The Context Engineering Playbook: 15 Copy-Paste Templates That Make Any AI 10x Smarter

Nav Toor 2026-04-08 37.8 KB 31.1K 350 𝕏 文章列表

文章提出了“上下文工程”的概念，认为它比单纯的提示词工程更能决定 AI 的输出质量。作者将 LLM 比作 CPU，将上下文窗口比作 RAM，而用户则是操作系统。文章提供了 15 个可直接复制粘贴的实用模板，分为身份设定、任务指令和上下文控制三个部分，旨在帮助用户通过预设背景、品牌语调和工作规则，无需代码即可显著提升 ChatGPT、Claude 等模型的输出质量。

AI 提示词工程 ChatGPT Claude 效率 LLM 上下文窗口工作流模板 Nav Toor

# The Context Engineering Playbook: 15 Copy-Paste Templates That Make Any AI 10x Smarter **作者**: Nav Toor **日期**: 2026-04-07T17:44:26.000Z **来源**: [https://x.com/heynavtoor/status/2041572825698615583](https://x.com/heynavtoor/status/2041572825698615583) --- ![Image](/doc/42/image/img_f2bb61125e84.jpg) (Last time I explained why context engineering killed prompt engineering. This time I am giving you the actual system. Copy these. Paste these. Watch the output change immediately.) My last article on context engineering went viral. Millions of people read it. Thousands sent me the same message: “I understand the concept now. But what do I actually type?” Fair enough. Context engineering is the discipline of designing everything an AI model sees before it generates a response. Not just your prompt. The system instructions. The memory. The tools. The retrieved knowledge. The dynamic assembly. Everything. Andrej Karpathy said it best: the LLM is a CPU, the context window is RAM, and you are the operating system. You decide what gets loaded into memory for each task. That article explained the theory. This one gives you the system. Fifteen templates you can copy, paste into ChatGPT, Claude, or Gemini, and start getting dramatically better results in the next five minutes. No code. No API. No technical skills required. Just copy, paste, and customize the brackets. ## PART 1: IDENTITY TEMPLATES (Who the AI thinks it is) These templates replace the first ten minutes of every AI session. Set them once. They apply to every conversation after. Template 1: The Personal Context File This is the single highest-impact thing you can do. It eliminates the cold start problem permanently. Copy this into Claude’s Custom Instructions, ChatGPT’s Memory, or Gemini’s Gems: I am [your name], [your role] at [company/context]. My audience: [who you create for — clients, readers, team, executives]. Current priorities: [list your top 2-3 active projects or goals]. Communication style: [direct/formal/casual/technical]. I value [brevity/thoroughness/data-driven reasoning]. When I say “draft,” I want a rough version I will edit. When I say “write,” I want something close to final. Default format: [bullet points / short paragraphs / numbered lists]. If you are unsure about something, ask me before guessing. Never fabricate data. Why it works: Every session now starts with the AI already knowing who you are, what you care about, and how you communicate. Anthropic’s own engineering team calls system instructions “the most important lever for controlling Claude’s behavior.” This is that lever. Template 2: The Brand Voice File This prevents the AI from sounding like a generic chatbot and makes it sound like you. My writing voice rules: Tone: [e.g., direct and confident, no hedging or filler phrases] Sentence length: [e.g., short. Two to four sentences per paragraph maximum] Words I use: [list 5-10 words or phrases that define your