deep-research

name	deep-research
description	Use when the user explicitly asks for deep research / 深度研究 / 深入调研, or when they need evidence-driven, multi-source research for a decision, report, due diligence, current-state analysis, technical comparison, recommendation, or contested/high-stakes question.

Deep Research

Deep research is a disciplined research workflow, not a longer web search. Use it to answer questions where factual accuracy, source quality, recency, disagreement, and decision usefulness matter.

Keep the run proportional. A narrow question should get a tight evidence brief. A broad, contested, or high-stakes question should get a fuller research report.

Activation

Use this skill when one of these is true:

• The user explicitly says "deep research", "深度研究", "深入调研", or "deep dive".
• The user asks for a research-backed report, memo, market scan, technical comparison, policy/current-state analysis, or due diligence.
• The answer depends on current facts, contested evidence, multiple stakeholders, or source credibility.
• The user will likely use the answer for planning, external writing, investment, compliance, medical, legal, product, or engineering decisions.

Do not use this skill for quick definitions, simple factual lookups, ordinary coding questions, casual opinions, or cases where one authoritative source fully resolves the question.

If the user's wording is ambiguous, infer the lightest mode that will still protect accuracy.

Execution Gate

When this skill is activated, do not answer from memory alone.

Use available research tools to gather fresh evidence unless the user-provided materials already contain enough source evidence to answer the question. At minimum:

• Search or inspect multiple source families appropriate to the task.
• Open and read the key sources used for central claims.
• Prefer primary, official, original, or user-provided sources for the claims that drive the conclusion.
• Record source date, data period, version, or effective date when it matters.
• Treat snippets, search-result summaries, AI summaries, and repeated secondary articles as leads, not evidence.

If tool access is unavailable, source access is blocked, or the user asks for an offline answer, state that limitation before giving conclusions and lower confidence accordingly.

Choose A Mode

Fast Deep Research

Use for narrow questions, initial briefs, short memos, or "quick research" requests.

Minimum standard:

• 2-4 focused subquestions.
• Usually 3-6 high-quality sources.
• At least one primary or official source when available.
• Cross-check the facts that drive the conclusion.
• Final output usually has 3-5 findings.

Do not use fast mode for high-stakes medical, legal, financial, compliance, safety, or broad contested questions.

Standard Deep Research

Use for broad, strategic, contested, high-stakes, or external-facing work.

Minimum standard:

• Explicit research frame and scope.
• Multiple source families, not just one search phrasing.
• Usually 8-15 strong sources, adjusted to topic size.
• Primary/official sources for central claims whenever available.
• Cross-check important facts, numbers, dates, roles, and versions.
• Treat material contradictions directly.
• Final output includes findings, analysis, uncertainty, implications, and references.

Stop Conditions

Stop collecting sources when the research question is answerable at the needed confidence level:

• Central claims are supported by primary or strong sources where available.
• Key numbers, dates, versions, roles, and definitions have been checked.
• Obvious counterqueries or likely objections have been searched.
• Material contradictions have been explained or marked unresolved.
• Additional searching mostly returns duplicate secondary summaries.

Do not stop just because the source count target was reached. Do not continue just to inflate the source count when the added sources are derivative or low value.

Research Frame

Before searching, define these internally. Show them only when helpful.

• Primary question: what must be answered.
• User purpose: decision, explanation, comparison, writing support, due diligence, etc.
• Scope: included and excluded topics, geography, time horizon, domain.
• Audience: general, executive, technical, academic, operational.
• Recency need: historical, current, latest, today, this year, version-specific.
• Output shape: brief, report, matrix, recommendation, memo, annotated bibliography.

Ask a clarification only when a missing boundary would materially change the research. Otherwise state reasonable assumptions in the answer.

Clarification Pass

Before researching, identify missing information that would materially change source selection, scope, or the final recommendation.

If clarification is needed, ask at most 3 questions. For each question, provide suggested choices so the user can answer quickly. Include one recommended default when reasonable.

Use this shape:

Before I start, these choices would materially change the research:

1. **Question:** [What boundary needs clarification?]
   - **Recommended default:** [Default choice and why]
   - **Options:** [A / B / C]
   - **Why it matters:** [How this changes source selection, scope, or recommendation]

Only block on clarification when proceeding would likely produce the wrong research. If the missing information is low-impact, state the assumed default and continue.

