Deep Research

Deep Research: Multi-Domain Research Framework

Execute this skill to perform comprehensive research on any topic using domain-aware scheduling and parallel information gathering.

Execution Instructions

When invoked, you will:

0. Resolve scope and context — invoke context skill (always), then preflight if confidence is low
1. Discover available research tools using ToolSearch
2. Analyze the research intent from working_scope to extract:
   • Primary research topic/question
   • Confirmed domains (from working_scope)
   • Research depth and scope indicators
   • Research methodology requirements (web search, crawling, interactive)

Dependency Check

Before executing any step, verify all required skills are present:

[skills-root]/context/SKILL.md
[skills-root]/preflight/SKILL.md
[skills-root]/domain-registry/README.md

Where [skills-root] is the parent of this skill's directory. Resolve with ls ../ from this skill's location.

If any required file is missing → stop immediately and output:

⚠ Missing required skills for deep-research:

  {missing-skill}
    Expected: {skills-root}/{missing-skill}/SKILL.md

Install the missing skill(s):
  git clone https://github.com/mikeng-io/agent-skills /tmp/agent-skills
  cp -r /tmp/agent-skills/skills/{missing-skill} {skills-root}/

Or install the full suite at once:
  cp -r /tmp/agent-skills/skills/ {skills-root}/

All dependencies present → proceed to Step 0.

Step 0: Scope & Context Resolution

Context (always required):

Invoke Skill("context") first. It classifies the research topic, detects relevant domains, and assesses confidence:

context_report:
  artifact_type: ""  # code | financial | marketing | creative | research | mixed
  domains: []        # matched domain names from domain-registry
  routing: ""        # routing recommendation (informational for research)
  confidence: ""     # high | medium | low

Preflight (conditional — triggered by context confidence):

Invoke Skill("preflight") only if context_report.confidence == "low" OR one or more signals remain unresolved:

• The research topic or question is not clearly stated
• Research intent is ambiguous (background vs. deep dive vs. comparison?)
• Domains cannot be inferred from available context

Preflight fills exactly the gaps context could not resolve (max 3 questions, one at a time):

scope_clarification:
  artifact: ""       # primary research topic or question
  intent: "research"
  domains: []        # supplements context_report.domains
  constraints: []    # explicit focus areas, excluded areas, depth preferences
  confidence: ""     # high | medium

If context_report.confidence == "high" → skip preflight entirely.

Merge into working scope:

working_scope:
  artifact: ""            # primary research topic or question
  domains: []             # from context_report (authoritative), supplemented by preflight
  constraints: []         # depth, focus, exclusion preferences
  context_summary: ""     # combined description for researcher agent prompts
  research_depth: ""      # BRIEF | STANDARD | COMPREHENSIVE (from constraints or defaults)

Use working_scope throughout this skill. Steps 1+ use it instead of re-extracting from conversation.

2. Create a research plan with domain-aware effort allocation:

   • Primary domains get 70% of research effort
   • Secondary domains get 30% of research effort
   • Generate 5-10 search queries per domain
   • Identify sites that may require browser automation
   • Plan crawling strategy for complex sources

3. Execute parallel research using multiple domain-focused agents:

   • Spawn domain researchers in parallel
   • Use web search for general information
   • Use browser automation for interactive sites, paywalls, or dynamic content
   • Use agent-browser skill for complex web interactions (multi-step flows, form submission, navigation) — install only, no MCP config needed
   • Fall back to mcp__playwright__* tools if agent-browser is not installed (requires Playwright MCP configured)
   • Fall back to Bash with curl/wget for simpler static requests
   • Collect and validate findings from each source

4. Synthesize findings across domains:

   • Identify patterns and consensus
   • Surface contradictions and debates
   • Generate evidence-based recommendations

5. Validate output format against schema

6. Generate and save report to .outputs/research/

Step 0: Discover Available Research Tools

Before beginning research, discover all available MCP tools and skills that can be used for information gathering.

Tool Discovery Process

Use ToolSearch to find available research tools:

# Search for web search tools
ToolSearch: "web search"
  → Returns: brave-search (independent index + news + local), web-search-prime, etc.