style] Words I never use: [list words you hate — e.g., “leverage,” “utilize,” “delve,” “I’d be happy to”] Formatting: [e.g., bold for emphasis, no italic, no emojis unless specified] Structure: [e.g., lead with the conclusion, then support with evidence] Here are two examples of my actual writing for reference: [Paste a real paragraph you wrote] [Paste another real paragraph you wrote] Match this voice in everything you write for me. Why it works: Two examples of your real writing do more than a thousand words of instruction. The AI pattern-matches your actual voice instead of guessing. Template 3: The Working Rules File This stops the AI from doing things you hate. Rules for every task: Ask clarifying questions before starting complex tasks. Do not assume. When presenting options, give me your recommendation first, then alternatives. Never use placeholder text. Every example must be specific and real. If a task requires current data, tell me you cannot verify it rather than fabricating numbers. Default output length: [e.g., 300-500 words unless I specify otherwise]. When I say “shorter,” cut by 40%. When I say “longer,” expand by 50%. [Add your own rule: something the AI keeps doing wrong that you want stopped permanently]. Why it works: Every complaint you have about AI output is a rule you have not written yet. This file turns complaints into constraints. Write the rule once and the mistake never happens again. ## PART 2: TASK TEMPLATES (How you ask) These templates replace vague prompts with structured requests that produce first-draft quality output. Template 4: The Research Brief Use this before any research task to prevent surface-level summaries. Research [topic]. Scope: [what specifically I need to know — not “everything about X”] Depth: [surface overview / detailed analysis / comprehensive deep-dive] Sources: [prioritize primary sources / academic papers / industry reports / recent news] Time frame: [only information from the last 6 months / last 2 years / all time] Output format: [bullet summary / structured report with sections / comparison table] For each claim, include the source name and date. Flag anything you are uncertain about. Do not include general background I already know. I know [state what you already know so the AI skips it]. Why it works: The line “Do not include general background I already know” eliminates the filler paragraphs that waste the first third of every AI research response. You are engineering the context of what should not appear. Template 5: The Writing Brief Use this for any content creation task. Write [type of content: email / article / report / social post]. Audience: [who will read this and what do they care about] Goal: [what should the reader think, feel, or do after reading] Key points to include: [list the 3-5 things that must appear] Key points to exclude: [what should NOT be mentioned] Tone: [reference my brand voice file / or specify: formal, casual, provocative, empathetic] Length: [specific word count or range] Structure: [how it should be organized — sections, numbered list, narrative] Reference: [paste or describe an example of what “good” looks like for this task] Why it works: The “Key points to exclude” field is the most underrated line. Most people tell AI what to write. Context engineers also tell it