 3. Assess evidence quality. 4. Document unresolved status if needed.

Source Quality Concerns

• Prefer primary sources.
• Flag outdated or biased sources.
• Note unclear methodology.

Technical Failures

• If web_fetch fails, document the URL and proceed.
• On rate limits, slow down and retry with smaller counts.
• If memory search fails, proceed and note limitation.

Parallel Research with sessions_spawn

Use parallel sub-agents when themes are independent.

When to Use

• Themes do not depend on each other.
• Time constraints require speed.
• Token budget allows parallel work.

Workflow

1. Spawn sub-agents with explicit instructions for two cycles. 2. Each sub-agent returns findings, confidence, gaps, and sources. 3. Integrate results across themes.

Example

Theme A: Market landscape
→ sessions_spawn(task="Research market landscape with 2 cycles... return findings, gaps, sources")

Theme B: Security
→ sessions_spawn(task="Research security/compliance with 2 cycles... return findings, gaps, sources")

Research Protocol

Phase 1: Initial Engagement (STOP POINT)

Before any research begins:

• Ask 2–3 essential clarifying questions.
• Reflect understanding back to the user.
• Wait for response before proceeding.

Phase 2: Research Planning (STOP POINT)

Present the complete plan to the user:

• Major themes (3–5) with key questions and approach.
• Execution plan with steps, tools, and outputs.
• Expected deliverables, format, and length.

Wait for explicit approval before Phase 3.

Phase 3: Mandated Research Cycles (NO STOPS)

For each theme, complete two cycles.

Cycle 1: Landscape Analysis

1. web_search broad query (count=20). 2. Synthesize patterns and initial hypotheses. 3. Identify gaps and contradictions.

Cycle 2: Deep Investigation

1. Targeted web_search on gaps. 2. web_fetch primary sources. 3. Re-evaluate assumptions; integrate new evidence.

Required Analysis Between Tool Calls

After each tool call, explicitly:

• Connect new findings to prior results.
• Show how understanding evolved.
• Highlight pattern changes.
• Address contradictions and evidence quality.
• Maintain a coherent narrative thread.

Tool Usage Sequence

Per theme: 1. Start with web_search for landscape. 2. Analyze and identify gaps. 3. web_fetch primary sources. 4. Analyze, then repeat with targeted searches.

Cross-Theme Integration

After all theme cycles:

• Identify shared conclusions.
• Note cross-theme contradictions.
• Map relationships across findings.
• Form a unified understanding.

Evidence and Quality Standards

Evidence Hierarchy

1. Systematic reviews & meta-analyses 2. Randomized controlled trials 3. Cohort / longitudinal studies 4. Expert consensus / guidelines 5. Cross-sectional / observational studies 6. Expert opinion / editorials 7. Media reports / blogs

Confidence Annotations

• [HIGH] Multiple high-quality sources agree
• [MEDIUM] Limited or mixed evidence
• [LOW] Single source or preliminary
• [SPECULATIVE] Hypothesis or emerging area

Red Flags

• Claims without cited sources
• Single-study findings presented as fact
• Conflicts of interest not disclosed
• Outdated information
• Cherry-picked statistics
• Overgeneralization

Final Report Template

# Research Report: [Topic]

## Executive Summary
Two to three substantial paragraphs capturing the research question, primary findings, and confidence.

## Knowledge Development
Narrative of how understanding evolved through the research cycles, including shifts and uncertainties.

## Comprehensive Analysis
Narrative covering primary findings, patterns, contradictions, and evidence strength.

## Practical Implications
Narrative on applications, risks, implementation considerations, and future research directions.

## References
Full APA-formatted reference list in alphabetical order.

## Appendices (optional)
Search strategy, source reliability assessment, excluded sources, research timeline.

Writing Style Requirements

Narrative Style (Final Report)

• Flowing narrative prose only.
• Academic but accessible.
• Integrated evidence and citations.
• Clear logical progression and transitions.

Structured Data Rules

Prohibited in Final Report

• Bullet points or numbered lists.
• Data tables.
• Isolated data points without narrative context.

Required in Final Report

• Topic sentences for each paragraph.
• Evidence with citations.
• Analysis and implications.
• Transitions between paragraphs.

Paragraph Pattern

1. Topic sentence 2. Evidence (citations) 3. Analysis/implications 4. Transition