LovisMind

第 03 章

把 AI 用成工作流：十一个常见问题

从给上下文、让回答不泛，到目标、日历、邮件与自动化，一套可照着做的日常 AI 使用方法。

Using AI as a Workflow: Eleven Common Questions

From supplying context and escaping generic answers to goals, calendars, email and automation — a set of practices you can actually run.

首发于 2026.08.19中 / EN
技术教程ai-workflowproductivityprompting
语言
字号
目录
篇内目录

这份资料解决什么问题

Dan Martell 的公司每天在造 AI 产品，他把被问得最多的问题集中回答了一遍。开场的比喻说明了多数人的用法：把 AI 当作花哨的搜索引擎，就像请了私厨却只让他做花生酱三明治。

他的答案不是「用哪个工具」，而是一组可以照做的流程：先把上下文喂足，再用结构化的提示拿到具体输出，然后把重复的部分固化成技能或自动化。工具清单会变，流程不会。

时间轴

时间内容
0:00多数人把 AI 当搜索引擎用
0:40用语音输入代替打字，让 AI 主动反问要上下文
1:49回答为什么总是很泛：它在平均所有答案
2:24MAPS 框架：任务、请求、参数、形态
3:29参数可以给到两三本书的信息量
4:20每次开新对话都要重新自我介绍的问题
6:35用 AI 加速学习：先找最新资源，再生成学习计划
7:08让它直接教，跳过无关章节
8:47五年目标：SMART 化，再生成愿景图
10:45靠系统而非意志力坚持目标
11:16每日仪表盘与红黄绿信号
12:50用日历反推目标，让 AI 当严厉导师
15:03让 AI 写出像你的邮件：先生成语气文档
16:10重复任务：做过两次就该自动化
17:35笔记散乱的解法：先定一个中枢库
19:27财务数据接入后的提问方式
21:24收藏的教程如何真正落地
22:49搭建个人 AI 助手的三层结构
25:02代码报错：先让错误可见，再交给 AI

一、给足上下文：说话比打字快

第一个问题的痛点是：想给 AI 足够背景，又不想打那么多字。

他的原则是口述——他说话的速度约为打字的三倍，所以无论在聊天工具、编程平台还是智能体里，他都用语音转文字把话讲进去，顺便积累了上下文。

第二个技巧是把主动权交出去：在提示末尾加一句「如果你需要更多上下文，就问我」。让 AI 主动提问，比你去猜它需要什么更省事。多轮的多选、清单式提问，也比一次性塞一大段描述更有效。

二、让回答不泛：MAPS

AI 被训练在人类已有的全部语料上，默认行为是把「最好的答案」取平均——所以不加约束时，输出必然泛泛。

他的框架叫 MAPS：

  • M — Mission（任务）：这次要达成什么，比动作本身更上位。不是「写个文案」，而是「帮我获取更多客户」。
  • A — Ask（请求）：具体要它产出什么，比如一份可外呼的潜在客户名单、一个收集邮箱的落地页。多数人只写了这一层，所以结果很泛。
  • P — Parameters（参数）：你手上一切能描述处境的信息——过去的客户资料、已经奏效的脚本、产品文档。他提醒：你可以给它两三本书的信息量，而大多数人只给一两段。
  • S — Shape（形态）：输出长什么样——表格、清单、还是一句话答案。

参数不够时，正好用口述把脑子里的背景一股脑说出来。另外有个技巧：用能切换多个模型的服务，把同一个问题交给不同模型，比较结果，相当于开一个小型评审会。

三、让 AI 认识你：master prompt

每次新对话都像在跟陌生人说话，是一个被普遍忍受的问题。

他的做法从「回答问题」改成「让工具自己去看」：

  1. 把存有你的信息的工具接进来——邮箱、云盘、笔记、Slack。
  2. 让它据此生成一份尽可能详细的 master prompt，描述你是谁、在做什么项目、身边有谁；不准确的地方手动改。
  3. 把这份文档存进 AI 的设置，或复制到其他工具里。

这样每次新对话都自带背景，回答就不再像通用建议。他提到自己选工具时会考虑数据信任问题：如果要把财务信息和内部文档交给模型，供应商是否拿你的数据做别的用途，是决策的一部分。