what NOT to write. Constraints produce focus. Focus produces quality. Template 6: The Decision Framework Use this when you need the AI to help you think, not just answer. I need to decide between [Option A] and [Option B] (and [Option C] if applicable). Context: [describe the situation and why this decision matters] Constraints: [budget, timeline, team size, technical limitations] My priorities: [what matters most — speed, cost, quality, risk reduction] What I am leaning toward: [your current instinct and why] Give me: 1. A comparison table with the most important criteria 2. The strongest argument FOR each option 3. The strongest argument AGAINST each option 4. Your recommendation with specific reasoning 5. What I might be overlooking Do not be neutral. Take a position. Why it works: “Do not be neutral. Take a position.” is the single most powerful instruction for decision support. Without it, AI defaults to “it depends” hedging. With it, you get a clear recommendation you can argue against, which is far more useful than a balanced summary that helps no one decide anything. ## PART 3: CONTEXT CONTROL TEMPLATES (Managing what the AI knows) These templates solve the biggest problem in AI: the model forgets, hallucinates, or loses track of your goals mid-conversation. Template 7: The Session Kickoff Use this at the start of any complex, multi-step session. Before we begin, here is the context for this session: Project: [name and one-line description] Current status: [where things stand right now] What we accomplished last session: [summary of prior work if applicable] Goal for this session: [specific deliverable or outcome] Files to reference: [list any documents, links, or prior outputs to use] Constraints: [time, word count, format, scope limitations] Confirm you understand the full context before starting. Why it works: This is context loading — the practice of frontloading everything the model needs before it generates a single token. LangChain’s context engineering framework calls this “selecting context.” You are pulling exactly the right information into the window at exactly the right time. Template 8: The Anti-Hallucination Guard Use this for any task involving facts, data, or claims. Important rules for this task: Only include information you can directly verify from the sources I provide or from your training data. If you are uncertain about a specific fact, number, or date, flag it with [UNVERIFIED] next to it. Do not invent statistics. Do not fabricate quotes. Do not create fake citations. If I ask about something beyond your knowledge cutoff, say so explicitly instead of guessing. Distinguish clearly between “this is a fact” and “this is my analysis.” Why it works: The [UNVERIFIED] tag system is the simplest anti-hallucination technique that exists. Instead of hoping the AI does not make things up, you give it an explicit escape valve. The model is far less likely to fabricate when it has permission to flag uncertainty. Anthropic’s own documentation recommends this approach for high-stakes factual tasks. Template 9: The Context Reset Use this when a conversation has gone off track. Stop. Let us reset this conversation. Here is what we are actually trying to accomplish: [restate the original goal] Here is what has gone wrong: [describe what went off