How to master graph engineering

Machina 2026-07-23 12.9 KB 548.3K 1.8K 𝕏 文章列表

本课程教授如何构建 AI 智能体图，涵盖图的基本概念、关键模式（如菱形模式）、停止规则及人工审批环节。包含三个实战案例：深度研究台、SEO 内容生成器和市场推广套件，旨在提升业务效率并控制成本。

# How to master graph engineering (Full Course) **作者**: Machina **日期**: 2026-07-22T14:20:41.000Z **来源**: [https://x.com/EXM7777/status/2079934660982047021](https://x.com/EXM7777/status/2079934660982047021) --- ![Image](/doc/3297/image/img_c14c7a6693dd.jpg) I'm going to show you how to build your first agent graph and put it to work in your business today: A team of AI agents that researches in parallel, tries to kill its own findings, and hands you one merged result with your approval as the last yes before anything ships this course covers the whole method: what a graph is, the one pattern that pays for itself, the stop rule that protects your bill, and 3 complete builds for your business: a research desk, an SEO content machine, and a go-to-market kit some quick context before we start, because this trend has a story the engineer who built Claude Code said he doesn't prompt it anymore, he writes loops that prompt it, and a few weeks later the same corner of the internet declared loops old news and moved on to graphs under all that noise there is a real craft, it is simpler than the posts make it look, and you don't need to be an engineer for any of it the course map: - lesson 1: what a graph actually is - lesson 2: the diamond pattern - lesson 3: the stop rule - lesson 4: the human gate - graph 1 - the deep research desk (build + prompt) - graph 2 - the SEO content machine (build + prompt) - graph 3 - the go-to-market kit (build + prompt) - the playbook if you want to go deeper than this article and learn how to turn systems like these into money, that is exactly what i built the real time AI ops community for: weeklyaiops.com ## lesson 1: what a graph actually is a graph is a plan for your AI work, drawn out so you can see it it answers two questions: which jobs need to happen, and which job has to wait for which every job is something you would hand to one assistant, like researching a competitor, writing one draft, or checking one claim when a job needs the result of another job before it can start, you connect the two with an arrow and a small set of running notes travels along with the work, holding what was found, what was decided, and what is still left to do i turned this whole vocabulary into the card below, so save it once and you will never need a definition again: ![Image](/doc/3297/image/img_0e88e314864b.jpg) now i would advise you to do one thing tonight, before you touch any new tool look at the AI system you run today and, for every "and then" inside it, ask whether the next job needs the last job's result an arrow is real only when work flows through it "summarize this file and then check my calendar" sounds like a sequence, but the calendar step never needs the summary, so those two jobs never had to wait for each other when the answer is no, the arrow is fake and the waiting is wasted, and you will find two or three of these fake edges in any system you draw right now your system runs as a straight line where every job waits for the one before it, and that works, but it is the slowest possible way to run things, because one stuck job blocks everything behind it one more thing before we build: engineers pushed back on this trend within hours, pointing out that it is a decades-old pattern wearing a new name, and they are right i see that as the good news, because a pattern that has run critical systems for decades is exactly what you want to trust with your business ## lesson 2: the