support-with-evidence

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Support With Evidence

用证据佐证

Take a body of text — an argument, a set of claims, a thesis, bullet points — and extract the falsifiable ideas from it. Then go out and find real evidence for or against each one. The goal is not confirmation bias — it's an honest evidence audit. If the evidence supports the idea, you'll see it. If the evidence contradicts the idea, you'll see that too.

接收一段文本——可以是论点、一组主张、一个论题或项目符号列表——从中提取所有可证伪的观点。然后搜索真实的证据来支持或反驳每个观点。我们的目标不是确认偏差，而是进行诚实的证据审计。如果证据支持该观点，会如实呈现；如果证据与该观点相矛盾，也会如实呈现。

When to Use

使用场景

User has claims, predictions, or assertions and wants to know what evidence exists
User asks "is this true?" or "can you find evidence for this?" or "support this with evidence"
User wants to fact-check or ground-truth a set of ideas before publishing or acting
User has a thesis and wants to know which parts are empirically supported and which are speculation

Do NOT use for: stress-testing a thesis for logical weaknesses (use stress-test), evaluating prompts (use think-critically), or surfacing insights from data (use surface-insight).

用户持有主张、预测或断言，希望了解相关的现有证据
用户询问“这是真的吗？”“你能为这个观点找证据吗？”或“用证据支撑这个观点”
用户希望在发布内容或采取行动前，对一组观点进行事实核查或真实性验证
用户有一个论题，想知道哪些部分有实证支持，哪些部分属于推测

请勿用于：针对逻辑缺陷检验论题（请使用stress-test技能）、评估提示词（请使用think-critically技能）或从数据中挖掘洞见（请使用surface-insight技能）。

The Honesty Rule

诚实原则

This skill's job is to find evidence — not to confirm what the user hopes is true. The single most important rule:

Report what you find, not what the user wants to hear
If evidence contradicts the idea, say so clearly and present the counter-evidence
If you cannot find meaningful evidence either way, say so — "no evidence found" is a legitimate and valuable output
Never fabricate, hallucinate, or overstate evidence — cite real sources or state that the search was inconclusive
ENFORCEMENT: Every evidence bullet must include a source — a named study, dataset, organization, publication, or URL. If you cannot name a source, the bullet is not evidence. Discard it.

本技能的职责是寻找证据，而非确认用户希望为真的内容。最重要的规则如下：

报告实际找到的内容，而非用户想听到的内容
如果证据与观点相矛盾，需清晰说明并呈现反证
如果无法找到支持或反对的有效证据，需如实告知——“未找到相关证据”是合理且有价值的输出
绝不能编造、虚构或夸大证据——需引用真实来源，或说明搜索无果
强制执行要求：每个证据条目必须包含来源——如指定的研究、数据集、机构、出版物或URL。如果无法命名来源，该条目不能作为证据，需予以舍弃。

Process

流程

Phase 1: Claim Extraction (Output)

阶段1：主张提取（输出）

Read the user's input. Extract all falsifiable ideas — claims that could in principle be shown true or false with evidence.

FALSIFIABILITY GATE: For each candidate claim, apply this test: "What observable evidence would confirm or deny this?" If you cannot answer that question, the claim is not falsifiable — it's an opinion, value judgment, or unfalsifiable abstraction. Do not research it. Instead, list it separately as "Non-falsifiable (skipped)" with a one-sentence explanation of why.

SHARPENING: If a claim is close to falsifiable but too vague as stated, sharpen it into a testable version. Show both the original wording and your sharpened version. Ask the user to confirm only if the sharpening substantially changes the meaning. Otherwise, proceed with the sharpened version and note what you did.

HARD RULE: Extract 1-10 claims. If the input contains more than 10 falsifiable claims, keep the 10 most substantive. If it contains zero, state: "No falsifiable claims found in the input — nothing to research."

