Back to Details

deep-research

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

Deep Research Guide

深度研究指南

Produce a trustworthy synthesized answer, not a pile of links. The output of research is a decision-grade brief: a clear claim, the evidence behind it, citations, and an honest map of what is still unknown. Speed without verification is a liability — a confident wrong answer costs more than a slow right one.
Core principle: default to skeptical. Treat every key claim as guilty until corroborated. Your job is not to confirm a hypothesis — it's to find what would break it.
生成可信的综合答案，而非一堆链接。研究的输出是一份决策级简报：清晰的结论、背后的证据、引用，以及对未知信息的如实说明。缺乏验证的速度是一种隐患——一个自信的错误答案比一个缓慢的正确答案代价更高。
核心原则：默认持怀疑态度。在得到证实之前，将每个关键主张视为不可信。你的任务不是证实一个假设，而是找到能推翻它的依据。

Workflow — plan → gather → verify → synthesize

工作流程——规划 → 收集 → 验证 → 综合

Run these as distinct phases. Don't start writing the report while you're still gathering, and don't gather before the question is scoped.
分阶段执行这些步骤。不要在收集信息的同时就开始撰写报告，也不要在问题界定清楚前就开始收集信息。

1. Plan — scope the question

1. 规划——界定问题范围

Before searching, pin down what's actually being asked. Most bad research answers a question nobody asked.
  • State the question in one sentence. If you can't, it's not scoped yet.
  • Name the decision it serves. "Which X should we pick?" needs different evidence than "Is X true?". The decision sets the bar for confidence.
  • List the answer's shape. A number? A recommendation? A comparison table? A yes/no with caveats? Knowing the shape tells you when you're done.
  • Set boundaries: time window (recency that matters), geography, scale, definitions of fuzzy terms.
Narrow an underspecified question before spending effort. If the ask is "what should I buy / which tool / is this a good idea" without budget, use-case, constraints, or context, ask 2–3 clarifying questions first. Researching the wrong question thoroughly is still wrong.
Then decompose into sub-questions — the 3–7 things that must each be answered for the whole to hold. Research the sub-questions; assemble the answer.
在搜索之前，明确实际要解决的问题。大多数糟糕的研究都是在回答一个没人关心的问题。
  • 用一句话表述问题。如果做不到，说明问题范围还未明确。
  • 明确服务的决策。“我们应该选择哪款X？”需要的证据与“X是否属实？”不同。决策决定了置信度的标准。
  • 列出答案的形式。是数字？推荐？对比表格？带附加条件的是/否？明确形式能让你知道何时完成研究。
  • 设定边界：时间范围（相关的时效性）、地域、规模、模糊术语的定义。
在投入精力前先缩小模糊的问题范围。如果需求是“我应该买什么/选哪个工具/这个主意好不好”但没有预算、使用场景、限制条件或背景信息，先提出2-3个澄清问题。彻底研究错误的问题依然是错误的。
然后将问题分解为子问题——需要逐一解答的3-7个关键问题，只有这些子问题都得到解答，整体问题才能解决。先研究子问题，再整合答案。

2. Gather — fan out across sources

2. 收集——多渠道信息拓展

  • Cast wide before going deep. Run several differently-worded queries; don't anchor on the first source's framing. Search for the counter-claim too ("X is overrated", "problems with X") — not just confirmation.
  • Go to primary sources. Prefer the original study, filing, spec, dataset, or official doc over an article summarizing it. Summaries drift; numbers get garbled in retelling.
  • Triangulate. A claim is only as strong as the number of independent sources that confirm it. Three outlets all citing one press release is one source, not three.
  • Capture as you go: for each fact, note the source, the date, and a direct quote/figure. You cannot cite what you didn't record.
  • 先广泛搜索再深入研究。尝试几种不同措辞的查询；不要局限于第一个来源的框架。也要搜索反主张（“X被高估了”“X的问题”）——不要只找支持自己观点的信息。
  • 优先使用原始来源。比起总结性文章，更倾向于原始研究、文件、规范、数据集或官方文档。总结会出现偏差；数据在转述中会被歪曲。
  • 交叉验证。一个主张的可信度取决于独立来源的数量。三个媒体都引用同一份新闻稿，只能算一个来源，而非三个。
  • 随时记录：对于每个事实，记录来源、日期和直接引用/数据。没有记录的内容无法引用。