diamond - the one pattern that pays watch any serious agent system work and the same picture keeps appearing the work splits, several workers dig side by side, something checks what they found, and everything merges back into one answer that picture is called the diamond, and i feel confident saying it is the only pattern you need this year the research feature inside Claude runs exactly this in production, one lead plans, workers gather in parallel, and the findings get checked before they reach you now the part everyone skips, the checking step, and you should treat it as non-negotiable every serious test of AI self-review lands on the same conclusion: models miss most of their own mistakes so you should never let the same agent grade its own homework give the checking to a separate job whose only task is to kill weak findings before they reach you, and give every checker a different question... one asks if it is correct, one asks if it is current, one asks if the source is real ![Image](/doc/3297/image/img_fe9ced6e62cb.jpg) ## lesson 3: the stop rule more agents is not automatically better, and this is the part of the course that saves you real money when researchers hand a team of agents and a single agent the same budget, the team wins clearly on work that splits into independent pieces, and loses just as clearly on work where each step needs the previous one so here is the stop rule a graph buys breadth, it does not buy better judgment my advice is simple: when every step needs the full picture, you should stay with one agent, and the moment the work splits into jobs that never read each other's results, that is when the graph starts paying so before you add a single agent, ask the only question that decides the bill... where does my work split ## lesson 4: the human gate you are the most important node in your own graph, and this lesson is about where exactly you should stand every serious system routes to a human before anything irreversible, because the send, the publish, the refund and the invoice are arrows that should end at you, and that approval is the last yes between a drafted pipeline and the world even Klarna learned this in public, going all in on AI support, admitting the cost cutting went too far, and bringing human service back as the premium tier the design rule fits in one line put your approval where a mistake would be expensive to undo, not on every step a gate on everything makes you the bottleneck, and a gate on nothing means nobody is watching i would also advise you to judge the whole system on numbers that cannot argue back, money that landed, customers that stayed, tests that ran, because a system that only grades its own reports is confidently wrong four small rules keep everything from becoming an expensive accident, and the card below is the version you save: every loop gets a maximum number of rounds, only one job writes to any one file, the routing lives in written steps while the AI fills the jobs, and there is always a cap on how many agents can spawn ![Image](/doc/3297/image/img_f111423a4d79.jpg) ## the builds - three graphs that pay for themselves every build here runs in Claude Code, ends with the exact prompt you paste, and keeps you as the gate one word in each prompt does the heavy lifting: "workflow" tells Claude to build a coordinated team instead of working through a single line of steps ## graph 1 - the deep research desk think of the last decision you made with real money behind it, a price change, a new offer, a market you almost entered you probably decided it on a week of googling, or an expensive analyst invoice, and this build