jixiaxue 知识库
evidence · 2026-04-15

Phase 1 开放编码 — T0 Anthropic

/Users/shanfang/Documents/pe/jixiaxuegong/research/提示工程教程/evidence/phase1-coding-T0.md

Phase 1 开放编码 — T0 Anthropic

提取统计

编码表

编号原则/洞察来源文件原文摘录初始分类
A001系统提示词用于在 user turn 之前提供上下文、指令和指南T0-1-ch01”A system prompt is a way to provide context, instructions, and guidelines to Claude before presenting it with a question or task in the ‘User’ turn.”通用
A002精心编写的系统提示词能多方面提升模型表现T0-1-ch01”A well-written system prompt can improve Claude’s performance in a variety of ways, such as increasing Claude’s ability to follow rules and instructions.”通用
A003消息必须以 user 角色开始，且 user/assistant 必须交替出现T0-1-ch01”user and assistant messages MUST alternate, and messages MUST start with a user turn.”Claude专属
A004可以在末尾 assistant 消息中预填文字，让模型从你设定的位置继续T0-1-ch01”You can also put words into a terminal assistant message for Claude to continue from where you left off.”Claude专属
A005像对待新入职员工一样向模型解释任务——它除了你告诉的内容之外没有任何上下文T0-1-ch02”Think of Claude like any other human that is new to the job. Claude has no context on what to do aside from what you literally tell it.”通用
A006清晰提示的黄金法则：把 prompt 给同事看，如果他们困惑，模型也会困惑T0-1-ch02”Show your prompt to a colleague or friend and have them follow the instructions themselves to see if they can produce the result you want. If they’re confused, Claude’s confused.”通用
A007直接要求你想要的行为——“just ask” 原则T0-1-ch02”How do we achieve that? We ask for it!” (关于跳过 preamble)通用
A008要求模型跳过前言，直接进入内容T0-1-ch02”Write a haiku about robots. Skip the preamble; go straight into the poem.”通用
A009如果想让模型做出明确选择，就明确要求它做出选择T0-1-ch02”if you absolutely had to pick one player, who would it be?”通用
A010角色提示可以改变模型的风格、语气和回答方式T0-1-ch03”Priming Claude with a role can improve Claude’s performance in a variety of fields, from writing to coding to summarizing.”通用
A011角色提示可以放在系统提示或 user turn 中T0-1-ch03”Role prompting can happen either in the system prompt or as part of the User message turn.”Claude专属
A012提供目标受众信息可以进一步调整回答T0-1-ch03”provide Claude context on its intended audience… ‘You are a cat’ produces quite a different response than ‘you are a cat talking to a crowd of skateboarders.’”通用
A013角色提示能让模型更擅长数学和逻辑任务T0-1-ch03”Role prompting can also make Claude better at performing math or logic tasks.”通用
A014多种提示工程技巧可以达到相似的结果，鼓励实验找到自己的风格T0-1-ch03”there are many prompt engineering techniques you can use to derive similar results. Which techniques you use is up to you and your preference!”通用
A015用 XML 标签将变量数据与指令分离T0-1-ch04”Wrap the input in XML tags… use specifically XML tags as separators for Claude, as Claude was trained specifically to recognize XML tags as a prompt organizing mechanism.”Claude专属
A016Claude 专门训练过识别 XML 标签作为 prompt 组织机制T0-1-ch04”Claude was trained specifically to recognize XML tags as a prompt organizing mechanism.”Claude专属
A017没有特殊的 magic XML 标签——标签名可以自定义T0-1-ch04”there are no special sauce XML tags that Claude has been trained on that you should use to maximally boost your performance.”Claude专属
A018小细节很重要：拼写和语法错误会影响模型表现T0-1-ch04”small details matter! It’s always worth it to scrub your prompts for typos and grammatical errors. Claude is sensitive to patterns… it’s more likely to make mistakes when you make mistakes, smarter when you sound smart.”通用
A019模型对模式敏感——你犯错它犯错，你聪明它聪明，你傻它傻T0-1-ch04”it’s more likely to make mistakes when you make mistakes, smarter when you sound smart, sillier when you sound silly”通用
A020Prompt 模板化：固定骨架 + 可替换变量，实现重复任务的简化T0-1-ch04”separating the fixed skeleton of the prompt from variable user input, then substituting the user input into the prompt”通用
A021确保模型知道变量的起止位置——用 XML 标签包裹变量T0-1-ch04”make sure Claude knows where variables start and end (vs. instructions or task descriptions)“通用
