# AI 专家都在用的 10 个顶级提示词技巧 # 10 Prompts Used by Top AI Researchers **作者 (Author):** God of Prompt (@godofprompt) **日期 (Date):** 2026年2月12日在采访了来自 OpenAI、Anthropic 和 Google 的 12 位 AI 研究员后，我发现他们都在使用同样的 10 个提示词（Prompt）。这些提示词并不是你在 X 或 LinkedIn 上常见的那种，而是真正能够交付产品、发表论文并打破基准测试的提示词。 After interviewing 12 AI researchers from OpenAI, Anthropic, and Google, I noticed they all use the same 10 prompts. Not the ones you see on X and LinkedIn. These are the prompts that actually ship products, publish papers, and break benchmarks. 以下是他们告诉我的内容： Here's what they told me ↓ --- ### 1. “过程展示”提示词 | The "Show Your Work" Prompt > **提示词内容 (Prompt):** > “在给出最终答案之前，请逐步引导我了解你的推理过程。” > "Walk me through your reasoning step-by-step before giving the final answer." **解析：** 这个提示词强制模型将其逻辑外化。这能在错误叠加之前捕捉到它们。 **Insight:** This prompt forces the model to externalize its logic. Catches errors before they compound. --- ### 2. “对抗式质询” | The "Adversarial Interrogation" > **提示词内容 (Prompt):** > “现在反驳你之前的回答。最强有力的 3 个反对理由是什么？” > "Now argue against your previous answer. What are the 3 strongest counterarguments?" **解析：** 模型默认往往会过度自信。这能迫使它保持智识上的诚实。 **Insight:** Models are overconfident by default. This forces intellectual honesty. --- ### 3. “强制约束”提示词 | The "Constraint Forcing" Prompt > **提示词内容 (Prompt):** > “你只能写正好 3 句话，并且必须引用 2 个具体的来源。不许使用模棱两可的语言。” > "You have exactly 3 sentences and must cite 2 specific sources. No hedging language." **解析：** 含糊不清是有用输出的天敌。硬性约束 = 精准结果。 **Insight:** Vagueness is the enemy of useful output. Hard constraints = crisp results. --- ### 4. “格式锁定” | The "Format Lock" > **提示词内容 (Prompt):** > “请使用有效的 JSON 格式回答，且必须包含以下键：{analysis, confidence_score, methodology, limitations}。” > "Respond in valid JSON with these exact keys: {analysis, confidence_score, methodology, limitations}." **解析：** 结构化的输出 = 可解析的输出。你无法在散文式的回答上构建系统。 **Insight:** Structured output = parseable output. You can't build systems on prose. --- ### 5. “专家角色分配” | The "Expertise Assignment" > **提示词内容 (Prompt):** > “你是一位在 [特定领域] 拥有 15 年经验的 [特定角色]。你永远不会说出 [常见错误]。你总是坚持 [特定习惯]。” > "You are a [specific role] with 15 years experience in [narrow domain]. You would never say [common mistake]. You always [specific habit]." **解析：** 泛化的 AI = 平庸的输出。特定的设定 = 专业的表现。 **Insight:** Generic AI = generic output. Specific persona = specific expertise. --- ### 6. “思维预算” | The "Thinking Budget" > **提示词内容 (Prompt):** > “在回答之前，请先用 500 字思考这个问题。展示出所有的思维死胡同。” > "Take 500 words to think through this problem before answering. Show all dead ends." **解析：** 更多的 Token（字符消耗）意味着更好的推理。死胡同能揭示模型的理解深度。 **Insight:** More tokens = better reasoning. Dead ends reveal model understanding. --- ### 7. “对比协议” | The "Comparison Protocol" > **提示词内容 (Prompt):** > “请从这 5 个维度对比方案 A 和方案 B：[速度、准确性、成本、复杂性、维护性]。请使用表格展示。” > "Compare approach A vs B across these 5 dimensions: [speed, accuracy, cost, complexity, maintenance]. Use a table." **解析：** 强制进行结构化分析。在技术决策中，表格优于段落。 **Insight:** Forces structured analysis. Tables > paragraphs for technical decisions. --- ### 8. “不确定性量化” | The "Uncertainty Quantification" > **提示词内容 (Prompt):** > “请为每个观点打分（0-100 分表示你的信心）。将任何低于 70 分的内容标记为推测性的。” > "Rate your confidence 0-100 for each claim. Flag anything below 70 as speculative." **解析：** 当幻觉被贴上标签时，它们的危险性就会降低。信心评分是强制性的。 **Insight:** Hallucinations are less dangerous when labeled. Confidence scoring is mandatory. --- ### 9. “边界情况猎人” | The "Edge Case Hunter" > **提示词内容 (Prompt):** > “哪 5 种输入会导致这个方案失效？请站在对抗者的角度思考。” > "What are 5 inputs that would break this approach? Be adversarial." **解析：** 模型往往会遗漏人类能捕捉到的边界情况。强制对抗式思维能揭示系统的脆弱性。 **Insight:** Models miss edge cases humans would catch. Forcing adversarial thinking reveals brittleness. --- ### 10. “验证链” | The "Chain of Verification" > **提示词内容 (Prompt):** > “首先，回答问题。其次，列出你的答案可能错误的 3 个原因。第三，核实每个担忧并更新你的答案。” > "First, answer the question. Second, list 3 ways your answer could be wrong. Third, verify each concern and update your answer." **解析：** 在提示词中内置自我修正机制。让模型自己修复自己的错误。 **Insight:** Self-correction built into the prompt. Models fix their own mistakes. --- | 提示词策略 (Prompt Type) | 来源 (Source) | 提示词原文与翻译 (Prompt Example & Translation) | | :--- | :--- | :--- | | **展示推理过程**
Show your work | **Anthropic** | "Walk me through your reasoning step-by-step before giving the final answer."