# Search for AI-synthesized search (Perplexity)
ToolSearch: "perplexity"
  → Returns: mcp__perplexity__search (AI-grounded answers with citations)
  → If found: use for synthesis-heavy queries or to cross-validate web findings

# Search for browser automation tools (agent-browser preferred, playwright as fallback)
ToolSearch: "browser automation"
  → Returns: agent-browser skill, playwright navigation, screenshot, interaction tools

# Search for content extraction tools
ToolSearch: "web reader content"
  → Returns: web-reader, content extraction tools

# Search for documentation query tools
ToolSearch: "documentation query"
  → Returns: context7, zread tools

# Search for DeepWiki codebase intelligence (optional)
ToolSearch: "devin"
  → Returns: mcp__devin__read_wiki_structure, mcp__devin__read_wiki_contents, mcp__devin__ask_question
  → If found: DeepWiki available — use for codebase-grounded research
  → If not found: fall back to local Glob/Grep/Read

Build Tool Inventory

Create an inventory of available tools for the research session:

tool_inventory:
  web_search:
    - mcp__brave-search__brave_web_search      # Brave Search MCP — independent index, no Google bias
    - mcp__brave-search__brave_news_search     # Brave news search — recent/breaking content
    - mcp__brave-search__brave_local_search    # Brave local search — geo-specific results
    - mcp__perplexity__search                  # Perplexity MCP — AI-synthesized answers with citations
    - mcp__web-search-prime__webSearchPrime    # WebSearch Prime — fallback general search

  web_reading:
    - mcp__web-reader__webReader

  browser_automation_primary:
    - agent-browser    # Preferred: skill-based, install only — no MCP config needed

  browser_automation_fallback:
    - mcp__playwright__browser_navigate     # Fallback: requires Playwright MCP configured
    - mcp__playwright__browser_snapshot
    - mcp__playwright__browser_click
    - mcp__playwright__browser_fill_form
    - mcp__playwright__browser_evaluate
    - mcp__browser-tools__takeScreenshot    # Visual capture supplement
    - mcp__browser-tools__getConsoleLogs

  documentation:
    - mcp__context7__query-docs
    - mcp__zread__search_doc

  codebase_intelligence:
    - mcp__devin__ask_question          # DeepWiki: AI-grounded Q&A about indexed GitHub repos
    - mcp__devin__read_wiki_structure   # DeepWiki: list documentation topics for a repo
    - mcp__devin__read_wiki_contents    # DeepWiki: read specific topic documentation
    # Optional — requires Devin API key. Falls back to Glob/Grep/Read if unavailable.
    # See: deepwiki/SKILL.md for setup and usage patterns.

Tool Selection Strategy

Choose tools based on research requirements:

Standard Web Research:

• Use web-search tools for general queries; prefer Brave Search for independent index coverage
• Use Perplexity (mcp__perplexity__search) for synthesis-heavy questions or cross-validation — it returns AI-synthesized answers with citations, useful for "what is the current consensus on X"
• Use web-reader for content extraction from specific URLs
• Multi-search strategy: run Brave Web + Perplexity in parallel, then cross-reference results to surface contradictions
• Fallback gracefully if tools unavailable

Dynamic/Interactive Content:

• Use agent-browser skill (preferred — install only, no MCP config required):
  • Invoke via: Skill("agent-browser")
  • Handles: JavaScript-heavy sites, interactive forms, dynamic content, paywalls
• If agent-browser not installed → fall back to mcp__playwright__* (requires Playwright MCP configured)
• If neither available → fall back to Bash with curl/wget for simpler static requests

Complex Interactions:

• Use agent-browser skill first for:
  • Multi-step workflows (login → navigate → extract)
  • Complex form filling
  • Sites requiring human-like interaction
  • Screenshot capture with analysis
• If agent-browser not installed, use mcp__playwright__* tools (requires MCP config)
• If neither available, use Bash with curl/wget

Tool Availability Handling

If preferred tools unavailable:

1. Log missing tools
2. Use available alternatives in fallback order
3. Adjust research strategy accordingly
4. Note limitations in final report