3. Verify — adversarial check (the step people skip)

3. 验证——反向核查（人们常跳过的步骤）

For each key claim (the ones the conclusion rests on), actively try to refute it:
  • Find the origin. Trace the claim to its source. Where did this number actually come from? Who measured it, how, and when?
  • Look for the strongest disagreement. Who says the opposite, and why? A claim you can't find any dissent on is either settled or you haven't looked hard enough.
  • Check the math and the units. Percentages without a base, totals that don't add up, growth rates with no time frame, and apples-to-oranges comparisons are the common tells.
  • Test recency. Is this still true? Prices, rankings, "fastest/largest/only" claims, and policy facts decay. A correct 2019 fact can be a wrong 2026 answer.
  • Watch for self-interest. Vendor benchmarks, sponsored studies, and anything selling something get an extra round of scrutiny.
If a claim survives a genuine attempt to break it, it's load-bearing. If it doesn't, demote it to "reported but unverified" or drop it.
对于每个关键主张（结论所依赖的主张），主动尝试推翻它：
  • 追溯起源。找到主张的源头。这个数据究竟来自哪里？谁测量的、如何测量的、何时测量的？
  • 寻找最强烈的反对意见。谁持相反观点，原因是什么？如果找不到任何异议，要么这个主张是既定事实，要么你还没足够深入地搜索。
  • 检查数据和单位。没有基数的百分比、加总不符的总数、没有时间范围的增长率，以及跨类别比较是常见的问题信号。
  • 验证时效性。这个信息现在还成立吗？价格、排名、“最快/最大/唯一”的主张以及政策事实都会随时间失效。2019年正确的事实在2026年可能就是错误的答案。
  • 警惕利益相关方。厂商基准测试、赞助研究以及任何带有销售目的的内容都需要额外一轮审查。
如果一个主张在真正的推翻尝试后依然成立，它就是支撑结论的关键。如果不成立，就将其降级为“已报告但未验证”或直接舍弃。

4. Synthesize — write the brief

4. 综合——撰写简报

  • Lead with the answer. First line: the conclusion / recommendation. Decision-makers read top-down and may stop after the first paragraph — make it count.
  • Then the why, structured by sub-question, each point carrying its citation.
  • Separate fact from inference explicitly (see below).
  • Close with confidence + unknowns.
  • 开门见山给出答案。第一句：结论/建议。决策者自上而下阅读，可能看完第一段就停止——所以要让开头有价值。
  • 然后说明原因，按子问题结构化呈现，每个观点都附带引用。
  • 明确区分事实与推论（见下文）。
  • 结尾说明置信度与未知信息。

Source-credibility checklist

来源可信度 checklist

Score each source before you lean on it:
  • Primary or secondary? Original data/document > reporting on it > commentary on the reporting.
  • Authority — does the author/org actually have standing on this topic, or are they out of their lane?
  • Recency — is it current enough for a claim that changes over time? Note the date, always.
  • Independence — funded by, owned by, or selling the thing in question? Conflicts bias.
  • Method transparency — can you see how they got the number (sample, methodology, sources), or are you trusting an assertion?
  • Corroboration — do independent sources agree? Outliers need explaining, not silent dropping.
  • Track record — has this source been reliable/retracted before?
Rough hierarchy (context-dependent, not absolute): peer-reviewed studies, official filings/standards, and primary datasets at the top; reputable journalism and expert analysis in the middle; anonymous posts, marketing, and AI-generated content summaries near the bottom. A low-tier source can still be right — it just needs corroboration before it carries weight.
在依赖某个来源之前，先对其评分：
  • 原始来源还是二次来源？ 原始数据/文档 > 基于原始来源的报道 > 对报道的评论。
  • 权威性——作者/机构是否在该领域有话语权，还是超出了他们的专业范围？
  • 时效性——对于随时间变化的主张，它是否足够新？始终标注日期。
  • 独立性——是否由相关事物的投资方、所有者或销售方资助？利益冲突会导致偏见。
  • 方法透明度——你能否看到他们获取数据的方式（样本、方法论、来源），还是只能盲目相信断言？
  • 交叉验证——独立来源是否一致？异常值需要解释，而非忽略。
  • 过往记录——这个来源之前是否可靠，有没有被撤回过内容？
大致层级（取决于上下文，非绝对）：顶级是同行评审研究、官方文件/标准和原始数据集；中间是知名新闻报道和专家分析；底部是匿名帖子、营销内容和AI生成的摘要。低层级来源也可能正确，但在成为支撑依据前需要交叉验证。