Source Strategy

Use the strongest source types available for the task:

• Official, primary, or original sources: laws, filings, standards, docs, reports, datasets, papers, press releases, project repositories.
• Peer-reviewed or formal technical sources: trials, reviews, standards, RFCs, specifications, benchmarks.
• Institutional sources: regulators, governments, universities, think tanks, industry bodies.
• Reputable journalism and analysis: useful for events, stakeholder reactions, and context, but not a substitute for primary evidence.
• Expert commentary: useful for interpretation; label it as interpretation.
• User-provided files or private sources: inspect them first when they are central to the question.

Prefer source fitness over source quantity. Do not treat repeated secondary summaries as independent confirmation.

Domain Evidence Rules

Apply these priority rules when relevant:

• Medical/biomedical: prefer systematic reviews, clinical guidelines, randomized trials, large cohort studies, regulator labels, and trial registries. Separate efficacy, safety, population, endpoint, and evidence certainty.
• Legal/policy/compliance: prefer statutes, regulations, agency guidance, court opinions, official notices, and effective dates. Treat commentary as secondary.
• Finance/investment/markets: prefer filings, audited statements, exchange/regulator data, central-bank or statistical-agency data, fund documents, and current prices/fees/premiums. State data periods and market accessibility.
• Software/API/standards: prefer official docs, source code, changelogs, release notes, specs, RFCs, and reproducible examples. Verify version and date.
• News/current events: distinguish event date, publication date, update time, and what is confirmed versus alleged.
• Company/product research: prefer official docs, pricing pages, filings, support docs, changelogs, status pages, and credible user/industry evidence.

High-Stakes Advice Boundaries

For medical, legal, financial, compliance, or safety-sensitive topics:

• Treat the output as research synthesis or educational analysis unless the user provides enough context for a bounded professional-style decision memo.
• Do not present individualized diagnosis, treatment, legal advice, tax advice, investment allocation, or safety instructions as a definitive recommendation.
• State the missing variables that would materially change the answer, such as jurisdiction, patient population, time horizon, risk tolerance, tax status, existing obligations, or operating environment.
• Separate what the evidence says from what a professional decision would require.
• Prefer conditional language: "the evidence supports X under these assumptions" rather than "you should do X."

Search Workflow

1. Map

• Start with broad discovery to identify terms, actors, source names, dates, and disputed points.
• Build a small topic map: dimensions, likely primary sources, gaps, and likely biases.
• Do not settle on the first plausible narrative.

2. Deepen

• Search each important dimension with varied phrasing.
• Open and read important sources beyond snippets.
• Capture claim, source, date, evidence type, limitation, and relevance.
• Seek counterevidence or criticism for major claims.

For current, technical, legal, medical, financial, or policy topics, include at least one query aimed at primary sources and one query aimed at criticism, contradiction, limitations, or recent changes.

3. Validate

For every central finding:

• Verify important numbers, names, dates, roles, versions, and definitions.
• Cross-check with independent sources where possible.
• Prefer primary evidence when secondary sources conflict.
• Separate facts, interpretations, tentative inferences, and unresolved disputes.
• Downgrade confidence when evidence is old, indirect, paywalled/unread, derivative, or inconsistent.

4. Synthesize

• Group findings by what matters to the user's purpose.
• Lead with conclusions, then evidence.
• Explain disagreement instead of flattening it.
• State practical implications and remaining uncertainty.

Temporal Rules

Never assume "latest" from memory. Use the actual current date in context.

When time matters, distinguish:

• Publication date.
• Event date.
• Data coverage period.
• Effective date or version date.
• Last updated date.

Precision should match the request:

• "today" or "just released": day-level verification.
• "this week": week-level verification.
• "latest" or "recent": usually month-level verification.
• "trend" or "this year": period and data-window verification.

Treat undated pages, stale statistics, superseded versions, and derivative articles cautiously.

Evidence Thresholds

Use these defaults unless the user asks for stricter standards:

• Low-stakes background fact: one credible source is enough.
• Important factual claim: two independent credible sources when possible.
• Numeric/statistical claim: source, timeframe, population/scope, and measurement method.
• High-stakes claim: primary source plus independent confirmation where possible.
• Contested claim: present competing interpretations and why sources disagree.
• Recommendation: link the recommendation to explicit evidence and assumptions.