track] Here is what I want you to do differently now: [specific instruction] Ignore everything above this message that contradicts these new instructions. Start fresh from this point. Why it works: Long conversations suffer from what researchers call “context rot” — the model’s attention drifts to recent messages and loses the original instructions. Claude Code triggers an automatic “compaction” when you hit 95% of the context window, summarizing everything to stay focused. This template lets you do the same thing manually in any AI tool. A clean reset is better than ten correction messages. ## PART 4: ADVANCED TEMPLATES (Building systems, not conversations) These templates move you from using AI as a chat tool to building AI-powered workflows. Template 10: The CLAUDE.md / Project Rules File For Claude Code, Cursor, or any AI coding tool. This is the context file that governs every interaction. Project: [Name] ## Architecture - Language: [e.g., TypeScript] - Framework: [e.g., Next.js 15] - Database: [e.g., PostgreSQL with Prisma ORM] - Testing: [e.g., Vitest for unit tests, Playwright for E2E] ## Code Standards - Use functional components. No classes. - All functions must have explicit return types. - Error handling: use Result types, not try-catch. - File naming: kebab-case. One component per file. ## Workflow Rules - Write tests before writing implementation code. - Run all tests before marking any task as complete. - Never modify files outside the scope of the current task. - When uncertain about architecture decisions, ask before implementing. ## Common Mistakes to Avoid - Do not add console.log statements in production code. - Do not install new dependencies without explicit approval. - Do not refactor existing code unless the task specifically requires it. Why it works: This file loads at the start of every Claude Code session. Every instruction applies automatically to every task. Context engineering for developers means never repeating the same correction twice — you encode the correction into the rules file and it becomes permanent. Template 11: The Multi-Step Task Planner Use this to prevent the AI from rushing through complex tasks. I need you to complete a complex task. Do NOT start executing immediately. First, create a numbered plan with every step required to complete this task: - What information you need - What order to work in - What the deliverable for each step looks like - What could go wrong at each step Present the plan for my approval before executing any step. After I approve, work through each step one at a time, confirming completion before moving to the next. Task: [describe what you need] Why it works: This is the Plan-Execute-Verify loop applied to any AI tool. The AI’s biggest failure mode is rushing to output without thinking. Forcing it to plan first, then wait for approval, dramatically improves the quality of complex multi-step work. This mirrors how agentic engineering workflows are designed — plan first, execute second, verify third. Template 12: The Self-Evaluation Loop Use this to make the AI check its own work before showing you. After completing this task, evaluate your own output before presenting it to me. Check against these criteria: 1. Does it fully address every requirement I listed? 2. Is anything factually questionable or unverified? 