**翻译：** “在给出最终答案之前，请逐步引导我了解你的推理过程。” | | **“对抗性审问”**
The "Adversarial Interrogation" | **OpenAI** | "Now argue against your previous answer. What are the 3 strongest counterarguments?"
**翻译：** “现在请反驳你之前的回答。最强有力的三个反论点是什么？” | | **“约束强制”提示词**
The "Constraint Forcing" Prompt | **Google** | "You have exactly 3 sentences and must cite 2 specific sources. No hedging language."
**翻译：** “你只能写 3 句话，并且必须引用 2 个具体来源。禁止使用模棱两可的套话。” | | **“格式锁定”**
The "Format Lock" | **OpenAI** | "Respond in valid JSON with these exact keys: {analysis, confidence_score, methodology, limitations}."
**翻译：** “请以有效的 JSON 格式回复，并包含以下确切的键：{分析, 置信度得分, 方法论, 局限性}。” | | **“专家角色分配”**
The "Expertise Assignment" | **Anthropic** | "You are a [specific role] with 15 years experience in [narrow domain]. You would never say [common mistake]. You always [specific habit]."
**翻译：** “你是一位在 [特定领域] 拥有 15 年经验的 [特定角色]。你绝不会说 [常见错误]。你总是坚持 [特定习惯]。” | | **“思维预算”**
The "Thinking Budget" | **Google** | "Take 500 words to think through this problem before answering. Show all dead ends."
**翻译：** “在回答之前，请先用 500 字深入思考这个问题。请展示出所有行不通的思路（死胡同）。” | | **“对比协议”**
The "Comparison Protocol" | **OpenAI** | "Compare approach A vs B across these 5 dimensions: [speed, accuracy, cost, complexity, maintenance]. Use a table."
**翻译：** “请从这 5 个维度对比方案 A 和方案 B：[速度、准确性、成本、复杂性、维护性]。请使用表格呈现。” | | **“不确定性量化”**
The "Uncertainty Quantification" | **Anthropic** | "Rate your confidence 0-100 for each claim. Flag anything below 70 as speculative."
**翻译：** “请为您提出的每项主张评分（0-100 分）。任何低于 70 分的内容请标注为‘推测性内容’。” | | **“极端情况猎手”**
The "Edge Case Hunter" | **OpenAI** | "What are 5 inputs that would break this approach? Be adversarial."
**翻译：** “哪 5 种输入会导致这个方案失效？请站在对抗者的角度思考。” | | **“验证链”**
The "Chain of Verification" | **Anthropic** | "First, answer the question. Second, list 3 ways your answer could be wrong. Third, verify each concern and update your answer."
**翻译：** “第一步，回答问题。第二步，列出该答案可能出错的 3 种情况。第三步，核实每一个疑虑并更新你的答案。” | --- ### 术语说明 (Glossary) * **Hedging language**: 指“可能”、“大概”、“在某种程度上”等为了逃避责任或不确定而使用的模糊语言。在提示词中禁止它可以让模型给出更果断、清晰的回答。 * **Adversarial**: 对抗性。要求模型模拟对手或批评者的视角，以发现逻辑漏洞。 * **Dead ends**: 死胡同。指在解决问题过程中尝试过但失败的思路，展示这些可以帮助用户理解模型排除掉的错误路径。 * **Chain of Verification (CoVe)**: 验证链。一种自我修正技术，通过让模型审视自己的错误来提高准确度。