四、学习与目标

学得更快。 他给自己定的目标是「任何主题 30 分钟内入门，快 20 倍」。步骤是：让它先找这个主题最好的、最新的资源（旧资料在快速变化的领域没有价值），再据这些资源生成学习计划并直接开讲。这样跳过无关章节，只学你真正需要的部分。工具上他提到 NotebookLM——能搜到资源、生成可对话的学习材料，还能输出音频、演示、测验。

五年目标。 目标是看不清的靶子，而多数人的目标太模糊（「多赚钱」「更健康」）。做法分三步：先要求它把目标改写成 SMART（具体、可衡量、可实现、相关、有时限）；再用语音回答它提出的追问；最后让它把达成目标后的生活描述成画面——看见什么、听见什么、什么感觉——并生成图片。一个实用细节：把图改成手机竖屏比例做壁纸，于是目标每天在眼前出现上百次。

五、执行与坚持

坚持靠系统，不靠意志力。 他给的三步是：

  1. 把目标交给 AI；
  2. 让它拆成带时间线的书面计划；
  3. 让它生成一个每日仪表盘：写下今天必须做的活动，并录入实际完成量；系统按剩余时间计算进度，用红黄绿标出状态——低于 10% 是红，差值在 10% 以内是黄，符合进度是绿。

他的观察很直白：如果界面又丑又难懂，没人会每天打开；把数据做得好看、游戏化，使用率才会上去。进阶做法是把仪表盘做成手机应用或保存到主屏的网页。

日历。 他的观点是：目标不是你想要的东西，而是时间流向的副产品。所以他会把所有事情放进日历，用它审计自己。做法是把日历和季度目标一起交给 AI，并给出明确的指令——像一位不留情面的导师那样审视我的日历，指出该删、该移的时间，并说明理由；标准是「对我的目标最好」，而不是「我现在想要」。

他提到多数创始人每周大约有 10 小时花在并不需要自己做的事上。

六、让输出像你：邮件

AI 写的邮件有明显的「AI 味」。修正方法是让它学你的语料：把邮箱接进来，让它分析你写过的东西，输出一份「语气文档」——说明你对老板、对朋友、对客户分别怎么写。之后每次写邮件都让文档兜底，并定期用新邮件更新它。

七、把重复工作交给 AI

他的规则借自编程：DRY，don't repeat yourself。同一件事做过两次，就停下来，把它变成自动化。

具体流程是：录屏并一边做一边口述你在做什么、为什么这么做；然后把这段记录交给 AI，转成一个可复用的「技能」；最后把它挂在定时任务上。

这一步的差别在于「接管方式」：有些工具是接管整块屏幕替你操作，你得坐着看它跑；也有工具把录制下来的流程放到后台执行并保存为技能，真正做到设好就忘。

八、信息与财务

笔记散乱。 三步：先选定一个中枢库，把所有零散信息汇进去；再让其他 AI 工具都指向它；最后让智能体在夜里自动整理——删掉不需要的、把该连的连起来、把资料放到其他智能体能取用的位置。他自己用 Obsidian 作为个人中枢，团队用 Notion。

财务。 障碍不在工具，而在数据：多数记账软件对 AI 不友好，多数经营者也不愿意把数据整理清楚，于是垃圾进、垃圾出。做法是先把银行流水、支付系统等接进 AI，然后直接提问——现金够不够做这笔投入、当前烧钱速度是多少、库存还能撑几个月——并要求用图表呈现。

收藏的教程不落地。 他把「收藏等于学习」称为「有组织的拖延」。解决办法是把链接交给 AI，问一句「这对我有用吗」——它知道你的目标和现有系统——有用就立刻落地；再进一步，把这个判断做成自动化技能，以后见到类似的链接自动处理。

九、助手与排错

搭建 AI 助手。 分三层：接好所有系统的 API；用一层 agent 把它们串起来持续运行（看日历、看邮件、看 Slack）；最后给这个助手一个身份——行为规则与安全边界。他区分了两种形态：给自己用的「单机模式」，和给团队用的「多人模式」——后者要把公司数据整合、清洗，形成统一的「事实来源」，再在其上做看板与任务执行。