replaces both your question splits into five angles, five researchers dig at once, a skeptic attacks every finding, and only the survivors reach the final report with their sources attached ![Image](/doc/3297/image/img_061d8aa397ea.jpg) 1. open Claude Code in an empty folder 2. write your real question in one sentence, the kind with money behind it... "should i raise my price" or "which of these two offers do i launch" 3. paste the prompt below with your question in it 4. watch the fan-out, five researchers working in parallel, then the skeptic pass 5. read the report, and only what survives your read enters the decision the prompt: "i need decision-grade research on: [your question]. use a workflow: split the question into 5 distinct angles, run one researcher per angle in parallel, every finding needs a source link and a date, then run a skeptic against each finding that tries to disprove it, drop what fails, merge the survivors into one report ranked by confidence, save it as research-report.md and show me the top findings" your first run takes minutes instead of days, and i feel like the skeptic pass is the real product here, because it is the difference between research and rumor collection ## graph 2 - the SEO content machine be honest about your publishing schedule for a second, because for a lot of businesses it either costs a monthly retainer or it simply stopped happening this build writes one ranking-ready draft per run, and it never publishes without you three researchers work side by side, one on what the current top pages cover, one on the real questions people ask, one on what everyone misses, then their work merges into an outline, a draft gets written and fact-checked, and it waits in your folder for your yes ![Image](/doc/3297/image/img_2783a3882e91.jpg) 1. pick the one topic your customers type into google 2. paste the prompt below 3. the three research jobs run at once, top pages, real questions, gaps 4. the draft lands in your drafts folder with every claim checked and the weak ones flagged at the top 5. you read it, fix what only you would know, and publish it yourself 6. then ask Claude to save the workflow, and every monday this becomes one command instead of one afternoon the prompt: "i want an article that can rank for: [topic]. use a workflow: run three research jobs in parallel, one lists what the current top-ranking pages cover, one collects the real questions people ask about this topic, one finds what the top pages skip, merge all three into an outline, write a full draft from the outline, then run a fact-checker that flags every claim without a source, save the draft to drafts/ with the flagged claims listed at the top, never publish anything" ## graph 3 - the go-to-market kit you are launching something, a product, an offer, a service, and the honest options were always weeks of prep or launching blind this build produces the full kit in one run, with you approving every piece three researchers profile the buyer, the channels and the competition in parallel, their work merges into a one-page positioning doc, and the system pauses there so you can read it, because that page becomes the foundation every asset gets checked against then three writers draft the landing copy, the launch posts and the outreach messages side by side, a checker flags anything that drifts from the positioning, and the kit lands in a folder where nothing goes live on its own ![Image](/doc/3297/image/img_0c9259e66ea1.jpg) 1. name the product and the person it is for, one line each 2. paste the prompt below 3. read the positioning page when the system pauses, this is the