Fallback chain:

agent-browser → mcp__playwright__* → mcp__browser-tools → web-reader → Bash curl/wget

Step 1: Intent Analysis & Domain Detection

Analyze the conversation to extract research intent and infer domains.

research_intent:
  primary_topic: ""           # Main research question/topic
  explicit_domains: []        # Domains directly mentioned
  inferred_domains: []        # Domains inferred from context
  research_depth: BRIEF | STANDARD | COMPREHENSIVE
  scope_indicators: []        # Keywords suggesting scope (e.g., "overview", "deep dive")

Domain Detection Strategy

Primary Domains (70% effort allocation): Extract from explicit mentions in conversation:

• Topics directly stated by user
• Subject areas named in research questions
• Fields explicitly referenced

Secondary Domains (30% effort allocation): Infer from contextual cues:

• Language patterns (technical vs. business vs. creative)
• Concerns expressed (performance, security, UX)
• Artifacts referenced (code, designs, documents)

Domain-registry integration: Read domain-registry/domains/*.md to supplement inferred domains. Research domains include: technical, financial, marketing, creative, legal, strategy. Select all domains matching the research topic signals.

Domain Inference Examples

User: "Research how to implement event sourcing with Kafka"

Explicit domains:
  - Software Architecture
  - Distributed Systems

Inferred domains:
  - Data Engineering (from "event sourcing")
  - Performance (implied by distributed systems context)

User: "I need to understand the business implications of AI regulation"

Explicit domains:
  - Business
  - Law/Regulation

Inferred domains:
  - Ethics (from "AI regulation")
  - Technology (AI context)

Domain Output Format

domain_analysis:
  primary_domains:
    - domain: "{domain name}"
      confidence: HIGH | MEDIUM | LOW
      rationale: "why detected"
      keywords: ["relevant", "terms"]

  secondary_domains:
    - domain: "{domain name}"
      confidence: HIGH | MEDIUM | LOW
      rationale: "why inferred"
      keywords: ["relevant", "terms"]

Step 2: Research Planning

Generate a structured research plan with domain-aware effort allocation.

Research Question Generation

For each domain, generate 5-10 targeted search queries:

research_plan:
  domains:
    - domain: "{domain name}"
      effort_allocation: 0.70  # or 0.30 for secondary
      search_queries:
        - "{broad overview query}"
        - "{specific technical query}"
        - "{implementation/practical query}"
        - "{comparison/alternatives query}"
        - "{challenges/limitations query}"
        - "{recent developments query}"
        - "{best practices query}"
        - "{case studies query}"

Query Generation Guidelines

Query Types:

1. Overview: "What is {topic}?", "{topic} overview"
2. Technical: "{topic} implementation", "{topic} how it works"
3. Practical: "{topic} best practices", "{topic} tutorial"
4. Comparative: "{topic} vs {alternative}", "{topic} alternatives"
5. Critical: "{topic} problems", "{topic} limitations"
6. Current: "{topic} 2025", "{topic} latest developments"

Effort Allocation

effort_distribution:
  primary_domains: 0.70
  secondary_domains: 0.30

# Per-domain calculation
queries_per_primary_domain: 8-10
queries_per_secondary_domain: 4-6

Step 3: Information Gathering

Execute parallel research using domain-focused agents.

Tool Availability Check

Use the tool inventory from Step 0 to determine research capabilities:

# Core tools (expected to be available)
core_tools:
  - web_search                    # Brave Search or WebSearchPrime
  - web_reader                    # For content extraction
  - sequential_thinking           # For logic validation

# Browser automation (for dynamic/interactive content)
browser_tools:
  - agent-browser                 # PRIMARY: skill-based, install only — no MCP config needed
  - playwright_navigate           # FALLBACK: requires Playwright MCP configured
  - playwright_snapshot
  - playwright_click
  - playwright_fill_form
  - playwright_evaluate
  - browser_screenshot            # Visual capture supplement
  - browser_console_logs          # Debug info

# Specialized tools (optional)
specialized_tools:
  - documentation_query           # context7 or zread
  - repository_search             # For code/GitHub research