代码报错。 核心是一句话：把问题展示给 AI，它才可能解决。找不到错误信息时，就让它生成一个开发时可见的错误窗口，截图发给它；修不好时，把截图交给另一个模型，让两个模型互查。

关于工具推荐的一点提醒

作者在片中多次推荐的工具中，Apex、Precision、HelloFrank 是他自己公司的产品（他也做了说明），《Buyback Your Time》是他的书。这些推荐不妨当作「他实际怎么用」的样本，而不是中立评测；方法本身与具体工具是分开的。

理解检验

  1. 为什么 AI 的默认回答通常很泛？ 参考答案：它被训练在人类已有的全部语料上，默认会把「最好的答案」平均化，缺少你处境里的具体信息；因此需要在提示里补充任务、参数与输出形态。

  2. MAPS 四个字母分别要求你补充什么？ 参考答案：Mission 是要达成的目标；Ask 是具体要产出的东西；Parameters 是与你处境相关的资料（可以给到两三本书的量）；Shape 是输出的形式。

  3. 怎么解决「每次新对话都要重新自我介绍」？ 参考答案：把邮箱、云盘、笔记等工具接入，让它据此生成一份详尽的 master prompt，人工修正后存进设置，并复制到其他 AI 工具中。

  4. 每日仪表盘的红黄绿分别代表什么？ 参考答案：低于应完成进度的 10% 为红（明显落后）；与应完成量相差在 10% 以内为黄；符合进度为绿。

  5. DRY 规则在 AI 工作流里怎么落地？ 参考答案：同一操作出现第二次就停止手动重复，录屏并口述过程，让 AI 转成技能，再挂到定时任务上自动执行。

  6. 财务数据为什么很难直接用 AI 处理？ 参考答案：多数记账软件缺少可用的 AI 接口，经营者也常不愿整理数据，导致输入质量差。解法是先把银行流水与支付系统接入，再做提问与可视化。

  7. 代码报错时，第一步该做什么？ 参考答案：先让错误可见——生成开发时可截图查看的错误窗口，或直接截图错误信息交给 AI；仍不奏效时，把截图交给另一个模型交叉验证。

与这本书其他内容的关系

第一章讲如何把想法提炼成别人听得懂的形状，这一章是把同样的思路用到与机器协作上：先给足上下文（相当于「参数」），再明确输出形态。黄仁勋那一章谈的是系统层面的判断，这一章则是把它落到日常：把重复劳动识别出来，交给可以重复执行的系统。

What this material addresses

Dan Martell builds AI products every day, and here he answers the questions he is asked most. The opening image sets up the problem: using AI as a fancy search engine is like hiring a private chef and only ever ordering a peanut butter sandwich.

His answers are less about which tool to pick and more about processes you can copy: supply enough context, use a structured prompt to get specific output, then turn the repetitive parts into skills or automations. Tool lists change; the process does not.

Timeline

TimeTopic
0:00Most people use AI like a search engine
0:40Dictating instead of typing, and asking the AI to ask you for context
1:49Why answers are generic: the model averages everything
2:24The MAPS framework: mission, ask, parameters, shape
3:29Parameters can be two or three books' worth of material
4:20The problem of re-introducing yourself in every new chat
6:35Learning faster: find recent resources, then build a plan
7:08Have it teach you directly, skipping irrelevant chapters
8:47Five-year goals: make them SMART, then generate a vision image
10:45Systems instead of willpower
11:16A daily dashboard with red, yellow and green signals
12:50Auditing the calendar against goals with a ruthless mentor
15:03Emails that sound like you: build a tone document first
16:10Repetitive work: the second time is the cue to automate
17:35Scattered notes: pick one home database
19:27Financial questions once the accounts are connected
21:24Making saved tutorials actually land
22:49Three layers of a personal AI assistant
25:02Code errors: make the error visible, then hand it over

1. Give enough context: talking beats typing

The first question is how to give AI the context it needs without typing pages of text.

His principle is to dictate — he talks about three times faster than he types, so across chat tools, coding platforms and agents he speaks and lets transcription do the work, which also accumulates context over time.