A022可以要求模型用 XML 标签包裹输出，便于程序化提取T0-1-ch05”having the output in XML tags allows the end user to reliably get the poem and only the poem by writing a short program to extract the content between XML tags.”通用
A023预填 assistant turn（Speaking for Claude）可以引导输出格式T0-1-ch05”put the first XML tag in the assistant turn… you’re basically telling Claude that Claude has already said something, and that it should continue from that point onward.”Claude专属
A024预填 { 可以近乎强制 JSON 输出格式T0-1-ch05”If you want to enforce JSON output (not deterministically, but close to it), you can also prefill Claude’s response with the opening bracket, {”Claude专属
A025将关闭标签传给 stop_sequences 参数可以节省成本和时间T0-1-ch05”pass the closing XML tag to the stop_sequences parameter to get Claude to stop sampling once it emits your desired tag. This can save money and time-to-last-token.”Claude专属
A026给模型时间思考（step by step）能提高复杂任务的准确性T0-1-ch06”Giving Claude time to think step by step sometimes makes Claude more accurate, particularly for complex tasks.”通用
A027思考只有在”大声”输出时才算数——不能要求只输出答案不输出思考过程T0-1-ch06”thinking only counts when it’s out loud. You cannot ask Claude to think but output only the answer - in this case, no thinking has actually occurred.”通用
A028让模型先列出正反论点再作判断，可以改善复杂判断T0-1-ch06”First, write the best arguments for each side in and XML tags, then answer.”通用
A029模型对选项顺序敏感——倾向于选择第二个选项T0-1-ch06”Claude is sometimes sensitive to ordering… Claude is more likely to choose the second of two options, possibly because in its training data from the web, second options were more likely to be correct.”Claude专属
A030明确拼出思考步骤，比简单说”think step by step”更有效T0-1-ch06”literally spelling out the steps that Claude should take in order to process and think through its task”通用
A031用 XML 标签（如 <brainstorm> <scratchpad>）隔离思考区域T0-1-ch06”think step by step, this time in tags”通用
A032Few-shot 示例极其有效——既能得到正确答案，又能得到正确格式T0-1-ch07”Giving Claude examples of how you want it to behave (or how you want it not to behave) is extremely effective for: Getting the right answer; Getting the answer in the right format.”通用
A033用示例比用冗长描述更高效地传达想要的语气T0-1-ch07”You could take the time to describe your desired tone, but it’s much easier just to give Claude a few examples of ideal responses.”通用
A034模型可以从示例中自动归纳格式规则（而不需要显式描述格式）T0-1-ch07”provide Claude with some correctly-formatted examples and Claude can extrapolate from there”通用
A035在 few-shot 中使用预填可以引导模型跟随示例格式T0-1-ch07(在提取名字的例子中使用 <individuals> 作为 prefill)Claude专属
A036模型有时会”幻觉”——编造不存在的信息，因为它试图尽可能帮忙T0-1-ch08”Claude sometimes ‘hallucinates’ and makes claims that are untrue or unjustified.”通用
A037给模型一个”退出机制”——告诉它可以说不知道T0-1-ch08”give Claude an out — tell Claude that it’s OK for it to decline to answer, or to only answer if it actually knows the answer with certainty.”通用
A038让模型先提取相关引用再回答，可以减少长文档场景下的幻觉T0-1-ch08”tell Claude to first extract relevant quotes, then base its answer on those quotes”通用
A039使用 scratchpad 先提取引用再判断——判断引用是否真的回答了问题T0-1-ch08”in tags, pull the most relevant quote from the document and consider whether it answers the user’s question or whether it lacks sufficient detail.”通用
A040降低 temperature 可以减少幻觉——temperature 0 最接近确定性输出T0-1-ch08”Claude’s hallucinations can be solved by lowering the temperature… temperature 0 will generally yield an almost-deterministic answer set.”通用
A041复杂 prompt 的推荐结构有 10 个元素，顺序对部分元素有影响T0-1-ch09”The following structure combines multiple prompt engineering elements and is a good starting point for complex prompts. The ordering matters for some elements, not for others.”通用
A042先用多种技巧让 prompt 工作，然后再精简优化T0-1-ch09”it is usually best to use many prompt elements to get your prompt working first, then refine and slim down your prompt afterward.”通用