# Fallback strategy
fallback: "Use available tools, gracefully degrade, note limitations"

Research Methodology Selection

Choose methodology based on source types:

Codebase-Grounded Research (use DeepWiki if available):

• Questions about how an existing system works before researching alternatives
• Architecture intent and design decisions behind code you're analyzing
• Cross-component relationships and implicit contracts in a repository
• Use mcp__devin__ask_question for synthesis; mcp__devin__read_wiki_contents for specific topics
• Falls back to Glob/Grep/Read if DeepWiki not configured

Static Content (use web_search + web_reader):

• News articles
• Blog posts
• Academic papers (PDFs)
• Documentation sites
• Wikipedia-style content

Dynamic Content (use browser_automation):

• JavaScript-heavy sites (SPAs)
• Infinite scroll pages
• Content loaded on interaction
• Sites with lazy loading

Interactive Content (use browser_automation via mcp__playwright__*):

• Sites requiring login (with authorization)
• Multi-step workflows
• Form submissions
• Search interfaces
• Filtered/paginated results

Paywalled/Gated Content:

• Use browser automation if access authorized
• Note limitations if access unavailable
• Look for alternative sources

Domain Researcher Template

Spawn a researcher for each domain using the Task tool:

Capability: high

You are a {DOMAIN} RESEARCHER. Your role is to gather comprehensive information about {topic} from a {DOMAIN} perspective.

## Research Focus
{domain_specific_context}

## Search Queries to Execute
{list_of_queries}

## Your Task
1. Execute each search query using available search tools
2. Read promising sources using appropriate method:
   - Static content: web-reader
   - Dynamic content: browser automation (Playwright)
   - Interactive content: agent-browser skill (fallback: mcp__playwright__* if agent-browser not installed)
3. Extract key findings, evidence, and sources with URLs
4. Assess source credibility (HIGH/MEDIUM/LOW)
5. Identify consensus vs. debate in the field

## Tool Strategy

**For Standard Web Research:**
- Use WebSearch tools for finding sources
- Use web-reader for extracting static content
- Use documentation queries for technical topics

**For Dynamic/Interactive Sites:**
- Use `agent-browser` skill (preferred — install only):
  - Invoke: `Skill("agent-browser")` with instructions for the specific interaction
  - Handles navigation, snapshots, screenshots, JavaScript execution
- If `agent-browser` not installed, fall back to Playwright MCP tools:
  - Navigating: `mcp__playwright__browser_navigate`
  - Capturing state: `mcp__playwright__browser_snapshot`
  - Screenshots: `mcp__playwright__browser_take_screenshot`
  - JavaScript execution: `mcp__playwright__browser_evaluate`

**For Complex Interactions:**
- Use `agent-browser` skill first — no MCP config needed, handles:
  - Multi-step workflows (login → navigate → extract)
  - Complex form filling
  - Human-like interaction
- If `agent-browser` not installed, use `mcp__playwright__*` tools (requires Playwright MCP configured)
- If neither available, use `Bash` with `curl`/`wget` for simpler static requests

**URL Requirements:**
- ALWAYS capture source URLs
- Use actual page URLs (not search result URLs)
- Include direct links for cross-referencing
- Note if URL requires authentication

**Graceful Degradation:**
- If browser tools unavailable, try web-reader
- If web-reader fails, note source as "inaccessible"
- Document tool limitations in findings

## Output Format (JSON)
{
  "agent": "domain-researcher-{domain}",
  "domain": "{domain name}",
  "queries_executed": ["list of queries executed"],
  "findings": [
    {
      "topic": "specific finding",
      "severity": "CRITICAL | HIGH | MEDIUM | LOW | INFO",
      "consensus": "STRONG | MODERATE | WEAK | DEBATE",
      "evidence": ["supporting points"],
      "sources": [
        {
          "url": "source URL",
          "title": "source title",
          "credibility": "HIGH | MEDIUM | LOW",
          "type": "academic | industry | blog | documentation | other",
          "date": "publication date if available",
          "key_points": ["extracted insights"]
        }
      ]
    }
  ],
  "contradictions": [
    {
      "topic": "what's debated",
      "viewpoints": ["conflicting perspectives"]
    }
  ],
  "gaps": ["information not found or unclear"]
}