3. Is the length within the specified range? 4. Does the tone match my brand voice? 5. Are there any filler sentences that add no value? Remove them. 6. Would I be comfortable putting my name on this? If not, what needs to change? If the output fails any check, revise it before showing me. Show me only your best version. Why it works: This single template can cut your editing time in half. Instead of being your first draft generator and reviewer, the AI becomes its own first reviewer and shows you its second draft. Anthropic calls this pattern “chain-of-thought self-verification.” You are engineering the context to include an evaluation step before delivery. ## PART 5: WORKFLOW AUTOMATION TEMPLATES These templates turn one-time conversations into reusable systems. Template 13: The Weekly Briefing System Set this as a recurring scheduled task or run it every Monday. Generate my weekly briefing using this structure: PRIORITY ACTIONS: What are the 3 most important things I need to act on this week? Base this on [my calendar / email / project status]. MEETINGS PREP: List my meetings for the week. For each, include: who is attending, what the meeting is about, and one thing I should prepare. FOLLOW-UPS: List anything from last week that I said I would follow up on but has not been completed yet. READING LIST: Summarize the 3 most relevant articles or updates in [my industry / topic area] from the past 7 days. Keep each section concise. No longer than 5 bullet points per section. Lead with the most urgent item. Why it works: This is context engineering applied to workflow automation. You are not writing a prompt — you are designing an information system that runs repeatedly and improves as your context files get better. The AI pulls from your connected tools (calendar, email, web) and assembles a briefing that would take you 30 minutes to compile manually. Template 14: The Meeting Notes Processor Use this after every meeting to extract structured value from raw notes. Here are my raw meeting notes: [paste or upload notes] Process these into the following structure: DECISIONS MADE: What was decided? List each decision as a clear statement. ACTION ITEMS: Who needs to do what, by when? Format: [Person] — [Action] — [Deadline] OPEN QUESTIONS: What was raised but not resolved? KEY QUOTES: Any important statements worth preserving verbatim. NEXT STEPS: What happens next and when is the next meeting? If any information is unclear from the notes, flag it with [UNCLEAR] rather than guessing. Why it works: Raw meeting notes are chaos. This template applies LangChain’s “compress context” strategy — it takes a large, unstructured input and compresses it into only the tokens that matter. The [UNCLEAR] tag applies the same anti-hallucination logic from Template 8 — better to flag uncertainty than fabricate details. Template 15: The Skill Builder Use this to turn any repeatable task into a permanent context template. I just completed a task that I will need to do again in the future. Help me turn it into a reusable template. The task: [describe what you just did] What worked well: [what aspects of the output were good] What I had to fix manually: [what the AI got wrong that I corrected] What I want different next time: [improvements for the next iteration] Create