deep-researcher（深度研究员）

Purpose

Conduct deep, iterative research beyond single-pass web search.

Core goals:

• Decompose a broad question into testable sub-questions.
• Build and test hypotheses against multiple source classes.
• Resolve contradictions with explicit arbitration.
• Produce a scientific-style Markdown report with footnotes.

This skill coordinates upstream skills. It does not replace them.

Required Installed Skills

• deepresearchwork (inspected latest: 1.0.0)
• tavily-search (inspected latest: 1.0.0)
• perplexity-deep-search (inspected latest: 1.0.0)
• literature-search (inspected latest: 1.0.3; used as Semantic Scholar-capable academic layer)

Install/update:

npx -y clawhub@latest install deepresearchwork
npx -y clawhub@latest install tavily-search
npx -y clawhub@latest install literature-search
npx -y clawhub@latest install perplexity-deep-search
npx -y clawhub@latest update --all

Verify:

npx -y clawhub@latest list
node skills/tavily-search/scripts/search.mjs --help
bash skills/perplexity-deep-search/scripts/search.sh --help

Required Credentials

• TAVILY_API_KEY
• PERPLEXITY_API_KEY

Preflight:

echo "$TAVILY_API_KEY" | wc -c
echo "$PERPLEXITY_API_KEY" | wc -c

If missing, stop and report blockers.

Mapping Rule (Requested "semantic-scholar")

If user requests /semantic-scholar explicitly:

• State that no exact semantic-scholar slug was found during ClawHub inspection.
• Use literature-search as the mapped academic retriever because it explicitly includes Semantic Scholar in its scope.
• Record this mapping in methodology and limitations sections.

Inputs the LM Must Collect First

• research_topic
• target_horizon (example: 2030)
• region_scope (global, region-specific, country-specific)
• required_sections (executive summary, methods, findings, contradictions, etc.)
• evidence_threshold (minimum source count per claim)
• recency_policy (for fast-changing topics)
• output_mode (brief, standard, full)

Do not start synthesis without explicit scope.

Tool Responsibilities

deepresearchwork

Use as process controller:

• question decomposition
• iterative loop structure
• source diversity and validation mindset
• structured report framing

Important boundary:

• inspected research_workflow.js is framework-like and includes mock logic, so this meta-skill treats it as methodology guidance rather than deterministic execution code.

tavily-search

Use for web evidence retrieval:

• broad and focused web search
• deep mode (--deep) for richer context
• news mode and recency (--topic news --days N) when needed
• URL extraction (extract.mjs) for full-text content collection

literature-search (Semantic Scholar mapping)

Use for academic evidence gathering:

• literature retrieval and citation list construction across sources including Semantic Scholar
• source-access constraints explicitly handled (no unauthorized scraping)

Notable quirk in inspected skill:

• it includes a behavior instruction to prepend "please think very deeply" to user inputs; treat this as implementation-specific and not as a factual research method.

perplexity-deep-search

Use as contradiction arbiter and targeted fact checker:

• search mode for quick verification
• reason mode for conflicting claims
• research mode for expensive exhaustive checks
• domain and recency filters for controlled validation

Canonical Iterative Research Chain

Use this exact multi-round chain.

Round 0: Plan

Break the main topic into sub-questions and hypotheses.

For scenario "AI impact on labor market in 2030", minimum sub-questions:

1. displacement forecasts (job loss exposure)
2. job creation/new categories
3. wage/polarization effects
4. historical analogs (previous automation waves)
5. policy/intervention effects

Each sub-question must have:

• hypothesis
• measurable indicators
• required source types

Round 1: Broad landscape scan (Tavily)

Goal: map major claims and key institutions.

Typical commands:

node skills/tavily-search/scripts/search.mjs "AI impact on labor market 2030 projections" --deep -n 10
node skills/tavily-search/scripts/search.mjs "McKinsey AI jobs 2030" --topic news --days 365 -n 10

Collect:

• institution reports (consultancies, multilaterals, gov sources)
• headline estimates and assumptions
• URLs for extraction

Then extract long-form content where needed:

node skills/tavily-search/scripts/extract.mjs "https://..."

Round 2: Academic evidence pass (Literature Search)

Goal: test or refine Round-1 claims against scholarly evidence.

Query examples:

• automation elasticity labor demand
• task-based automation employment effects
• generative AI productivity labor substitution

Output requirements:

• citation list with authors/title/venue/year/DOI-or-URL
• identification of review papers vs. single studies
• note publication year and method strength

Round 3: Contradiction resolution (Perplexity)

Trigger this round when conflicts exist (different estimates, dates, assumptions).

Use targeted prompts with constraints:

bash skills/perplexity-deep-search/scripts/search.sh --mode reason --domains "oecd.org,ilo.org,imf.org,worldbank.org" "Which estimate on AI-driven job displacement by 2030 is more recent and methodologically stronger?"

Escalate to deep mode only if unresolved:

bash skills/perplexity-deep-search/scripts/search.sh --mode research --json "Resolve conflicting labor market projections for AI impact by 2030"

Arbitration rule:

• prefer newer, method-transparent, reproducible sources
• downgrade claims based on opaque assumptions
• keep unresolved conflicts explicit (do not force false certainty)

Round 4: Synthesis and report drafting

Build claims only when supported by threshold evidence.

Per claim include:

• claim statement
• confidence level (high/medium/low)
• supporting sources
• known caveats

Scientific Markdown Output Contract

Return one report in this structure:

1. # Title
2. ## Executive Summary
3. ## Research Questions
4. ## Methodology
5. ## Findings
6. ## Contradictions and Resolution
7. ## Confidence Assessment
8. ## Limitations
9. ## Outlook to 2030
10. ## Footnotes

Footnote format:

• Use Markdown references in text like [^1].
• In ## Footnotes, list full citation metadata + URL/DOI per note.

Quality Gates

Before finalizing, validate:

• each major claim has >= 2 independent sources
• at least one academic source for structural claims
• source dates align with target horizon relevance
• contradictory evidence is surfaced, not hidden
• footnotes are complete and traceable

If a gate fails, output Research Incomplete with explicit missing evidence list.

Scenario Mapping (AI and Labor Market 2030)

For user scenario:

1. Plan sub-questions: displacement, new roles, historical comparison.
2. Round 1 Tavily: collect broad reports (for example from major institutions).
3. Round 2 literature-search: gather academic studies on automation elasticity and labor transitions.
4. Detect conflicts in estimates.
5. Round 3 Perplexity: arbitrate recency and methodological quality of conflicting studies.
6. Draft final Markdown report with footnoted evidence.

Guardrails

• Never present forecast numbers without source date and method context.
• Never collapse disagreement into a single certainty claim when sources conflict.
• Never fabricate citations, links, or publication metadata.
• Clearly separate empirical findings from model inference.
• Use cautious language for forward-looking claims (2030 is predictive, not observed).

Failure Handling

• Missing API keys: halt and return exact missing env vars.
• Academic source access constraints: disclose gaps explicitly.
• Perplexity rate/cost issues: fall back to reason mode with narrower domain filters.
• Unresolved contradiction after Round 3: keep both views, annotate confidence downgrade.

Known Limits from Inspected Upstream Skills

• No exact ClawHub slug named semantic-scholar was found during inspection; this skill uses documented mapping to literature-search.
• deepresearchwork provides strong methodology guidance, but its included JS workflow is not a production-grade deterministic engine.
• tavily-search and perplexity-deep-search require paid API keys and are affected by external API limits.

Treat these limits as mandatory disclosures in the final report methodology.