Parallel Execution

Spawn all domain researchers in parallel:

execution_strategy:
  mode: parallel
  max_concurrent: 10
  timeout: 300  # seconds per researcher

researchers:
  - "{primary_domain_1}"
  - "{primary_domain_2}"
  - "{secondary_domain_1}"
  - "{secondary_domain_2}"

Browser-Based Research Crawling

For sources requiring browser automation, use this workflow:

When to Use Browser Automation

Use browser tools when encountering:

• ❌ web-reader returns incomplete content
• ❌ "JavaScript required" messages
• ❌ Dynamic content not loading
• ❌ Search interfaces requiring interaction
• ❌ Paginated results needing navigation
• ❌ Content behind forms or filters

Playwright Research Workflow

Basic Navigation & Extraction:

1. Navigate to URL
   → mcp__playwright__browser_navigate(url)

2. Capture page state
   → mcp__playwright__browser_snapshot
   → Returns: HTML, visible text, accessibility tree

3. Extract specific content
   → mcp__playwright__browser_evaluate(script)
   → Execute JavaScript to extract data

Interactive Research (Forms, Search, Filters):

1. Navigate to site
   → browser_navigate(url)

2. Fill search/filter forms
   → browser_fill_form(selector, value)

3. Click search/submit buttons
   → browser_click(selector)

4. Wait for results to load
   → browser_evaluate("check if loaded")

5. Extract results
   → browser_snapshot or browser_evaluate

Multi-Page Crawling:

For paginated results:
1. Extract page 1
2. Click "Next" button
3. Extract page 2
4. Repeat until complete or limit reached
5. Aggregate all findings

Complex Interaction via agent-browser (with Playwright MCP fallback)

For complex multi-step workflows, use agent-browser skill (preferred — install only, no MCP config needed). If agent-browser is not installed, fall back to mcp__playwright__* tools (requires Playwright MCP configured). If neither is available, use Bash with curl/wget.

Preference order:

agent-browser → mcp__playwright__* → Bash curl/wget

Example agent-browser workflow:

Invoke Skill("agent-browser") with instructions:
- Go to {research_site}
- Search for: {query}
- Extract top results: title, URL, summary, date
- Return as structured JSON

Fallback — Playwright MCP (if agent-browser not installed):

1. mcp__playwright__browser_navigate(url="{research_site}")
2. mcp__playwright__browser_fill_form(selector="input[name=q]", value="{query}")
3. mcp__playwright__browser_click(selector="button[type=submit]")
4. mcp__playwright__browser_snapshot() → extract top results with title, URL, summary, date
5. Return as structured JSON

URL Capture Requirements

CRITICAL: Always capture actual content URLs:

✅ Correct:

{
  "url": "https://example.com/article/actual-content",
  "title": "Article Title",
  "method": "playwright-browser-automation"
}

❌ Wrong:

{
  "url": "https://google.com/search?q=...",
  "title": "Search results",
  "method": "web-search"
}

URL Validation:

• URL must point to actual content (not search results)
• URL must be clickable/accessible
• URL must be permanent (not session-specific)
• If URL requires auth, note: "access": "requires-authentication"

Browser Automation Best Practices

Performance:

• Use browser automation only when necessary
• Prefer web-reader for static content
• Limit concurrent browser sessions (max 3)
• Set reasonable timeouts (30s per page)

Ethics & Legal:

• Respect robots.txt
• Honor rate limiting
• Don't bypass paywalls without authorization
• Note access requirements in findings

Error Handling:

• If browser automation fails → try web-reader
• If web-reader fails → note as "inaccessible"
• Log failures in research quality assessment
• Don't block research on single source failure

Step 4: Cross-Domain Exploration

After domain researchers complete, spawn cross-domain analysts:

Cross-Domain Analyst Template

Capability: high

You are a CROSS-DOMAIN ANALYST. Your role is to explore intersections and connections between domains.