one step i would never skip 4. let the three writers produce the assets in parallel 5. review what the checker flagged 6. approve piece by piece, nothing ships without you the prompt: "i'm launching [product] for [audience]. use a workflow: run three research jobs in parallel, one profiles the buyer and collects the exact words they use, one maps where these buyers spend time online, one collects how competitors pitch them, merge into a one-page positioning doc and pause to show it to me, then run three writers in parallel from that doc: landing page copy, a week of launch posts, a set of outreach messages, then run a checker that compares every asset against the positioning doc and flags anything off, save everything to launch-kit/ and change nothing after that without asking me" ## the playbook - fake edges first: nothing new runs until the no-work arrows are gone - split only what never reads back, one agent keeps everything step-by-step - no finding travels unchecked, and no two checkers ask the same question - every loop has a maximum number of rounds - one writer per file, the plan lives in written steps, the AI fills the jobs - the last yes sits exactly where a mistake would be expensive to undo - one tool is enough until you can name the reason it isn't skip that last yes and the graph ships its first confident mistake straight to a customer draw your current AI system tonight, just the jobs and the arrows between them count the fake edges, then delete them that is the first move of the whole craft, it costs nothing, and it usually removes more waiting than any tool you could buy ## 相关链接 - [Machina](https://x.com/EXM7777) - [@EXM7777](https://x.com/EXM7777) - [234K](https://x.com/EXM7777/status/2079934660982047021/analytics) - [weeklyaiops.com](https://weeklyaiops.com/) - [Upgrade to Premium](https://x.com/i/premium_sign_up) - [10:20 PM · Jul 22, 2026](https://x.com/EXM7777/status/2079934660982047021) - [234.4K Views](https://x.com/EXM7777/status/2079934660982047021/analytics) - [View quotes](https://x.com/EXM7777/status/2079934660982047021/quotes) --- *导出时间: 2026/7/23 10:58:59* --- ## 中文翻译 # 如何精通图工程（完整课程） **作者**: Machina **日期**: 2026-07-22T14:20:41.000Z **来源**: [https://x.com/EXM7777/status/2079934660982047021](https://x.com/EXM7777/status/2079934660982047021) --- ![Image](/doc/3297/image/img_c14c7a6693dd.jpg) 我将向你展示如何构建你的第一个智能体图，并让它今天就为你的业务工作：这是一个 AI 智能体团队，它们并行进行研究，试图推翻自己的发现，并将最终合并的结果交给你，而你的批准是任何事情发布前的最后一道关卡本课程涵盖整个方法：什么是图，那个能回本的模式，保护你账单的停止规则，以及为你的业务提供的 3 个完整构建：一个研究台，一个 SEO 内容机器，以及一个上市工具包在开始之前先简单交代一下背景，因为这个趋势是有故事的构建 Claude Code 的工程师说他不再直接提示它了，他编写循环来提示它，几周后，互联网的同一个角落宣称循环已是旧闻，并转向了图在所有的喧嚣之下，这是一种真正的手艺，它比帖子看起来要简单，而且你不需要成为工程师也能掌握其中的任何部分课程地图： - 第 1 课：图到底是什么 - 第 2 课：钻石模式 - 第 3 课：停止规则 - 第 4 课：人工关卡 - 图 1 - 深度研究台（构建 + 提示） - 图 2 - SEO 内容机器（构建 + 提示） - 图 3 - 上市工具包（构建 + 提示） - 实战手册如果你想深入本文之外，了解如何将这样的系统转化为金钱，这正是我建立实时 AI 运营社区的目的：weeklyaiops.com ## 第 1 课：图到底是什么图是你 AI 工作的蓝图，画出来以便你查看它回答两个问题：哪些工作需要发生，以及哪个工作必须等待哪个工作每个工作都是你会交给一个助手的任务，比如研究竞争对手、写一份草稿或核实一个声明当一个工作在开始之前需要另一个工作的结果时，你用箭头将两者连接起来一小份运行笔记会随工作一起流动，记录着发现了什么、决定了什么以及还有什么待办我把这整套术语变成了下面的卡片，保存一次，你就再也不需要定义了： ![Image](/doc/3297/image/img_0e88e314864b.jpg) 现在我建议你今晚做一件事，在你接触任何新工具之前看看你今天运行的 AI 系统，对于其中的每一个“然后”，问问下一个工作是否需要上一个工作的结果只有当工作流经时，箭头才是真实的 “总结这个文件，然后检查我的日历”听起来像是一个序列，但日历检查步骤从来不需要摘要，所以这两个工作从来不需要互相等待当答案是“否”时，箭头就是假的，等待也是浪费，而在你画的任何系统中，你都会发现两三个这样的虚假边现在你的系统作为一条直线运行，每个工作都在等待前一个，这虽然行得通，但这可能是运行事情的最慢方式，因为一个卡住的工作会挡住它后面的一切在我们构建之前还有一件事：工程师们在几小时内就对这个趋势提出了反对意见，指出这是一个戴着新名字的几十年前的模式，他们是对的我将其视为好消息，因为一个运行关键系统几十年的模式正是你想要信任并用于你业务的东西 ## 第 2 课：钻石——唯一的回本模式观察任何严肃的智能体系统工作，同样的画面会不断出现工作分叉，多个工作者并排挖掘，有什么东西检查他们的发现，然后一切合并回一个答案那个画面被称为钻石模式，我有信心说这是你今年唯一需要的模式 Claude 