desk-research（案头研究）

Desk Research

Execute this workflow for any desk-research request.

0) Load methodology checklist (first)

Read references/methodology.md, references/deep-writing-patterns.md, and references/quality-checklist.md and apply all as guardrails.

1) Define the research brief

Write 4 lines before searching:

• Research question (1 sentence)
• Scope (time, geography, industry)
• Must-answer sub-questions (3-6 bullets)
• Output format needed by user

If the question is vague, propose assumptions explicitly and continue.

2) Build a source plan

Collect evidence in this priority order:

1. Primary/official sources (government, regulator, company filings, product docs)
2. Reputable secondary analysis (major research firms, established media)
3. Community signals (forums/social) only as supporting evidence

Require at least 2 independent sources for every key claim.

3) Gather evidence fast

For each sub-question:

• Find 3-8 candidate sources
• Keep the highest-signal sources
• Extract only claim + evidence + date + link

Reject sources that are undated, anonymous, or purely opinionated unless the user asked for sentiment.

4) Score source reliability

Tag each source:

• A = official primary source
• B = credible secondary source
• C = weak/indicative source

When claims conflict, prefer newer A/B sources and explicitly note uncertainty.

5) Synthesize insights

Convert notes into:

• Facts (well-supported)
• Interpretations (reasoned but inferential)
• Unknowns (gaps needing validation)

Never present interpretation as fact.

Source hard rule (critical)

For all final research outputs:

• Every major factual claim must include source + date.
• Every key viewpoint, analytical framework, stage model, or category split must either:

1. cite a source, or 2. be explicitly labeled as the report author's synthesis / working model.

• Do not present an unsupported framework as if it were an established industry fact.
• If a conclusion combines multiple sources, cite the main supporting sources inline or in the same paragraph.

5.5) Deepening loop (mandatory)

Before final delivery, run at least 2 rounds of self-questioning:

Round A — Coverage challenge

• What did I miss by source type, time window, or geography?
• Which category/conclusion is over-dependent on one source?
• What contradicts my current conclusion?

Round B — Decision challenge

• If this conclusion is wrong, what evidence would prove it wrong?
• Which part is descriptive but not decision-useful?
• What next data pull would most change the recommendation?

After each round, update findings and confidence.

6) Deliver in concise structure

Use this exact section order:

1. Core Questions (2 questions)
2. One-sentence Verdict
3. Executive Summary (5-8 bullets)
4. Key Findings by sub-question (with metric anchors)
5. Evidence Table (claim | source | date | reliability)
6. Confidence tags (High/Medium/Low per major claim)
7. Risks / Uncertainty
8. What would falsify this conclusion
9. Next Verification Steps / Todo

Inline citation rule:

• In the body, append source/date after key claims whenever it helps verification.
• For analytical frameworks or stage models, add Source: or Author synthesis based on: explicitly.
• Do not leave major frameworks floating without attribution.

For output shape and compact template, use references/output-template.md.

7) Quality bar before sending

Check all items:

• Every major claim has source/date
• No single-source critical claim
• Time/geography scope matches user ask
• Clear separation of fact vs interpretation
• Actionable takeaway included
• Each promising case uses the full 9-part deep case framework
• Each promising case includes one final case-summary paragraph: what it does / who pays / business model / why pay
• Each key section ends with decision implication (so-what)

8) Case-depth hard rule (for startup/case research)

When the task is startup/use-case research, apply these hard requirements:

• For each promising case, collect at least 3 website evidence snippets (feature/pricing/use-flow)
• Add at least 1 metric anchor from trusted dataset (revenue/MRR/growth)
• Include at least 1 risk point and 1 falsification condition
• Do not submit if any case is only descriptive without judgment