A043元素1：Task context（任务上下文）应放在 prompt 最前面T0-1-ch09”Give Claude context about the role it should take on or what goals and overarching tasks you want it to undertake. It’s best to put context early in the body of the prompt.”通用
A044元素2：Tone context（语气上下文）T0-1-ch09”If important to the interaction, tell Claude what tone it should use.”通用
A045元素3：详细任务描述和规则——包含给模型的”退出机制”T0-1-ch09”Expand on the specific tasks you want Claude to do, as well as any rules… This is also where you can give Claude an ‘out’ if it doesn’t have an answer.”通用
A046元素4：示例是知识工作中让模型表现如预期的最有效工具T0-1-ch09”Examples are probably the single most effective tool in knowledge work for getting Claude to behave as desired.”通用
A047给出边缘案例的示例和 scratchpad 示例T0-1-ch09”Make sure to give Claude examples of common edge cases. If your prompt uses a scratchpad, it’s effective to give examples of how the scratchpad should look.”通用
A048示例越多越好（more examples = better）T0-1-ch09”Generally more examples = better.”通用
A049元素5：输入数据用 XML 标签包裹T0-1-ch09”If there is data that Claude needs to process within the prompt, include it here within relevant XML tags.”通用
A050元素6：在长 prompt 末尾重申即时任务（reiterate immediate task）T0-1-ch09”It generally doesn’t hurt to reiterate to Claude its immediate task. It’s best to do this toward the end of a long prompt. This will yield better results than putting this at the beginning.”通用
A051用户的查询/问题放在 prompt 末尾附近效果更好T0-1-ch09”It is also generally good practice to put the user’s query close to the bottom of the prompt.”通用
A052元素7：思考（precognition）放在长 prompt 末尾，紧接即时任务之后T0-1-ch09”For tasks with multiple steps, it’s good to tell Claude to think step by step… it’s best to do this toward the end of a long prompt.”通用
A053元素8：输出格式说明放在 prompt 末尾T0-1-ch09”If there is a specific way you want Claude’s response formatted… putting it toward the end of the prompt is better than at the beginning.”通用
A054元素9：预填可以引导模型进入特定角色或格式T0-1-ch09”A space to start off Claude’s answer with some prefilled words to steer Claude’s behavior or response.”Claude专属
A055法律场景下 prompt 结构可以灵活调整元素顺序T0-1-ch09”We’ve changed around the ordering of a few elements to showcase that prompt structure can be flexible!”通用
A056Prompt 工程是科学的试错过程T0-1-ch09”Prompt engineering is about scientific trial and error.”通用
A057用”Before you give your answer…”来确保模型先思考T0-1-ch09”Sometimes, you might have to even say ‘Before you give your answer…’ just to make sure Claude does this first.”通用
A058法律 prompt 中预填 <relevant_quotes> 引导模型先提取证据T0-1-ch09(Legal example: PREFILL = “<relevant_quotes>“)场景特定
A059Prompt chaining：让模型在第二轮修正自己的错误T0-1-ch10a”Claude can often improve the accuracy of its response when asked to do so… The ways that feel natural to ask a human to double check their work will also generally work for Claude.”通用
A060当模型答案已正确时给出退出机制，防止不必要的修改T0-1-ch10a”If all the words are real words, return the original list.” — 给模型一个保持原答案的选项通用
A061Prompt chaining 可以用来让模型改进自己的创作T0-1-ch10a”ask Claude to make its responses better… first write a story, and then improve the story it wrote.”通用
A062前一轮输出可以作为变量插入下一轮 prompt（substitution chaining）T0-1-ch10a”taking the results of one call to Claude and plugging it into another, longer call”通用
A063Tool use 本质是 substitution + prompt chainingT0-1-ch10b”tool use… is really just a combination of substitution and prompt chaining.”通用
A064Tool use 三步流程：模型输出工具名和参数→暂停→重新输入结果T0-1-ch10b”1. Output the tool name and arguments it wants to call; 2. Halt any further response generation while the tool is called; 3. Then we reprompt with the appended tool results”通用
A065工具定义用 XML 结构描述（名称、描述、参数列表）T0-1-ch10b(tool_description XML with tool_name, description, parameters)Claude专属
A066用 stop_sequences 检测模型何时发起工具调用T0-1-ch10b”use </function_calls> in stop_sequences to detect if and when Claude calls the function.”Claude专属