内部的研究功能在生产环境中正是这样运行的，一个负责人规划，工作者并行收集，发现在到达你之前会经过检查现在是大家都跳过的部分，即检查步骤，你应该将其视为不可妥协的每一次关于 AI 自我审查的严肃测试都得出同样的结论：模型会遗漏它们自己的大部分错误所以你永远不应该让同一个智能体批改它自己的作业把检查交给一个单独的工作，其唯一的任务就是在弱发现到达你之前将其扼杀，并给每个检查者一个不同的问题……一个问它是否正确，一个问它是否最新，一个问来源是否真实 ![Image](/doc/3297/image/img_fe9ced6e62cb.jpg) ## 第 3 课：停止规则更多的智能体并不自动意味着更好，而这是课程中真正能为你省钱的部分当研究人员给一个智能体团队和一个单一智能体同样的预算时，在分裂成独立部分的工作上，团队明显获胜；而在每一步都需要前一步的工作上，团队则同样明显地落败所以这里是停止规则图购买的是广度，它买不来更好的判断力我的建议很简单：当每一步都需要全貌时，你应该坚持使用一个智能体；而当工作分裂成从不读取彼此结果的任务时，那就是图开始回本的时候所以在你添加一个智能体之前，问问那个唯一决定账单的问题……我的工作在哪里分裂 ## 第 4 课：人工关卡你是自己图中最重要的节点，而这节课是关于你到底应该站在哪里每个严肃的系统在任何不可逆转的操作之前都会路由给人类，因为发送、发布、退款和发票是应该终结于你的箭头，那个批准是草稿管线与现实世界之间的最后一道“是” 就连 Klarna 也公开地吸取了这一教训，在 AI 支持上全力以赴，承认削减成本走得太远，并将人工服务作为高级层级带了回来设计规则只需一行把你的批准放在撤销代价高昂的地方，而不是每一步什么都设关卡会让你成为瓶颈，什么都不设关卡则意味着没人看着我还建议你用那些无法反驳的数字来评判整个系统：到账的钱、留住的客户、运行的测试，因为一个只给自己报告打分的系统会自信地犯错四条小规则能防止一切变成昂贵的意外，下面的卡片是你保存的版本：每个循环都有最大轮数，只有一个工作写入任何一个文件，路由存在于书面步骤中而 AI 填充工作，并且对于可以生成多少智能体总是有一个上限 ![Image](/doc/3297/image/img_f111423a4d79.jpg) ## 构建篇——三个能回本的图这里的每一个构建都在 Claude Code 中运行，以你需要粘贴的确切提示结束，并让你作为关卡每个提示中的一个词起到了关键作用：“workflow”（工作流）告诉 Claude 构建一个协调的团队，而不是通过单行步骤来工作 ## 图 1 - 深度研究台想想你最后一次背后有真金白银支持的决策，比如价格变动、新优惠、你差点进入的市场你可能是在一周的谷歌搜索后决定的，或者是支付了昂贵的分析师账单，而这个构建可以同时替代这两者你的问题分裂成五个角度，五个研究人员同时挖掘，一个怀疑论者攻击每一个发现，只有幸存者带着来源进入最终报告 ![Image](/doc/3297/image/img_061d8aa397ea.jpg) 1. 在一个空文件夹中打开 Claude Code 2. 用一句话写下你真实的问题，就是那种背后有钱的问题……“我应该提价吗”或“这两个优惠我该推出哪一个” 3. 将下面的提示连同你的问题一起粘贴进去 4. 观察发散过程，五个研究人员并行工作，然后是怀疑论者的审查 5. 阅读报告，只有通过你阅读的内容才会进入决策提示： “我需要针对以下问题的决策级研究：[你的问题]。使用工作流：将问题拆分为 5 个不同的角度，为每个角度并行运行一个研究员，每个发现都需要来源链接和日期，然后针对每个发现运行一个怀疑论者试图反驳它，丢弃未通过的，将幸存者合并为一个按置信度排序的报告，将其保存为 research-report.md 并向我展示主要发现” 你的第一次运行需要几分钟而不是几天，我觉得怀疑论者审查才是这里真正的产品，因为它是研究和收集谣言的区别 ## 图 2 - SEO 内容机器诚实一点地审视一下你的发布时间表，因为对于很多企业来说，它要么需要每月的预付费用，要么干脆就不发生了这个构建每次运行写一篇排名就绪的草稿，并且没有你就绝不发布三个研究人员并排工作，一个负责当前顶页的内容，一个负责人们问的真实问题，一个负责每个人都错过的东西，然后他们的工作合并成一个大纲，草稿被写出并经过事实核查，然后它在你的文件夹中等待你的“是” ![Image](/doc/3297/image/img_2783a3882e91.jpg) 1. 挑选你的客户会在谷歌中输入的那一个主题 2. 粘贴下面的提示 3. 三个研究工作同时运行：顶页、真实问题、空白点 4. 草稿落入你的草稿文件夹，每个声明都经过核查，弱的那些在顶部被标记 5. 你阅读它，修正只有你知道的事情，并自己发布 6. 然后让 Claude 保存工作流，到了每个周一这就变成一个命令而不是一个下午提示： “我想要一篇能为以下主题排名的文章：[主题]。使用工作流：并行运行三个研究工作，一个列出当前排名靠前的页面涵盖的内容，一个收集人们关于这个主题问的真实问题，一个找出顶页跳过的内容，将所有三个合并成一个大纲，根据大纲写出完整的草稿，然后运行一个事实核查器，标记每一个没有来源的声明，将草稿保存到 drafts/ 文件夹，并将被标记的声明列在顶部，绝不发布任何东西” ## 图 3 - 上市工具包你要发布某样东西，一个产品、一个优惠、一项服务，而诚实的选项从来都是数周的准备或是盲目发布这个构建在一次运行中产出完整的工具包，并由你批准每一部分三个研究人员并行分析买家、渠道和竞争对手，他们的工作合并成一页定位文档，系统在那里暂停以便你阅读，因为那一页将成为每个资产据以检查的基础然后三个撰写者并行起草着陆页文案、发布帖子和外联消息，一个检查器标记任何偏离定位的内容，工具包落入一个文件夹，没有任何东西会自动上线 ![Image](/doc/3297/image/img_0c9259e66ea1.jpg) 1. 命名产品以及它针对的人，各一行 2. 粘贴下面的提示 3. 当系统暂停时阅读定位页面，这是我绝不会跳过的一步 4. 让三个撰写者并行产出资产 5. 审查检查器标记的内容 6. 逐个批准，没有你什么都不会发货提示： “我正在为 [受众] 发布 [产品]。使用工作流：并行运行三个研究工作，一个分析买家并收集他们使用的确切词汇，一个绘制这些买家在线上花费时间的地方，一个收集竞争对手如何向他们推销，合并成一页定位文档并暂停向我展示，然后基于该文档并行运行三个撰写者：着陆页文案，一周的发布帖子，一套外联消息，然后运行一个检查器，将每个资产与定位文档进行比较并标记任何不一致的地方，将所有内容保存到 launch-kit/ 文件夹，未经询问我不得更改任何内容” ## 实战手册 - 先消除虚假边：直到无工作的箭头消失前，不运行任何新东西 - 只分裂那些从不回读的内容，一个智能体保持一切按部就班 - 没有发现能在未检查的情况下通过，且没有两个检查者问同一个问题 - 每个循环都有最大轮数 - 每个文件一个撰写者，计划存在于书面步骤中，AI 填充工作 - 最后的“是”恰好位于撤销代价高昂的地方 - 一个工具就足够了，直到你能说出它不是的原因跳过那个最后的“是”，图就会把它第一个自信的错误直接发给客户今晚画出你当前的 AI 系统，只是工作和它们之间的箭头数出虚假边，然后删除它们这是整个手艺的第一步，它不花一分钱，而且通常比你能买到的任何工具都能消除更多的等待 ## 相关链接 - [Machina](https://x.com/EXM7777) - [@EXM7777](https://x.com/EXM7777) - [234K](https://x.com/EXM7777/status/2079934660982047021/analytics) - [weeklyaiops.com](https://weeklyaiops.com/) - [Upgrade to Premium](https://x.com/i/premium_sign_up) - [10:20 PM · Jul 22, 2026](https://x.com/EXM7777/status/2079934660982047021) - [234.4K Views](https://x.com/EXM7777/status/2079934660982047021/analytics) - [View quotes](https://x.com/EXM7777/status/2079934660982047021/quotes) --- *导出时间: 2026/7/23 10:58:59*