Separating fact from inference

区分事实与推论

Be ruthless about which is which; conflating them is how research misleads.
  • Fact — directly stated by a credible source, with a citation. ("Revenue was $4.2M in 2025 [source].")
  • Inference — your reasoning from facts. Label it. ("This implies ~30% YoY growth, assuming the 2024 figure of $3.2M is comparable.")
  • Assumption — something you're taking as given without evidence. Name it so the reader can challenge it. ("Assuming the same accounting basis across years.")
  • Unknown — a gap you couldn't fill. State it; don't paper over it.
Phrases that signal you're doing it right: "according to…", "this suggests…", "I could not find…", "sources disagree on…".
严格区分二者；混淆它们是研究产生误导的原因。
  • 事实——由可信来源直接陈述，并带有引用。（“2025年营收为420万美元[来源]。”）
  • 推论——你从事实中得出的推理。标注清楚。（“假设2024年的320万美元数据具有可比性，这意味着同比增长约30%。”）
  • 假设——你未经证据证实就视为既定的内容。明确说明，以便读者提出质疑。（“假设各年采用相同的会计基准。”）
  • 未知信息——你无法填补的空白。明确说明；不要掩盖。
正确做法的标志性表述：“根据……”“这表明……”“我无法找到……”“来源在……上存在分歧”。

Reporting confidence and unknowns

报告置信度与未知信息

End every brief with an explicit confidence statement. Vague hedging ("seems like") is useless; calibrated confidence is actionable.
  • High — multiple independent primary sources agree; recent; verified the underlying math.
  • Medium — corroborated but with gaps, dated data, or some reliance on secondary sources.
  • Low — single source, conflicting evidence, stale data, or heavy inference. Say so loudly.
Always include a short "What I couldn't verify / what would change this answer" section. Naming the unknowns is a feature: it tells the decision-maker where the risk lives and what to check before betting on it.
每份简报结尾都要有明确的置信度声明。模糊的措辞（“似乎是”）毫无用处；校准后的置信度才具有可操作性。
  • 高——多个独立原始来源一致；时效性强；已验证底层数据。
  • 中——已交叉验证但存在空白、数据过时，或部分依赖二次来源。
  • 低——单一来源、证据冲突、数据陈旧，或大量依赖推论。要明确说明。
始终包含一个简短的**“我无法验证的内容/会改变答案的因素”**部分。指出未知信息是优势：它能告诉决策者风险所在，以及在采取行动前需要核查的内容。

Report template

报告模板

ANSWER:        <the conclusion / recommendation, one or two sentences>
CONFIDENCE:    High / Medium / Low — <one-line why>

KEY FINDINGS
  1. <claim> [source, date]
  2. <claim> [source, date]   (note: sources disagree — see below)
  ...

REASONING / INFERENCE
  <what you concluded from the facts, with assumptions named>

CONTEXT & CAVEATS
  <scope, definitions, anything that bounds the answer>

UNKNOWNS / WHAT WOULD CHANGE THIS
  <gaps you couldn't fill; what to verify before acting>

SOURCES
  [1] <title> — <publisher/author>, <date>, <url> — primary/secondary, why trusted
  ...
ANSWER:        <结论/建议，1-2句话>
CONFIDENCE:    High / Medium / Low — <一句话说明原因>

KEY FINDINGS
  1. <主张> [来源，日期]
  2. <主张> [来源，日期]   （注：来源存在分歧——见下文）
  ...

REASONING / INFERENCE
  <你从事实中得出的结论，明确说明假设>

CONTEXT & CAVEATS
  <范围、定义、任何限制答案的因素>

UNKNOWNS / WHAT WOULD CHANGE THIS
  <你无法填补的空白；采取行动前需要验证的内容>

SOURCES
  [1] <标题> — <发布方/作者>，<日期>，<链接> — 原始/二次来源，可信原因
  ...

Anti-patterns

反模式

  • Confirmation hunting — searching only for what you hope is true. Search the opposite.
  • Citation laundering — three articles citing one origin presented as three sources.
  • Stale-fact trap — quoting a "current" superlative that quietly expired.
  • Burying the answer — making the reader mine paragraphs for the conclusion.
  • False precision — "$4,231,847" from a source that said "about $4M".
  • Unlabeled inference — presenting your reasoning as if it were a sourced fact.
  • Over-hedging — refusing to give an answer when one is warranted. Calibrate, don't dodge.
  • 确认偏误——只搜索你希望为真的内容。也要搜索相反的信息。
  • 引用洗白——将引用同一来源的三篇文章当作三个独立来源。
  • 过时事实陷阱——引用已经失效的“当前”最高级表述。
  • 隐藏答案——让读者在段落中寻找结论。
  • 虚假精确——从一个说“约400万美元”的来源中引用“4,231,847美元”。
  • 未标注推论——将你的推理当作有来源的事实呈现。
  • 过度模糊——在有必要给出答案时拒绝回答。要校准置信度，不要回避。