A067在开始 prompt engineering 之前，先定义成功标准和评估方法T0-2-pe-overview”This guide assumes that you have: 1. A clear definition of the success criteria; 2. Some ways to empirically test against those criteria; 3. A first draft prompt”通用
A068不是所有问题都适合用 prompt engineering 解决——有时换模型更有效T0-2-pe-overview”Not every success criteria or failing eval is best solved by prompt engineering. For example, latency and cost can be sometimes more easily improved by selecting a different model.”通用
A069明确要求”上佳”行为而非依赖模型推断T0-2-best-practices”If you want ‘above and beyond’ behavior, explicitly request it rather than relying on the model to infer this from vague prompts.”通用
A070提供指令时用编号列表或要点列表，当顺序和完整性重要时T0-2-best-practices”Provide instructions as sequential steps using numbered lists or bullet points when the order or completeness of steps matters.”通用
A071提供指令背后的原因/动机能让模型更好地泛化T0-2-best-practices”Providing context or motivation behind your instructions, such as explaining to Claude why such behavior is important, can help Claude better understand your goals.”通用
A072解释原因比单纯禁止更有效——例如解释为什么不能用省略号T0-2-best-practices”Instead of: ‘NEVER use ellipses’; Try: ‘Your response will be read aloud by a text-to-speech engine, so never use ellipses since the text-to-speech engine will not know how to pronounce them.’”通用
A073示例要满足三个要求：相关、多样、结构化T0-2-best-practices”Make them: Relevant: Mirror your actual use case closely. Diverse: Cover edge cases and vary enough… Structured: Wrap examples in tags”通用
A074建议使用 3-5 个示例以获得最佳效果T0-2-best-practices”Include 3–5 examples for best results.”通用
A075可以让模型评估你的示例的相关性和多样性T0-2-best-practices”You can also ask Claude to evaluate your examples for relevance and diversity, or to generate additional ones based on your initial set.”通用
A076XML 标签使用一致的、描述性的命名T0-2-best-practices”Use consistent, descriptive tag names across your prompts.”通用
A077XML 标签可以嵌套以反映内容层级T0-2-best-practices”Nest tags when content has a natural hierarchy (documents inside , each inside ).”通用
A078长文档放在 prompt 顶部，查询/指令放在底部T0-2-best-practices”Put longform data at the top: Place your long documents and inputs near the top of your prompt, above your query, instructions, and examples.”通用
A079查询放在末尾可以提高响应质量达 30%T0-2-best-practices”Queries at the end can improve response quality by up to 30% in tests, especially with complex, multi-document inputs.”通用
A080多文档场景用 <document index="n"> 结构化包裹T0-2-best-practices”…<document_content>…</document_content>“通用
A081长文档任务中让模型先引用再执行任务（Ground responses in quotes）T0-2-best-practices”For long document tasks, ask Claude to quote relevant parts of the documents first before carrying out its task.”通用
A082告诉模型”不要做什么”不如告诉它”做什么替代”T0-2-best-practices”Tell Claude what to do instead of what not to do” — Instead of ‘Do not use markdown’, Try: ‘Your response should be composed of smoothly flowing prose paragraphs.’”通用
A083Prompt 风格会影响输出风格——去掉 prompt 中的 markdown 可以减少输出中的 markdownT0-2-best-practices”The formatting style used in your prompt may influence Claude’s response style… removing markdown from your prompt can reduce the volume of markdown in the output.”通用
A084Claude 4.6 起不再支持最后一个 assistant turn 的预填T0-2-best-practices”Starting with Claude 4.6 models, prefilled responses on the last assistant turn are no longer supported.”Claude专属
A085预填的替代方案：用系统 prompt 指令消除前言、用 Structured Outputs 强制格式T0-2-best-practices”Use direct instructions in the system prompt: ‘Respond directly without preamble.’… The Structured Outputs feature is designed specifically to constrain Claude’s responses.”Claude专属
A086如果想让模型行动而非建议，用明确的行动指令T0-2-best-practices”For Claude to take action, be more explicit… ‘Change this function’ vs ‘Can you suggest some changes‘“通用
A087对于想要主动行动的场景，添加 <default_to_action> 指令块T0-2-best-practices”<default_to_action>By default, implement changes rather than only suggesting them.</default_to_action>“Claude专属