a detailed instruction template that I can save and reuse. Include: - The exact context the AI needs to know before starting - Step-by-step instructions for completing the task - Quality standards and constraints - Common mistakes to avoid (based on what I had to fix this time) - The output format and structure Write it so that I can paste this template into any AI tool and get consistent results every time. Why it works: This is meta context engineering — using the AI to build your own context templates. Every task you repeat becomes a template. Every template eliminates one prompt you never have to write again. LangChain calls this the “write” strategy — saving context externally so it can be loaded later. After a few weeks of building templates, your entire workflow is encoded in reusable files. The system gets smarter without the model getting smarter. That is the whole point. The system behind the templates Here is the mental model that ties all fifteen together. Templates 1-3 are your identity layer. They answer: “Who am I talking to?” Set these once and the AI knows you permanently. Templates 4-6 are your task layer. They answer: “What exactly do you need right now?” Use these for every new request. Templates 7-9 are your control layer. They answer: “How do I keep the AI on track?” Use these when things go wrong or get complex. Templates 10-12 are your systems layer. They answer: “How do I build workflows, not conversations?” Use these for professional and development work. Templates 13-15 are your automation layer. They answer: “How do I make this run without me?” Use these to turn one-time work into repeatable systems. Stack the layers. Template 1 plus Template 5 plus Template 12 means the AI knows who you are, understands exactly what you need, and checks its own work before showing you. Three templates. Ten seconds to load. Output that takes two minutes to review instead of twenty. The people who copy one template will get better results tomorrow. The people who build the full five-layer system will get compounding results for the rest of the year. Context engineering is not about writing better prompts. It is about building better systems. The templates are here. The system is explained. The only thing left is to start. ## 相关链接 - [Nav Toor](https://x.com/heynavtoor) - [@heynavtoor](https://x.com/heynavtoor) - [25K](https://x.com/heynavtoor/status/2041572825698615583/analytics) - [Upgrade to Premium](https://x.com/i/premium_sign_up) - [1:44 AM · Apr 8, 2026](https://x.com/heynavtoor/status/2041572825698615583) - [25.5K Views](https://x.com/heynavtoor/status/2041572825698615583/analytics) - [View quotes](https://x.com/heynavtoor/status/2041572825698615583/quotes) --- *导出时间: 2026/4/8 22:58:39* --- ## 中文翻译 # 上下文工程实战手册：15 个可直接复制粘贴的模板，让任何 AI 变聪明 10 倍 **作者**：Nav Toor **日期**：2026-04-07T17:44:26.000Z **来源**：[https://x.com/heynavtoor/status/2041572825698615583](https://x.com/heynavtoor/status/2041572825698615583) --- ![Image](/doc/42/image/img_f2bb61125e84.jpg) （上次我解释了为什么上下文工程取代了提示词工程。这次我要把整套系统交给你。复制这些模板。粘贴这些模板。立刻看到输出质量的变化。）我上一篇关于上下文工程的文章爆了。数百万人阅读，成千上万的人给我发了同样的话："我理解这个概念了。但我到底该输入什么？" 问得有道理。上下文工程是一门学科，研究的是如何在 AI 模型生成回复之前，设计好它所看到的一切。不仅仅是你的提示词。系统指令。记忆。工具。检索到的知识。动态组装的一切。 Andrej Karpathy 说得最到位：LLM 是 CPU，上下文窗口是内存，而你是操作系统。你来决定每个任务要加载什么进入内存。上一篇文章讲了理论。这一篇给你系统。