## Domains to Analyze
{list_of_domains}

## Domain Findings Summary
{summary_of_findings_from_each_domain}

## Your Task
1. Identify intersections between domains
2. Find where domains agree and disagree
3. Surface insights that emerge only from cross-domain perspective
4. Identify trade-offs and tensions

## Output Format (JSON)
{
  "agent": "cross-domain-analyst",
  "intersections": [
    {
      "domains": ["domain1", "domain2"],
      "connection": "how they relate",
      "agreements": ["where domains align"],
      "tensions": ["where domains conflict"],
      "emergent_insights": ["insights from intersection"]
    }
  ],
  "domain_mapping": {
    "domain1": ["related domains"],
    "domain2": ["related domains"]
  }
}

Step 5: Analysis & Synthesis

Analyze findings for quality and generate synthesis.

Validation Strategy (Three-Layer)

Layer 1: Source Credibility Assessment

credibility_criteria:
  HIGH:
    - Academic papers with peer review
    - Official documentation
    - Industry standards bodies
    - Recognized experts in field

  MEDIUM:
    - Industry blogs (established companies)
    - Technical tutorials (reputable sources)
    - Conference presentations
    - Books from known publishers

  LOW:
    - Personal blogs without credentials
    - Forum discussions
    - Social media posts
    - Unverified claims

Layer 2: Cross-Reference Validation

validation_method: triangulation

# Finding is validated if:
triangulation_criteria:
  - Mentioned by 3+ independent sources
  - Appears in HIGH credibility sources
  - Consistent across domains

Layer 3: Internal Consistency Check

Use sequential-thinking tool to validate:
- Logical consistency of findings
- Cause-effect relationships
- Assumption validity

Synthesis Generation

Generate synthesis by:

1. Identify Patterns: What themes emerge across sources?
2. Surface Debates: Where do sources disagree?
3. Assess Evidence: Which findings are well-supported?
4. Map Connections: How do domains relate?

Step 6: Recommendations & Reporting

Generate structured report with evidence-based recommendations.

Report Structure

# Deep Research Report: {Topic}

**Generated:** {timestamp}
**Research Duration:** {duration}
**Domains Analyzed:** {list of domains}
**Sources Consulted:** {count}
**Validation Status:** {VALIDATED | PARTIAL | PRELIMINARY}

## Executive Summary

{3-5 sentence overview of key findings and recommendations}

---

## Research Intent & Scope

**Primary Research Question:**
{main question/topic}

**Scope:**
- Primary Domains: {list}
- Secondary Domains: {list}
- Research Depth: {BRIEF | STANDARD | COMPREHENSIVE}

---

## Key Findings by Domain

### {Domain 1}

#### {Finding 1}

**Consensus:** {STRONG | MODERATE | WEAK | DEBATE}

**Evidence:**
- {point 1}
- {point 2}

**Sources:**
- [{Title}]({URL}) - {credibility} - {key insight}
- [{Title}]({URL}) - {credibility} - {key insight}

#### {Finding 2}

{repeat pattern}

---

## Cross-Domain Insights

### Intersection: {Domain 1} + {Domain 2}

**Connection:** {how domains relate}

**Agreements:**
- {where domains align}

**Tensions:**
- {where domains conflict}

**Emergent Insights:**
- {insights from intersection}

---

## Synthesis & Patterns

### Key Patterns
{patterns identified across domains}

### Contradictions & Debates
{areas of disagreement with viewpoints}

### Information Gaps
{what could not be found or needs more research}

---

## Recommendations

### Recommendation 1: {Actionable recommendation}

**Rationale:** {evidence-based reasoning}

**Confidence:** {HIGH | MEDIUM | LOW}

**Supporting Evidence:**
- {finding from domain/source}

### Recommendation 2: {Another recommendation}

{repeat pattern}

---

## Research Quality Assessment

**Validation Method:** Triangulation across sources
**Source Credibility Distribution:**
- HIGH: {count} sources
- MEDIUM: {count} sources
- LOW: {count} sources