A088对于想要保守行为的场景，添加 <do_not_act_before_instructions> 指令块T0-2-best-practices”<do_not_act_before_instructions>Do not jump into implementation… default to providing information…</do_not_act_before_instructions>“Claude专属
A089新模型对系统提示更敏感——之前激进的语言现在需要调低T0-2-best-practices”If your prompts were designed to reduce undertriggering on tools or skills, these models may now overtrigger. The fix is to dial back any aggressive language.”Claude专属
A090从 “CRITICAL: You MUST use this tool” 降级为正常表述 “Use this tool when…”T0-2-best-practices”Where you might have said ‘CRITICAL: You MUST use this tool when…’, you can use more normal prompting like ‘Use this tool when…’”Claude专属
A091鼓励并行工具调用以提高效率T0-2-best-practices”Claude’s latest models excel at parallel tool execution… you can boost this to ~100%“Claude专属
A092”通用指令优于规定性步骤”——‘think thoroughly’ 往往比手写步骤更好T0-2-best-practices”A prompt like ‘think thoroughly’ often produces better reasoning than a hand-written step-by-step plan. Claude’s reasoning frequently exceeds what a human would prescribe.”通用
A093在 few-shot 示例中使用 <thinking> 标签来展示推理模式T0-2-best-practices”Use tags inside your few-shot examples to show Claude the reasoning pattern. It will generalize that style to its own extended thinking blocks.”Claude专属
A094让模型自检：“Before you finish, verify your answer against [test criteria]“T0-2-best-practices”Ask Claude to self-check. Append something like ‘Before you finish, verify your answer against [test criteria].’ This catches errors reliably, especially for coding and math.”通用
A095避免过度思考：让模型选择方案后坚持执行，不反复推敲T0-2-best-practices”When you’re deciding how to approach a problem, choose an approach and commit to it. Avoid revisiting decisions unless you encounter new information that directly contradicts your reasoning.”通用
A096用 adaptive thinking 代替固定的 extended thinking budgetT0-2-best-practices”In internal evaluations, adaptive thinking reliably drives better performance than extended thinking.”Claude专属
A097工具结果反馈后让模型反思质量再决定下一步T0-2-best-practices”After receiving tool results, carefully reflect on their quality and determine optimal next steps before proceeding.”通用
A098多上下文窗口场景：第一个窗口搭建框架，后续窗口迭代T0-2-best-practices”Use the first context window to set up a framework (write tests, create setup scripts), then use future context windows to iterate on a todo-list.”场景特定
A099让模型先写测试再开始工作——并用结构化格式跟踪测试T0-2-best-practices”Ask Claude to create tests before starting work and keep track of them in a structured format (e.g., tests.json).”场景特定
A100”不要因为 token 预算快用完就提前停止任务”T0-2-best-practices”do not stop tasks early due to token budget concerns… Always be as persistent and autonomous as possible and complete tasks fully”Claude专属
A101用 git 进行状态跟踪——跨多会话保持进度T0-2-best-practices”Git provides a log of what’s been done and checkpoints that can be restored. Claude’s latest models perform especially well in using git to track state.”场景特定
A102Prompt improver 生成的模板含有详细的 CoT 指令、XML 组织和标准化示例格式T0-2-tools”Detailed chain-of-thought instructions… Clear organization using XML tags… Standardized example formatting that demonstrates step-by-step reasoning”通用
A103Prompt improver 适合需要高准确性的复杂任务，但会生成更长更慢的回复T0-2-tools”For latency or cost-sensitive applications, consider using simpler prompts. The prompt improver creates templates that produce longer, more thorough, but slower responses.”通用
A104初始简单 prompt 导致不一致的输出——格式、长度、内容差异大T0-3-real-world-02”The summaries are all over the place: Some are long paragraphs, others are bulleted lists… formatting, length, and types of summary data are inconsistent”场景特定
A105改进 prompt 的核心策略组合：系统提示 + XML 标签 + 明确指令 + 格式要求 + 示例 + 输出标签T0-3-real-world-02”Adding a clear system prompt… Structuring input data using XML tags… Provide clear, specific instructions… Specify the desired format… Include an example… Use output XML tags”通用