十五个模板，你可以复制、粘贴到 ChatGPT、Claude 或 Gemini 中，在接下来的五分钟内就开始获得显著更好的结果。不需要代码。不需要 API。不需要任何技术能力。只需要复制、粘贴，替换括号里的内容。 ## 第 1 部分：身份模板（AI 认为自己是谁）这些模板可以替代每次 AI 会话的前十分钟。设置一次，之后的所有对话都适用。模板 1：个人上下文文件这是你能做的影响力最大的单件事。它能永久消除冷启动问题。把这段内容复制到 Claude 的自定义指令、ChatGPT 的记忆功能，或 Gemini 的 Gems 中：我是 [你的名字]，[公司/背景] 的 [你的角色]。我的受众：[你为谁创作——客户、读者、团队、高管]。当前优先事项：[列出你最重要的 2-3 个进行中的项目或目标]。沟通风格：[直接/正式/随意/技术性]。我重视 [简洁/详尽/数据驱动的推理]。当我说“起草”时，我想要一个我会再修改的粗略版本。当我说“撰写”时，我想要接近成稿的内容。默认格式：[要点列表/短段落/编号列表]。如果你对某件事不确定，先问我，不要瞎猜。绝不编造数据。为什么有效：现在每次会话开始时，AI 都已经知道你是谁、你在意什么、你如何沟通。Anthropic 自己的工程团队称系统指令是“控制 Claude 行为的最重要杠杆”。这就是那个杠杆。模板 2：品牌声音文件这能防止 AI 听起来像个千篇一律的聊天机器人，让它听起来像你本人。我的写作声音规则：语气：[例如：直接而自信，不用含糊其辞或废话填充] 句子长度：[例如：短句。每段最多两到四句话] 我使用的词：[列出 5-10 个能定义你风格的词或短语] 我绝不使用的词：[列出你讨厌的词——例如：“赋能”、“利用”、“深挖”、“我很乐意”] 格式：[例如：用加粗表示强调，不用斜体，除非特别说明否则不用表情符号] 结构：[例如：先给出结论，再用证据支撑] 以下是我真实写作的两个示例，供参考： [粘贴一段你真实写过的段落] [再粘贴一段你真实写过的段落] 在为我写作的所有内容中，都要贴合这种声音。为什么有效：两个你真实写作的示例，胜过一千字的语言描述。AI 会对你的真实声音进行模式匹配，而不是靠猜。模板 3：工作规则文件这能阻止 AI 做那些你讨厌的事。适用于每个任务的规则：在开始复杂任务之前，先提出澄清性问题。不要自行假设。在呈现选项时，先给我你的推荐，然后再给其他备选方案。绝不使用占位文本。每个例子都必须具体且真实。如果任务需要最新数据，告诉我你无法验证，而不是编造数字。默认输出长度：[例如：300-500 字，除非我另有说明]。当我说“短一点”，就删减 40%。当我说“长一点”，就扩充 50%。 [添加你自己的规则：某些 AI 总是做错、你想永久制止的事情]。为什么有效：你对 AI 输出的每一条抱怨，都是一条你还没写下的规则。这个文件把抱怨转化为约束。规则写一次，错误就永远不会再发生。 ## 第 2 部分：任务模板（你如何提问）这些模板用结构化的请求取代模糊的提示词，产出初稿级别的输出。模板 4：研究简报在任何研究任务之前使用这个模板，防止得到浮于表面的总结。研究 [主题]。范围：[我具体需要知道什么——不是“关于 X 的一切”] 深度：[浅层概览/详细分析/全面深挖] 来源：[优先一手来源/学术论文/行业报告/近期新闻] 时间范围：[仅限最近 6 个月的信息/最近 2 年/不限时间] 输出格式：[要点总结/带章节的结构化报告/对比表格] 每个论断都要注明来源名称和日期。标出任何你不确定的内容。不要包含我已经知道的通用背景。我知道 [说明你已经了解的内容，让 AI 跳过这部分]。为什么有效：“不要包含我已经知道的通用背景”这句话，消除了那些浪费 AI 研究回复前三分之一的填充段落。你是在对“不该出现的内容”进行上下文工程。模板 5：写作简报在任何内容创作任务中使用这个模板。撰写 [内容类型：电子邮件/文章/报告/社交帖子]。受众：[谁会读这篇文章，他们关心什么] 目标：[读者读完后应该怎么想、感受或做什么] 必须包含的要点：[列出必须出现的3-5件事] 必须排除的要点：[不应提及的内容] 语气：[参考我的品牌语音文件 / 或具体指定：正式、随性、犀利、共情] 长度：[具体字数或范围] 结构：[应如何组织——分节、编号列表、叙述式] 参考：[粘贴或描述一个能体现“优秀”标准的范例] 为什么有效：“必须排除的要点”这一栏是被低估最多的一行。大多数人只告诉AI要写什么。上下文工程师还会告诉它不要写什么。约束产生专注，专注产生质量。模板6：决策框架当你需要AI帮你思考而不仅仅是回答时，使用此模板。我需要在[选项A]和[选项B]之间做出决定（如适用，还包括[选项C]）。背景：[描述情况以及这个决定为何重要] 约束条件：[预算、时间线、团队规模、技术限制] 我的优先事项：[什么最重要——速度、成本、质量、降低风险] 我目前的倾向：[你现在的直觉以及原因] 请给我： 1. 一张包含最重要评判标准的对比表 2. 每个选项最有力的支持论据 3. 每个选项最有力的反对论据 4. 你的建议及具体理由 5. 我可能忽略了什么不要保持中立。要表明立场。为什么有效：“不要保持中立。要表明立场。”是决策支持中最有力的一条指令。没有它，AI会默认使用“视情况而定”的模棱两可的说法。有了它，你会得到一个可以反驳的明确建议，这比帮不了任何人做出任何决定的平衡总结有用得多。 ## 第三部分：上下文控制模板（管理AI所知道的信息）这些模板解决AI最大的问题：模型会遗忘、产生幻觉，或在对话中途丢失你的目标。模板7：会话启动在任何复杂的多步骤会话开始时使用此模板。在开始之前，以下是本次会话的背景：项目：[名称和一行描述] 当前状态：[目前进展如何] 上次会话完成的内容：[如有，先前工作的总结] 本次会话的目标：[具体的交付成果或产出] 需参考的文件：[列出任何需要使用的文档、链接或先前输出] 约束条件：[时间、字数、格式、范围限制] 开始前请确认你已理解完整背景。为什么有效：这就是上下文加载——在模型生成任何内容之前，预先提供它所需的一切信息。LangChain的上下文工程框架将此称为“选择上下文”。你在恰到好处的时间将恰到好处的信息拉入窗口。模板8：反幻觉防护在任何涉及事实、数据或声明的任务中使用此模板。本任务的重要规则：只包含你能从我提供的来源或你的训练数据中直接验证的信息。如果你对某个具体的事实、数字或日期不确定，在其旁边标注[未验证]。不要编造统计数据。不要捏造引言。不要创建虚假引用。如果我询问超出你知识截止日期的内容，请明确说明，而不是猜测。清晰区分“这是事实”和“这是我的分析”。为什么有效：[未验证]标签系统是现存最简单的反幻觉技术。与其寄希望于AI不胡编乱造，不如给它一个明确的逃生阀。当模型获得标记不确定性的许可时，它编造内容的可能性会大大降低。Anthropic自己的文档也针对高风险事实任务推荐了这种方法。模板9：上下文重置当对话偏离轨道时使用此模板。停下。让我们重置这次对话。我们实际上要完成的是：[重述最初的目标] 出了什么问题：[描述哪里偏离了轨道] 我现在希望你做何改变：[具体指令] 忽略本消息之上所有与这些新指令相矛盾的内容。从此刻起重新开始。为什么有效：长对话会遭受研究人员所说的“上下文腐烂”——模型的注意力会漂移到最近的消息上，丢失最初的指令。当达到上下文窗口的95%时，Claude Code会触发自动“压缩”，总结所有内容以保持专注。这个模板让你可以在任何AI工具中手动做同样的事情。一次干净的重置胜过十条纠正消息。 ## 第四部分：高级模板（构建系统，而非对话）这些模板让你从把AI当作聊天工具，进阶到构建AI驱动的工作流。模板10：CLAUDE.md / 项目规则文件适用于Claude Code、Cursor或任何AI编程工具。这是管理每一次交互的上下文文件。