**Cross-Domain Validation:**
- {percentage}% of findings validated across multiple domains

**Limitations:**
- {constraints or gaps in research}

---

## Sources Bibliography

### {Domain 1} Sources
1. [{Title}]({URL}) - {credibility} - {date}
2. [{Title}]({URL}) - {credibility} - {date}

### {Domain 2} Sources
{repeat pattern}

---

## Appendix: Research Methodology

**Search Strategy:**
- Total queries executed: {count}
- Domains researched: {list}
- Tools used: {list}

**Quality Controls:**
- Source credibility assessment: ✓
- Cross-reference validation: ✓
- Internal consistency check: ✓

Step 7: Validate Output Format

Before finalizing the report, validate it against the required format specification to ensure consistency.

Validation Gate

Spawn an output validator sub-agent using the Task tool:

Capability: standard

You are an OUTPUT VALIDATOR for deep-research reports. Your role is to ensure format compliance.

## Files to Validate
- Markdown: {path_to_markdown_file}
- JSON: {path_to_json_file}

## Validation Instructions
Follow the validation procedure defined in: skills/deep-research/validators/output-validator.md

## Schema Location
JSON Schema: skills/deep-research/schemas/research-report-schema.json

## Tasks
1. Load and validate JSON against schema
2. Validate markdown structure and required sections
3. Verify source quality and credibility distribution
4. Cross-check consistency between JSON and markdown
5. Generate validation report

## Output Format
Return validation result as JSON with:
- validation_status: PASS or FAIL
- Specific errors and warnings
- Source quality assessment
- Suggestions for fixes

## Strictness
FAIL on any critical errors:
- Missing required fields
- Invalid enum values
- Type mismatches
- Missing required sections
- Poor source quality (>70% LOW credibility)
- Insufficient sources (<3 total)

Handling Validation Results

If validation PASSES:

• Proceed to Step 8 (Save Report)

If validation FAILS:

1. Display all errors and warnings to user
2. Provide specific suggestions for each violation
3. DO NOT save report as "latest"
4. Ask user if they want to:
   • Fix the issues and regenerate
   • Override validation (with explicit confirmation)
   • Cancel research

Example failure output:

❌ Validation FAILED

JSON Errors:
- Missing required field: recommendations
- Invalid research_depth value: 'DEEP' (must be BRIEF, STANDARD, or COMPREHENSIVE)
- executive_summary only 35 characters (minimum: 50)

Markdown Errors:
- Missing required section: ## Research Quality Assessment
- Domain 'Machine Learning' in metadata but no findings section

Source Quality Issues:
- Only 2 sources consulted (minimum: 3)
- 80% of sources are LOW credibility (threshold: 70%)
- Only 1 HIGH credibility source (recommended: at least 30%)

Suggestions:
1. Add at least one recommendation with rationale and evidence
2. Change research_depth to BRIEF, STANDARD, or COMPREHENSIVE
3. Expand executive_summary to at least 50 characters
4. Add ## Research Quality Assessment section
5. Add findings section for Machine Learning or remove from domains
6. Conduct more research - consult at least 3 sources total
7. Include more HIGH credibility sources (academic papers, official docs)

Would you like to regenerate the report with corrections?

Step 8: Save Report

Artifact Output

Save to .outputs/research/{YYYYMMDD-HHMMSS}-research-{slug}.md with YAML frontmatter:

---
skill: deep-research
timestamp: {ISO-8601}
artifact_type: research
domains: [{domain1}, {domain2}]
context_summary: "{brief description of what was researched}"
session_id: "{unique id}"
---

Also save JSON companion: {timestamp}-research-{slug}.json

No symlinks. To find the latest artifact:

ls -t .outputs/research/ | head -1

QMD Integration (optional, progressive enhancement):

qmd collection add .outputs/research/ --name "deep-research-artifacts" --mask "**/*.md" 2>/dev/null || true
qmd update 2>/dev/null || true

Note: Only save reports that pass validation.

