ericgandrade/claude-superskills/deep-research

Deep Research

Purpose

Run structured, multi-step web research with evidence-first synthesis, source traceability, skeptical evaluation, and explicit confidence/gap analysis.

This skill works in native mode with built-in web research tools and does not require Google/Gemini, OpenAI, Anthropic, OpenRouter, or any other paid provider setup. When the user explicitly wants maximum depth, use dense/frontier mode to recommend the strongest available model and run a wider parallel research topology where the host platform supports subagents.

When to Use This Skill

Use this skill when:

• Performing market analysis
• Conducting competitive landscaping
• Building literature or source reviews
• Doing technical due diligence
• Preparing decision memos with citations
• Producing dense, high-confidence, Perplexity-like research with adversarial critique

Requirements

• Access to built-in web research tools (WebSearch, WebFetch)
• Clear research question and scope

No external API key is required for native mode.

Progress Tracking

Display a progress gauge at each research phase:

[████░░░░░░░░░░░░░░░░] 20% — Phase 1/5: Objective, Scope & Decomposition
[████████░░░░░░░░░░░░] 40% — Phase 2/5: Parallel Source Collection
[████████████░░░░░░░░] 60% — Phase 3/5: Evidence Ledger & Triangulation
[████████████████░░░░] 80% — Phase 4/5: Synthesis & Confidence/Gaps
[████████████████████] 100% — Phase 5/5: Citation Audit & Final Review

Operating Modes

Native Research

Use native mode by default.

• Use built-in WebSearch and WebFetch.
• Decompose complex topics into 3-5 sub-questions.
• Run parallel ResearchScout agents where subagents are available.
• Produce an evidence ledger, citations, and Confidence & Gaps.

Dense / Frontier Research

Use dense/frontier mode when the user asks for maximum depth, exhaustive research, "frontier model", "Perplexity-like" research, "turbinado", adversarial review, high-confidence evidence, or says cost is secondary to research quality.

• Recommend the strongest available model or model class before execution.
• Prefer the host platform's frontier model for synthesis and critique.
• Run a wider parallel topology for sub-questions, primary-source harvesting, contrarian evidence, recency checks, citation audit, and synthesis.
• Add critique and rebuttal rounds before final synthesis.
• Increase source quota and require deeper source extraction.
• Include a model/tooling note in the final report.

If dense/frontier mode is requested but frontier tooling is unavailable, continue with native mode and disclose the limitation.

Frontier Model Recommendation

For dense/frontier research, recommend the strongest available model before execution:

• Use an Opus-class Claude model for high-stakes synthesis, adversarial judgment, and nuanced source conflict analysis when available.
• Use a GPT frontier reasoning/chat model for broad synthesis, structured reasoning, and tool-heavy research when available.
• Use a Codex-class model for code-heavy technical research, API analysis, repository analysis, and engineering claims.
• Use a Gemini Pro-class thinking model for long-context, multimodal, PDF-heavy, and Google-grounded research when available.
• Use OpenRouter when the user wants access to multiple frontier model families through one API or wants a latest-family alias.

State the recommendation in this form:

Recommended research model: <model or model class>.
Reason: <why this model fits the research task>.
Fallback: <native mode or next best model>.

Do not hard-code one model as universally best. Use the strongest available model/tooling path and disclose limitations.

Research Modes

Before starting, present the user with three modes:

Which research mode do you want?
 
  [1] Standard    — 12–15 sources · 1.500–3.000 words · findings by sub-question · single Confidence & Gaps block
  [2] Verbose     — 20–30 sources · 4.000–6.000 words · per-source deep-dives · per-sub-question confidence · Devil's Advocate section · Appendices
  [3] Exhaustive  — 40–60+ sources · 8.000+ words · full dense/frontier agent topology · block quotes allowed · stakeholder map · chronology · multiple critique rounds

If the user does not specify, default to Standard.

Research Protocol

1. Present mode options and confirm with the user before proceeding.

0.5. Create and present a Research Plan

Before launching any agents, produce a structured research plan and show it to the user. Proceed unless they want changes.

## Research Plan: [Topic]
 
**Core Question:** [one sentence — what exactly needs answering]
**Decision Context:** [how results will be used: strategy, due diligence, comparison, etc.]
**Scope:** [geography, time horizon, industry, technology boundaries]
 
### Research Dimensions
1.[Dimension] — [what to find and why it matters]
2.[Dimension] — [what to find and why it matters]
3.[Dimension] — [what to find and why it matters]
...
 