Output:

undefined

阅读用户输入内容，提取所有可证伪的观点——即原则上可以用证据证明为真或假的主张。

可证伪性检验：对于每个候选主张，进行如下测试：“哪些可观察的证据可以证实或否定该主张？”如果无法回答这个问题，那么该主张不具备可证伪性——它属于观点、价值判断或无法证伪的抽象概念。无需对其进行研究，而是将其单独列为“不可证伪（已跳过）”，并附上一句解释原因。

优化处理：如果某个主张接近可证伪性，但表述过于模糊，需将其优化为可检验的版本。同时展示原始表述和优化后的版本。只有当优化后的内容大幅改变原意时，才需要请求用户确认；否则，可直接使用优化后的版本，并说明优化内容。

硬性规则：提取1-10个主张。如果输入内容包含超过10个可证伪的主张，保留其中最具实质性的10个。如果没有可证伪的主张，需说明：“输入内容中未找到可证伪的主张——无内容可研究。”

输出格式：

undefined

Extracted Claims

提取的主张

#	Claim	Falsifiable?
1	[claim as stated or sharpened]	Yes
2	[claim]	Yes
...	...	...
N	[non-falsifiable claim]	No — [reason]

undefined

序号	主张内容	是否可证伪
1	[原始或优化后的主张]	是
2	[主张内容]	是
...	...	...
N	[不可证伪的主张]	否——[原因]

undefined

Phase 2: Deep Research (Silent)

阶段2：深度研究（无输出）

For each falsifiable claim, conduct deep research using web search tools. This is not a surface-level check — dig into it.

Research procedure per claim:

Search broadly. Use multiple search queries — rephrase the claim, search for supporting evidence, then search for contradicting evidence. Do not stop at the first result.
Seek primary sources. Prefer peer-reviewed studies, government data, established datasets, named expert opinions, and reputable journalism over blog posts, opinion pieces, or anonymous forums.
Check both sides. For every claim, actively search for evidence AGAINST it, not just for it. If the first few results all confirm, search harder for disconfirming evidence — and vice versa.
Assess source quality. A single blog post is not equivalent to a meta-analysis. Weight evidence by source credibility.
Note recency. Evidence from 2024-2026 is stronger than evidence from 2015 for claims about current states of affairs. Flag when evidence is dated.

RESEARCH DEPTH: Spend meaningful effort on each claim. Use at least 2-3 distinct search queries per claim. Read actual results, not just titles. If initial searches are inconclusive, try different angles — related statistics, adjacent research, expert commentary.

Do not output Phase 2 reasoning.

针对每个可证伪的主张，使用网络搜索工具进行深度研究。这不是表面级别的检查，而是深入挖掘。

每个主张的研究流程：

广泛搜索：使用多个搜索查询词——重新表述主张，先搜索支持证据，再搜索反对证据。不要在得到第一个结果后就停止搜索。
优先选择原始来源：优先选择同行评审研究、政府数据、已确立的数据集、知名专家观点和权威新闻报道，而非博客文章、评论文章或匿名论坛内容。
兼顾正反两方：对于每个主张，主动搜索反对它的证据，而非仅搜索支持证据。如果前几个结果均为支持证据，需更努力地寻找反证——反之亦然。
评估来源质量：一篇博客文章不能与元分析等同。需根据来源可信度对证据进行加权。
注意时效性：对于有关当前状况的主张，2024-2026年的证据比2015年的证据更具说服力。如果证据过时，需标注说明。

研究深度：针对每个主张投入足够精力。每个主张至少使用2-3个不同的搜索查询词。阅读实际搜索结果内容，而非仅看标题。如果初始搜索无果，尝试不同角度——相关统计数据、相邻研究、专家评论等。

无需输出阶段2的推理过程。

Phase 3: Evidence Assembly & Rating (Output)

阶段3：证据整理与评级（输出）

For each falsifiable claim, present the evidence found and rate it.

Evidence Strength Scale:

Very Strong: Multiple independent, high-quality sources directly confirm. Peer-reviewed research, replication, broad expert consensus, or authoritative data.
Strong: Credible sources with direct evidence. Established facts, solid data, or well-sourced reporting from reputable outlets.
Moderate: Some evidence exists but with caveats — limited sources, indirect evidence, small sample sizes, or some conflicting data.
Weak: Thin evidence — anecdotal, from low-quality sources, speculative, or a single unreplicated finding.

Direction:

Supported: Evidence found predominantly in favor of the claim.
Contested: Mixed evidence — meaningful evidence exists on both sides.
Unsupported: Little or no evidence found either way.
Contradicted: Evidence found predominantly against the claim.