.outputs/research/
├── 20250115-143022-research-event-sourcing-kafka.md
├── 20250115-143022-research-event-sourcing-kafka.json
├── 20250116-091545-research-ai-regulation-business.md
└── 20250116-091545-research-ai-regulation-business.json

Step 9: Configuration (Optional)

The system uses these defaults unless overridden:

Default Configuration

# Research execution
research:
  parallel_execution: true
  max_concurrent_researchers: 10
  timeout_seconds: 300

# Effort allocation
effort_distribution:
  primary_domains: 0.70
  secondary_domains: 0.30

# Queries per domain
queries_per_domain:
  primary: 8-10
  secondary: 4-6

# Output
output_directory: ".outputs/research/"
output_format: "markdown"
include_json: true

Configuration Override Order

1. Built-in defaults
2. Environment variables
3. Config files in .outputs/research/config.yaml
4. Command-line arguments

Environment Variables

# Execution
export DEEP_RESEARCH_PARALLEL="true"
export DEEP_RESEARCH_MAX_CONCURRENT="10"
export DEEP_RESEARCH_TIMEOUT="300"

# Effort allocation
export DEEP_RESEARCH_PRIMARY_RATIO="0.70"
export DEEP_RESEARCH_SECONDARY_RATIO="0.30"

# Output
export DEEP_RESEARCH_OUTPUT_DIR=".outputs/research/"
export DEEP_RESEARCH_OUTPUT_FORMAT="markdown"

Notes

• Model-agnostic: Uses capability levels ("highest", "high", "standard") not specific model names
• Domain-agnostic: Works for any domain detected from conversation
• Conversation-driven: Extracts research intent and domains from what was discussed
• Tool-agnostic: Gracefully handles unavailable tools with fallbacks
• Evidence-based: All findings tied to sources with credibility assessment
• Cross-domain: Identifies insights that emerge from domain intersections
• Deep Council: For cross-runtime/cross-council research synthesis, see deep-council
• Domain-Aware: Research agents adapt to domain context via domain-registry
• Context Routing: If the research scope is complex or multi-domain, invoke the context skill first to classify artifact type and determine optimal routing (parallel-workflow vs debate-protocol vs deep-council)
• Severity vs Consensus: severity (CRITICAL/HIGH/MEDIUM/LOW/INFO) reflects the finding's importance; consensus (STRONG/MODERATE/WEAK/DEBATE) reflects source agreement strength — both are recorded per finding

Integration Boundary: Research Dossiers, Not Editorial Pipelines

deep-research is a generic evidence-acquisition and synthesis framework. It may produce artifacts consumed by editorial systems, strategy workflows, implementation planners, or humans, but it must not embed consumer-specific workflow logic.

Boundary rule:

deep-research → research_dossier → downstream consumer

Do not turn this into:

deep-research ↔ Cortex/editorial combined pipeline

Deep Research Owns

• research framing and domain detection
• source discovery and extraction
• credibility assessment
• claim/evidence mapping
• contradiction and gap analysis
• confidence scoring
• generic reports, dossiers, source bundles, debate packets, and decision briefs

Downstream Consumers Own

• editorial thesis and angle selection
• audience and voice
• narrative structure
• headline/lede/publishability
• channel packaging
• product-specific workflow gates

Generic Dossier Contract

Prefer consumer-neutral artifact names:

• research_dossier
• source_bundle
• claim_matrix
• evidence_log
• debate_packet
• decision_brief

Avoid consumer-specific names in this skill, such as cortex_packet, editorial_brief, article_seed, or publication_dossier. A downstream system may transform the dossier into those artifacts outside deep-research.

Optional Research Brainstorm Mode

When the research topic is broad or ambiguous, use a generic brainstorm stage before gathering evidence:

1. Frame — clarify question, domains, constraints, and excluded scope.
2. Diverge — generate candidate questions, hypotheses, source classes, stakeholder lenses, and likely controversies.
3. Plan — choose research questions, source strategy, and evidence thresholds.
4. Gather — collect sources and evidence.
5. Triangulate — map claims to evidence for/against.
6. Contest — produce debate packets for contradictions and weak claims.
7. Synthesize — emit dossier/report with confidence and limitations.

Hypotheses from brainstorm mode are unverified until supported by evidence. Do not convert research brainstorm output directly into publication copy or implementation plans without a separate downstream workflow.