项目：[名称] ## 架构 - 语言：[例如：TypeScript] - 框架：[例如：Next.js 15] - 数据库：[例如：PostgreSQL与Prisma ORM] - 测试：[例如：Vitest用于单元测试，Playwright用于端到端测试] ## 代码规范 - 使用函数式组件。不使用类。 - 所有函数必须有显式返回类型。 - 错误处理：使用Result类型，不用try-catch。 - 文件命名：kebab-case（短横线命名法）。每个文件一个组件。 ## 工作流规则 - 先写测试，再写实现代码。 - 在将任何任务标记为完成之前，运行所有测试。 - 永远不要修改当前任务范围之外的文件。 - 对架构决策不确定时，先询问再实现。 ## 需避免的常见错误 - 不要在生产代码中添加console.log语句。 - 未经明确批准，不要安装新的依赖项。 - 除非任务特别要求，不要重构现有代码。为什么有效：这个文件在每次 Claude Code 会话开始时加载。每条指令都会自动应用到每个任务。对开发者而言，上下文工程意味着同一种纠正永远不需要重复两次——你把纠正编码进规则文件中，它就变成永久性的。模板 11：多步骤任务规划器用它来防止 AI 匆忙草率地处理复杂任务。我需要你完成一个复杂任务。不要立即开始执行。首先，创建一个编号计划，列出完成此任务所需的每个步骤：- 你需要哪些信息 - 应按什么顺序进行 - 每个步骤的交付成果是什么 - 每个步骤可能出什么问题在执行任何步骤之前，先提交计划给我审批。我批准后，一次执行一个步骤，确认完成后再进入下一步。任务：[描述你的需求] 为什么有效：这是将“计划-执行-验证”循环应用到任何 AI 工具上。AI 最大的失败模式是不经思考就急于输出。强制它先制定计划、再等待批准，能显著提高复杂多步骤工作的质量。这反映了智能体工程工作流的设计方式——先计划，其次执行，第三验证。模板 12：自我评估循环用它让 AI 在向你展示成果之前先检查自己的工作。完成此任务后，在向我展示之前先评估你自己的输出。对照以下标准检查：1. 它是否完全满足我列出的每项要求？2. 是否有事实存疑或未经核实的内容？3. 篇幅是否在指定范围内？4. 语气是否符合我的品牌调性？5. 是否有毫无价值的填充句？将其删除。6. 我是否愿意署上自己的名字发布它？如果不愿意，需要修改什么？如果输出未通过任何一项检查，先修改再展示给我。只展示你最好的版本。为什么有效：仅这一个模板就能将你的编辑时间减少一半。AI 不再只是你的初稿生成器和审阅者，而是成为自己的第一轮审阅者，向你展示的是它的第二稿。Anthropic 将这种模式称为“思维链自我验证”。你是在工程化地设计上下文，在交付之前加入一个评估步骤。 ## 第 5 部分：工作流自动化模板这些模板将一次性对话转化为可复用的系统。模板 13：每周简报系统将其设置为定期计划任务，或每周一运行一次。按照以下结构生成我的每周简报：优先行动：本周我需要采取行动的 3 件最重要的事情是什么？基于[我的日历 / 邮件 / 项目状态]。会议准备：列出我本周的会议。每场会议包括：参会人员、会议内容，以及我应该准备的一件事。跟进事项：列出上周我说过要跟进但尚未完成的事项。阅读清单：总结过去 7 天内[我的行业 / 主题领域]最相关的 3 篇文章或更新。每个部分保持简洁。每个部分不超过 5 个要点。最紧急的事项放在最前面。为什么有效：这是将上下文工程应用于工作流自动化。你写的不只是一个提示词——你是在设计一个可以反复运行、并随着上下文文件质量提升而不断改进的信息系统。AI 从你连接的工具（日历、邮件、网络）中提取信息，汇编出一份你手动整理需要 30 分钟的简报。模板 14：会议纪要处理器每次会议后使用，从原始笔记中提取结构化价值。这是我的原始会议笔记：[粘贴或上传笔记] 将它们整理为以下结构：已做出的决定：决定了什么？将每项决定列成清晰的陈述。行动项：谁需要在什么时间之前做什么？格式：[人员] — [行动] — [截止日期] 未决问题：提出了哪些但尚未解决的事项？关键引述：任何值得原样保留的重要陈述。后续步骤：接下来会发生什么？下次会议是什么时候？如果笔记中有任何信息不清晰，请用 [UNCLEAR] 标记，而不是猜测。为什么有效：原始会议笔记是混乱的。这个模板应用了 LangChain 的“压缩上下文”策略——它将大量非结构化的输入压缩成只包含关键标记的内容。[UNCLEAR] 标签应用了与模板 8 相同的反幻觉逻辑——标记不确定性总比捏造细节好。模板 15：技能构建器用它将任何可重复的任务转化为永久的上下文模板。我刚刚完成了一项将来还需要再次执行的任务。帮我把它转换成一个可复用的模板。任务：[描述你刚刚做了什么] 效果好的地方：[输出的哪些方面是好的] 我必须手动修复的地方：[AI 出了什么错、我做了什么纠正] 下次我希望有所不同的地方：[对下次迭代的改进建议] 创建一个详细的指令模板供我保存和复用。包括：- AI 开始前需要知道的准确上下文 - 完成任务的分步说明 - 质量标准和约束条件 - 要避免的常见错误（基于我这次不得不修复的问题） - 输出格式和结构写得让我能把此模板粘贴到任何 AI 工具中，并每次都获得一致的结果。为什么有效：这是元上下文工程——用 AI 来构建你自己的上下文模板。你重复的每个任务都会变成一个模板。每个模板都消除了一条你再也不用写的提示词。LangChain 将此称为“写入”策略——将上下文保存在外部，以便日后加载。经过几周的模板构建，你的整个工作流都被编码在可复用的文件中。系统变得更聪明，而模型本身并没有变聪明。这正是全部要义所在。模板背后的系统以下是将所有十五个模板串联起来的思维模型。模板 1-3 是你的身份层。它们回答的问题是：“我在和谁对话？”设置一次，AI 就会永久了解你。模板 4-6 是你的任务层。它们回答的问题是：“你现在具体需要什么？”每次新的请求都使用它们。模板 7-9 是你的控制层。它们回答的问题是：“我如何让 AI 保持正轨？”当事情出错或变复杂时使用它们。模板 10-12 是你的系统层。它们回答的问题是：“我如何构建工作流，而不是对话？”在专业和开发工作中使用它们。模板 13-15 是你的自动化层。它们回答的问题是：“我如何让这一切无需我参与就能运行？”用它们把一次性工作变成可重复的系统。将这些层级叠加起来。模板 1 加模板 5 加模板 12，意味着 AI 知道你是谁，准确理解你的需求，并在展示给你之前自查其工作。三个模板。十秒钟加载。审阅输出只需两分钟，而不是二十分钟。只复制一个模板的人，明天就会得到更好的结果。而构建完整五层系统的人，将在这一年余下的时间里获得复利式的成果。上下文工程不是要写出更好的提示词，而是要构建更好的系统。模板已经在这里，系统已经解释清楚。剩下的只有开始行动了。 ## 相关链接 - [Nav Toor](https://x.com/heynavtoor) - [@heynavtoor](https://x.com/heynavtoor) - [25K](https://x.com/heynavtoor/status/2041572825698615583/analytics) - [升级至 Premium](https://x.com/i/premium_sign_up) - [1:44 AM · Apr 8, 2026](https://x.com/heynavtoor/status/2041572825698615583) - [25.5K 次浏览](https://x.com/heynavtoor/status/2041572825698615583/analytics) - [查看引用](https://x.com/heynavtoor/status/2041572825698615583/quotes) --- *导出时间: 2026/4/8 22:58:39*