If a threshold cannot be met, say so and lower certainty.

Failure And Access Limits

If research is limited by tool access, paywalls, missing PDFs, blocked pages, unavailable search, weak source quality, or time constraints:

• Say what could and could not be checked.
• Do not imply unread sources were read.
• Prefer "the available evidence suggests" over stronger language.
• List the specific follow-up evidence that would change confidence.

Internal Evidence Log

Track important evidence internally using this shape. Do not show it unless useful.

Claim:
Source:
Source type:
Date / data period / version:
What it supports:
Limitations or bias:
Confidence: high | medium | low
Status: consensus | disputed | tentative

For complex research, keep a small contradiction list:

Issue:
Source A says:
Source B says:
Likely reason for difference:
How final answer handles it:

Output Formats

Choose the format that best fits the user.

Evidence Brief

Use for fast mode.

## Answer

## Key Findings

## Evidence And Caveats

## References

Decision Memo

Use when the user needs to choose or act.

## Recommendation

## Rationale

## Options Compared

## Risks And Unknowns

## References

Research Report

Use for standard mode.

## Research Question

## Executive Summary

## Key Findings

## Detailed Analysis

## Areas Of Consensus

## Areas Of Disagreement Or Uncertainty

## Implications

## References

Comparison Matrix

Use when alternatives are central.

## Short Answer

| Option | Strengths | Weaknesses | Evidence | Best Fit |
|---|---|---|---|---|

## Notes

## References

Adapt headings freely. Do not force an academic paper shape when a memo, table, or short brief is more useful. Translate template headings into the user's language unless the user asks for a different language.

Citation Rules

Every key finding needs traceable support.

Place citations close to the claims they support. Do not rely only on a final References list for central findings, numbers, dates, legal/policy conclusions, version claims, medical claims, financial claims, or recommendations.

Use the citation style that is clearest for the context:

• Markdown links inline for short chat answers.
• Footnotes for long reports.
• A table column for comparison matrices.
• A compact source list for briefs.

For each important source, include enough metadata to audit it:

• Author or institution.
• Title or page name.
• Publication/update/effective date when available.
• URL or retrievable identifier.
• What the source establishes.
• Major limitation when relevant.

Do not overquote. Quote only short passages when wording matters; otherwise paraphrase and cite.

If a claim is based on synthesis rather than a single source, cite the main sources and say that the conclusion is synthesized from them.

Language And Math

Match the user's language and expertise.

• Translate output headings into the user's language unless preserving the English term improves precision.
• Explain technical terms when the target audience needs it.
• Skip basic definitions for expert audiences.
• Use analogies only when they reduce cognitive load.
• Prefer plain language, active voice, and direct conclusions.
• Use LaTeX for real formulas. Ordinary percentages, ratios, and simple comparisons can stay in normal text.

Completion Checklist

Before finalizing, verify:

• The answer directly addresses the research question.
• Scope, assumptions, and time window are clear.
• Fresh evidence was gathered or a source-access limitation was disclosed.
• Central claims have citations.
• Citations are near the claims they support, not only in a final source list.
• Source quality is proportional to the stakes.
• Important dates, versions, numbers, and roles were checked.
• Disagreement and uncertainty are visible.
• Recommendations are tied to evidence and assumptions.
• High-stakes advice is conditional and does not exceed the available context.
• Access limits or weak evidence are disclosed.
• The output format serves the user's purpose instead of showcasing process.

Common Failure Modes

Avoid:

• Turning deep research into a link dump.
• Answering from memory after activating deep research.
• Writing a long background section before answering.
• Treating snippet text as if the source was read.
• Citing sources only at the end while leaving key findings unsupported in context.
• Counting repeated secondary reports as independent confirmation.
• Ignoring primary sources.
• Hiding uncertainty to sound decisive.
• Using fast mode to bypass validation.
• Giving individualized medical, legal, financial, compliance, or safety advice without the needed context.
• Forgetting dates on current or versioned claims.
• Forcing footnotes or academic headings where a brief or matrix would be clearer.