A106在示例中同时提供输入和对应输出，模型能学到格式和内容要求T0-3-real-world-02(Complete example with patient_record input + summary output inside tags)通用
A107告诉模型需要 JSON 格式并在示例中给出 JSON 输出T0-3-real-world-02”The easiest way to ‘force’ a JSON response is through Claude’s tool use functionality” and explicit JSON schema instruction通用
A108Prompt 工程不是”基本提问”——是精确的、迭代的、可扩展的过程T0-3-real-world-03”basic prompting often involves single-turn interactions with simple queries. Prompt engineering… involves multi-turn conversations, complex instructions, and carefully structured inputs”通用
A109Prompt 工程生命周期：Plan → Draft → Evaluate → Refine → DeployT0-3-real-world-03”Plan… Draft… Evaluate… Refine… Deploy… This cycle repeats as you encounter new patterns or requirements.”通用
A110六个关键决策点：角色、上下文量、是否用示例、输出格式、如何处理不确定性、约束条件T0-3-real-world-03”Decision 1: What role… Decision 2: How much context… Decision 3: Should you use examples… Decision 4: What format… Decision 5: How should Claude handle uncertainty… Decision 6: What constraints”通用
A111上下文量的平衡：太少导致错误假设，太多浪费 token 可能迷惑模型T0-3-real-world-03”Too little: Claude may make incorrect assumptions; Too much: Wastes tokens and may confuse the model; Just right: Provides necessary context”通用
A112自我批评模式：先生成回复，再要求批评准确性/清晰度/完整性T0-3-real-world-03”Generate a response. Then, critique it for accuracy, clarity, and completeness. Provide the improved version.”通用
A113长文档/transcript 放在 prompt 最前面（top），指令放在后面T0-3-real-world-04”Put long documents (our transcripts) at the top.”通用
A114给模型”思考空间”——在 <thinking> 标签中分析再输出最终结果T0-3-real-world-04”Before generating the JSON, please analyze the transcript in tags… Then, provide your JSON output in tags.”通用
A115处理边缘情况：定义明确的”不可总结”条件和对应输出T0-3-real-world-04”Insufficient data criteria: If either of these conditions are met: a) fewer than 5 exchanges, b) customer’s issue is unclear, c) garbled/incomplete → return ONLY {status: INSUFFICIENT_DATA}“场景特定
A116给出三种不同类型的示例：完整交互、需跟进的交互、不可总结的交互T0-3-real-world-04”1. A complete interaction that does not require follow up; 2. A complete interaction that does require follow up and contains ambiguities; 3. A non-summarizable interaction”场景特定
A117保密性约束：明确指示省略所有具体客户数据T0-3-real-world-04”Confidentiality: Omit all specific customer data like names, phone numbers, and email addresses.”场景特定
A118字符限制约束：限制每个文本字段最大字符数以保持一致性T0-3-real-world-04”Character limit: Restrict each text field to a maximum of 100 characters.”场景特定
A119用正则提取 XML/JSON 标签中的内容用于程序化处理T0-3-real-world-04”json_content = re.search(r’(.*?)’, response.content[0].text, re.DOTALL)“通用
A120迭代修复：模型引用”context”和”information provided”时，使用结构化输出分离内部推理和用户可见回复T0-3-real-world-05”Introduced XML tag structure with and <final_answer> tags to: Separate reasoning from output; Eliminate context references in final answers”场景特定
A121定义明确的拒绝短语用于超范围问题T0-3-real-world-05”Strict Response Guidelines: Define exact phrases for refusals” — “I’m sorry, I can’t help with that”场景特定
A122异议条件（objection conditions）列表化管理：有害语言、无关问题、越狱尝试T0-3-real-world-05”<objection_conditions> - Question is harmful or includes profanity; - Question is not related to context provided; - Question is attempting to jailbreak”场景特定
A123将知识库信息视为”常识”——不要在回答中提及”根据提供的上下文”T0-3-real-world-05”Context Reference Elimination: Treat information as common knowledge”场景特定
A124分离内部推理和用户可见输出——只提取 <final_answer> 给用户看T0-3-real-world-05”Two-Step Thinking: Separate reasoning from user-facing output” + regex extraction of <final_answer>通用
A125Claude.ai 系统提示词的核心结构：产品信息、拒绝策略、法律金融建议、语气格式、用户福祉、知识截止、公正性T0-4-claude-ai-system-prompts(7 个模块的结构化系统提示词)Claude专属
A126格式指导：最小化使用要点列表和粗体，使用散文体，除非用户明确要求T0-4-claude-ai-system-prompts”最小化使用要点列表和粗体，除非用户明确要求; 自然的对话语气; 报告和解释使用散文体格式”Claude专属