Output format per claim:

undefined

针对每个可证伪的主张，呈现找到的证据并进行评级。

证据强度等级：

极强：多个独立、高质量来源直接证实。包括同行评审研究、重复实验、广泛的专家共识或权威数据。
强：可靠来源提供直接证据。包括已确立的事实、可靠数据或权威媒体的详实报道。
中等：存在一些证据，但存在局限性——来源有限、间接证据、样本量小或存在一些相互矛盾的数据。
弱：证据单薄——如轶事证据、低质量来源、推测性内容或单一未重复的研究发现。

证据倾向：

支持：找到的证据主要支持该主张。
有争议：证据混合——支持和反对的证据均有重要依据。
无支持：未找到支持或反对的有效证据。
反驳：找到的证据主要反对该主张。

每个主张的输出格式：

undefined

Claim [#]: [claim text]

主张[#]：[主张文本]

Evidence for:

[Evidence point with source name/URL] — [1 sentence explaining relevance]
[Evidence point with source] — [1 sentence] ...

Evidence against:

[Evidence point with source name/URL] — [1 sentence explaining relevance]
[Evidence point with source] — [1 sentence] ...

Assessment: [2-3 sentences: your honest reading of what the evidence says. Does it support, contradict, or leave the claim uncertain? What's the strongest evidence on each side? What evidence is missing that would be decisive?]


HARD RULES:
- Include at least 1 evidence bullet per claim (for or against). If you truly found nothing, state "No meaningful evidence found in search" and rate as Unsupported/Weak.
- Every evidence bullet must name its source. No unnamed evidence.
- If all evidence points in one direction, actively state what counter-evidence you searched for and didn't find — this is more honest than pretending the search was balanced.
- Order evidence bullets by strength (strongest first).

证据倾向：[支持 | 有争议 | 无支持 | 反驳] 证据强度：[极强 | 强 | 中等 | 弱]

支持证据：

[证据内容及来源名称/URL] —— [1句话说明相关性]
[证据内容及来源] —— [1句话说明相关性] ...

反对证据：

[证据内容及来源名称/URL] —— [1句话说明相关性]
[证据内容及来源] —— [1句话说明相关性] ...

评估： [2-3句话：如实解读证据所说明的内容。证据是支持、反驳还是让主张存疑？正反两方最有力的证据是什么？哪些关键证据缺失？]


**硬性规则**：
- 每个主张至少包含1个证据条目（支持或反对）。如果确实未找到任何证据，需说明“搜索未找到有效证据”，并评级为“无支持/弱”。
- 每个证据条目必须注明来源。未注明来源的内容不能作为证据。
- 如果所有证据均倾向同一方向，需主动说明已搜索过的反证类型但未找到——这比假装搜索平衡更诚实。
- 证据条目需按强度排序（最强的排在最前面）。

Phase 4: Summary Scorecard

阶段4：汇总评分卡

After all claims, output a summary table:

undefined

完成所有主张的分析后，输出汇总表格：

undefined

Evidence Scorecard

证据评分卡

#	Claim	Direction	Strength	Sources
1	[short form]	Supported	Strong	4
2	[short form]	Contradicted	Very Strong	6
...	...	...	...	...

Overall: [X] of [Y] claims supported, [Z] contested, [W] contradicted, [V] unsupported.


If any claims were non-falsifiable, add a note:

[N] claims were non-falsifiable and excluded from research.

undefined

序号	主张（简写）	证据倾向	证据强度	来源数量
1	[简写内容]	支持	强	4
2	[简写内容]	反驳	极强	6
...	...	...	...	...

总体情况： [X]个主张得到支持，[Z]个存在争议，[W]个被反驳，[V]个无支持证据。


如果存在不可证伪的主张，需添加说明：

[N]个主张不可证伪，未纳入研究范围。

undefined

Phase 5: Bottom Line

阶段5：核心结论

End with a short synthesis — 2-4 sentences max. What's the overall evidence picture? Which claims are on solid ground and which are shaky? If the user's input was a coherent argument, does the evidence support the argument as a whole, or only parts of it?

undefined

最后用简短的总结收尾——最多2-4句话。整体证据情况如何？哪些主张依据充分，哪些主张站不住脚？如果用户输入是一个连贯论点，证据是否整体支持该论点，还是仅支持部分内容？

undefined

Bottom Line

核心结论

[2-4 sentences: honest synthesis of the evidence picture]

---

[2-4句话：对证据情况的诚实总结]

---

Quality Standards

质量标准

The quality bar is the evidence test: could someone follow your citations and verify what you reported?