### Source Strategy
-Primary: [official docs, filings, government data, company pages]
-Secondary: [analyst reports, established media, expert orgs]
-Validation: [how conflicting claims will be resolved]
 
### Known Unknowns
-[What we know we don't know yet — data that may be hard to find]

Show the plan and wait for user confirmation before proceeding.

1. Define objective, scope, and decomposition

• Restate the original research question in one sentence.
• Identify audience, decision context, time horizon, geography/market scope, and requested output format.
• Decompose the topic into 3-5 concrete sub-questions.
• Each sub-question must cover a distinct dimension of the original question.
• Together, the sub-questions should be sufficient to answer the main question.
• If the topic is too broad to decompose cleanly, narrow the scope before searching.

Example decomposition:

• Market landscape: who are the major players?
• Evidence base: what reliable data exists?
• Differentiation: what capabilities or positions differ?
• Risks: what claims are uncertain, contested, or under-sourced?
• Decision implication: what should the target audience do with this evidence?

1. Build search strategy by sub-question

• Create targeted query variants for each sub-question.
• Include broad, narrow, comparative, opposing, primary-source, and recent-source queries when relevant.
• Prefer sources in this order:

1. Primary sources: official docs, filings, regulatory pages, standards bodies, government data, company pages
2. Scholarly sources: peer-reviewed papers, preprints with clear methodology, university research
3. Reputable secondary sources: analyst reports, established media, expert organizations
4. Supporting context only: blogs, newsletters, aggregators, marketing pages, review sites

• Use low-authority sources only when they provide unique firsthand evidence or market/customer sentiment.

Reject or down-rank sources that:

• Do not identify publisher, author, or date when those details matter
• Make numerical claims without methodology
• Repeat another source without adding primary evidence
• Are mostly promotional unless the research question is about vendor positioning
• Conflict with stronger primary evidence

Agent Status Notifications

For every agent, emit three status lines in this exact format:

⟳  <Agent Name> — <what it is doing>
✅  <Agent Name> — Done · <N> sources collected · <key finding in one sentence>

Example:

⟳  🔍 Scout Alpha — Scanning primary sources for sub-question 1...
✅  🔍 Scout Alpha — Done · 8 sources collected · Three vendors dominate with >60% market share combined.
 
⟳  ⚔️ Devil's Advocate — Searching for contradictions and critical evidence...
✅  ⚔️ Devil's Advocate — Done · 4 sources collected · Two analyst reports dispute the market share figures.

Emit the ⟳ line immediately before launching each agent. Emit the ✅ line immediately after the agent returns results.

Parallel Query Execution

Do NOT run searches sequentially. Launch one agent per sub-question or major query type simultaneously in a single block where the host platform supports subagents.

Each agent prompt begins with:

# <Agent Name> — Deep Research Agent
Role: Execute assigned web research using WebSearch/WebFetch. Collect authoritative sources, extract claim-level evidence, identify contradictions, and return structured results.
 
Required output:
- Sub-question or query assignment
- Queries attempted
- Source inventory with URL, title, publisher, date, and relevance
- Key claims and data points with source attribution
- Quote directly when the exact wording is material — definitions, methodologies, thresholds, or claims where paraphrasing would change the meaning. Avoid decorative quotes that add length without adding precision. (block quotes allowed in Verbose/Exhaustive mode)
- Evidence quality notes
- Conflicts, gaps, and inaccessible/paywalled sources

Wait for all agents to complete. Emit a summary banner before deduplication:

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
  All agents finished · Deduplicating results...
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Then deduplicate results by canonical URL and run a Gap Analysis before proceeding to the evidence ledger.

Gap Analysis & Iteration Check

After each agent batch, reason explicitly about findings vs needs before moving to synthesis:

### Gap Analysis — Iteration [N]
 
**Coverage so far:** [summary of dimensions addressed and confidence level per dimension]
 
**Gaps identified:**
-[ ] [Missing data point] — needed because [reason]
-[ ] [Conflicting claims unresolved] — need [source type] to adjudicate
-[ ] [Shallow coverage] — [dimension] needs deeper investigation
 
**Decision:** [Stop — all dimensions sufficient] / [Continue — targeted follow-up on gaps above]

Stop Criteria

Stop iterating when any of these conditions are met:

1. Sufficient coverage — all research dimensions have confirmed or contested (not weak) evidence
2. Diminishing returns — last iteration added less than 10% new distinct information
3. Maximum iterations reached — Standard: 1 follow-up, Verbose: 2, Exhaustive: 3
4. Source saturation — same sources appearing repeatedly across independent searches
5. Time budget exceeded — stop and synthesize with what's available, note gaps explicitly