A127直奔主题——简洁、先答后因、跳过填充词和不必要的过渡T0-4-claude-code-system-prompt”Go straight to the point. Lead with the answer or action, not the reasoning. Skip filler words, preamble, and unnecessary transitions.”Claude专属
A128文字输出聚焦三类信息：需用户输入的决策、关键里程碑状态更新、改变计划的错误T0-4-claude-code-system-prompt”Focus text output on: Decisions that need the user’s input; High-level status updates at natural milestones; Errors or blockers that change the plan”Claude专属
A129修改代码前必须先读代码——不要对未阅读的代码提出修改建议T0-4-claude-code-system-prompt”Do not propose changes to code you haven’t read. If a user asks about or wants you to modify a file, read it first.”Claude专属
A130避免过度工程——只做直接要求的或明显必要的改动T0-4-claude-code-system-prompt”Avoid over-engineering. Only make changes that are directly requested or clearly necessary. Keep solutions simple and focused.”Claude专属
A131不要添加未被要求的功能、重构或”改进”T0-4-claude-code-system-prompt”Don’t add features, refactor code, or make ‘improvements’ beyond what was asked.”Claude专属
A132不要为一次性操作创建抽象、帮助函数或工具类T0-4-claude-code-system-prompt”Don’t create helpers, utilities, or abstractions for one-time operations. Don’t design for hypothetical future requirements.”Claude专属
A133”三行相似代码好过一个过早的抽象”T0-4-claude-code-system-prompt”three similar lines of code is better than a premature abstraction.”Claude专属
A134只在系统边界做验证（用户输入、外部 API）T0-4-claude-code-system-prompt”Validate exclusively at system boundaries (user input, external APIs)“Claude专属
A135被阻塞时不要暴力重试——考虑替代方案或询问用户T0-4-claude-code-system-prompt”If your approach is blocked, do not attempt to brute force your way to the outcome… consider alternative approaches”Claude专属
A136破坏性操作执行前确认——“暂停确认的成本低，误操作的成本高”T0-4-claude-code-system-prompt”The cost of pausing to confirm is low, while the cost of an unwanted action can be very high.”Claude专属
A137用户批准某操作一次不代表在所有上下文中都批准T0-4-claude-code-system-prompt”A user approving an action once does NOT mean they approve it in all contexts.”Claude专属
A138子代理架构：专门化代理处理聚焦任务，用干净的上下文窗口，返回压缩摘要T0-4-claude-code-system-prompt(Explore Agent: read-only; Plan Mode: 5-phase; Worker Fork: parallel execution)Claude专属
A139工具描述本身就是 prompt engineering——描述必须精确指导工具使用行为T0-4-claude-code-tools(57 个工具描述文件，每个都是精心设计的指令)Claude专属
A140上下文工程 vs Prompt 工程：管理整个上下文状态，而非只关注写好指令T0-5-context-engineering”Context engineering encompasses strategies for curating and maintaining the optimal token set during inference… This differs from prompt engineering, which focuses specifically on writing effective instructions.”通用
A141上下文腐烂（context rot）：随着上下文长度增加，模型性能下降T0-5-context-engineering”LLMs exhibit performance degradation as context length increases—a phenomenon called ‘context rot.’”通用
A142上下文是有限资源——需要像管理稀缺资源一样精心策划T0-5-context-engineering”context must be treated as a finite resource requiring careful curation.”通用
A143系统提示既要足够具体以指导行为，又要足够灵活以允许模型发展强启发T0-5-context-engineering”Should be extremely clear and specific enough to guide behavior effectively, yet flexible enough to allow models to develop strong heuristics.”通用
A144工具返回值应是 token 经济的（token-economical returns）T0-5-context-engineering”Must promote efficiency through token-economical returns and encourage efficient agent behaviors.”场景特定
A145Just-in-time 上下文策略：保持轻量级标识符，动态加载相关信息T0-5-context-engineering”Agents maintain lightweight identifiers (file paths, URLs, queries) and dynamically load relevant information using tools.”场景特定
A146三种管理长任务的技术：压缩（compaction）、结构化笔记、子代理架构T0-5-context-engineering”Compaction: Summarizing conversation contents… Structured Note-Taking: Agents write persistent notes outside the context window… Sub-agent Architectures: Specialized agents handle focused tasks”场景特定
A147核心原则：“找到最大化预期结果概率的最小高信号 token 集合”T0-5-context-engineering”find ‘the smallest set of high-signal tokens that maximize the likelihood of your desired outcome.’”通用