Fail — No source. An evidence bullet without a named source. REJECT.
Fail — Fabricated. Evidence that doesn't correspond to a real source. REJECT.
Pass — Named source, relevant. A real source that speaks to the claim. MINIMUM.
Good — Primary source. A study, dataset, or authoritative report directly on point.
Strong — Multiple corroborating sources. Independent sources converging on the same finding.
Exceptional — Decisive evidence. A single source so authoritative it settles the question (e.g., a large meta-analysis, official government statistics, or a landmark ruling).

Aim for "Good" or above on at least half of evidence bullets.

质量评判标准为证据可验证性：他人能否通过你的引用验证你所报告的内容？

不合格——无来源：证据条目未注明来源。予以驳回。
不合格——编造内容：证据与真实来源不符。予以驳回。
合格——注明相关来源：真实且与主张相关的来源。最低要求。
良好——原始来源：与主张直接相关的研究、数据集或权威报告。
优秀——多个相互佐证的来源：独立来源得出相同结论。
卓越——决定性证据：单个来源权威性极高，足以解决问题（如大型元分析、官方政府统计数据或具有里程碑意义的裁决）。

至少一半的证据条目需达到“良好”或以上标准。

Anti-Patterns

反模式

Confirmation bias. Only searching for evidence that supports the claim. The #1 failure mode. Always search both directions.
Source inflation. Treating a blog post citing a study as equivalent to the study itself. Trace to primary sources.
False balance. Presenting one weak blog post against a claim as equal to five peer-reviewed studies for it. Weight evidence honestly.
Fabricated citations. Inventing studies or statistics that don't exist. If you're unsure a source exists, say "based on search results" and describe what you found without inventing specifics.
Vague sourcing. "Studies show" or "experts say" without naming them. Name the study. Name the expert.
Recency neglect. Citing a 2012 study for a claim about 2025 market conditions without noting the gap.
Strength inflation. Rating weak evidence as strong because it supports the claim. Calibrate honestly.
Exhaustive but shallow. Listing 15 low-quality sources instead of 3 high-quality ones. Quality over quantity.

确认偏差：仅搜索支持主张的证据。这是最常见的失败模式。务必兼顾正反两方搜索。
来源膨胀：将引用研究的博客文章等同于研究本身。需追溯到原始来源。
虚假平衡：将一篇反对主张的薄弱博客文章与五篇支持主张的同行评审研究视为同等重要。需诚实地对证据进行加权。
编造引用：编造不存在的研究或统计数据。如果不确定来源是否存在，需说明“基于搜索结果”，并描述找到的内容，不要编造细节。
模糊来源：使用“研究表明”或“专家认为”等表述但未指明具体来源。需指明研究名称或专家姓名。
忽视时效性：引用2012年的研究来支持2025年的市场状况主张，却未说明时间差距。
强度夸大：因证据支持主张而将弱证据评级为强证据。需诚实地校准评级。
面面俱到但流于表面：列出15个低质量来源，而非3个高质量来源。质量优先于数量。

Key Principles

核心原则

Honesty over comfort. Report what the evidence says, not what the user wants to hear. A well-sourced "your claim is contradicted" is more valuable than a vague "seems plausible."
Sources or silence. Every evidence bullet names its source. Unsourced claims are not evidence — they're opinion.
Both directions. For every claim, search for AND against. Asymmetric research is dishonest research.
Strength calibration. A blog post is not a meta-analysis. Rate evidence by what it actually is, not what you wish it were.
Less is more. Three well-sourced evidence points beat ten weakly sourced ones.
Transparent gaps. When evidence is missing or inconclusive, say so. "We don't know" is a finding.

诚实优先于讨好：报告证据所显示的内容，而非用户想听到的内容。来源详实的“你的主张被反驳”比模糊的“看似合理”更有价值。
要么注明来源，要么保持沉默：每个证据条目必须注明来源。未注明来源的主张不是证据，只是观点。
兼顾正反两方：对于每个主张，既要搜索支持证据，也要搜索反对证据。不对称研究是不诚实的研究。
如实校准强度：博客文章不等同于元分析。需根据证据的实际情况进行评级，而非按期望评级。
少而精：三个来源详实的证据条目胜过十个来源薄弱的条目。
透明说明空白：当证据缺失或无结果时，需如实告知。“我们不知道”也是一种发现。

Input

输入内容

[User provides text containing claims, ideas, or assertions below]

[用户在下方提供包含主张、观点或断言的文本]