The second trick hands initiative to the model: end a prompt with "ask me questions if you need more context". Letting the AI pull information beats guessing what it needs, and multiple-choice or checklist questions are easier to answer than one long description.

2. Escaping generic answers: MAPS

AI is trained on everything humanity has produced, so by default it condenses and averages the best answer — which is why unconstrained output feels generic.

His framework is MAPS:

  • M — Mission: the goal behind the task. Not "write copy" but "help me get more customers".
  • A — Ask: specifically what to produce, such as a call list of leads or a landing page to capture emails. Most people supply only this layer, which is why answers stay vague.
  • P — Parameters: everything about your situation — past client information, scripts that already worked, product documents. He notes you can hand over two or three full books of material, while most people give a paragraph or two.
  • S — Shape: the form of the output — spreadsheet, bulleted list, or a short answer.

When documents are scarce, dictate instead: talk through everything you know. He also suggests a service that answers with several models side by side, effectively convening a small review panel.

3. Letting AI know you: the master prompt

The familiar annoyance is that every new chat starts from zero, like talking to a stranger.

His approach moves from answering questions to letting the tool look for itself:

  1. Connect the tools where your information already lives — email, cloud drive, notes, Slack.
  2. Ask it to build the most detailed master prompt about who you are, what you are working on and who is in your life; correct the parts it got wrong.
  3. Save that document into the AI's settings, or paste it into your other AI tools.

Every new chat then starts with context, and the answers stop reading like generic advice. He adds a trust consideration: if financial information and internal documents are going in, whether the vendor treats your data as payment for a "free" product is part of the decision.

4. Learning and goals

Learning faster. His target is to learn any topic in under 30 minutes, roughly 20 times faster. The steps: have it find the best recent resources on the topic — old material is not useful in fast-moving fields — then build a learning plan from those and teach it to you. That skips irrelevant chapters so you learn only the part you need. He names NotebookLM as his daily tool: it searches resources, builds a chat-able study set, and outputs audio, presentations and quizzes.

Five-year goals. A goal you cannot see is a goal you cannot hit, and most goals are too vague ("make more money", "get healthier"). Three steps: ask it to rewrite the goal as SMART (specific, measurable, attainable, realistic, timely); answer its follow-up questions out loud with dictation; then ask it to describe what life looks like when the goal is hit — sights, sounds, details, feelings — and turn that into an image. One practical detail: re-render it at phone-wallpaper proportions so the goal is in front of you a hundred times a day.

5. Execution and consistency

Systems instead of willpower. Three steps:

  1. Give the AI the goals.
  2. Have it break them into a written plan with timelines.
  3. Have it build a daily dashboard: list the activities that must happen today, enter what you actually did, and let it compute progress against the time remaining using red, yellow and green — red below 10% of the required activity, yellow within 10% of it, green on track.

His observation is blunt: if the information is ugly and confusing, nobody will open it daily; make it pretty and gamified and you will. The advanced version turns the dashboard into a mobile app or a web page saved to the home screen.

Calendar. His premise: goals are not things you want, they are the by-product of where your time goes, which is why everything in his life is in his calendar so he can audit it. The method is to connect the calendar and give it this quarter's goals, then instruct it to review the calendar like a ruthless mentor — argue about what needs restructuring, say why, and optimise for the goals rather than what he wants right now.

He estimates most founders have about ten hours a week in their calendar on things they do not need to be doing.

6. Output that sounds like you: email

AI-written email has a recognisable texture. The fix is to let it learn your corpus: connect the mailbox, have it analyse how you write, and produce a tone document describing how you write to your boss, your friends and your clients differently. Let that document steady every draft, and refresh it with new messages periodically.

7. Handing repetitive work to AI

His rule is borrowed from programming: DRY — don't repeat yourself. The second time you do something, stop and turn it into an automation.

The process: record your screen while narrating what you are doing and why; hand the recording to the AI to convert into a reusable skill; then schedule that skill to run.

The difference between tools lies in how they take over. Some seize the whole screen and make you sit and watch; others capture the process, run it in the background and save it as a skill, so you can set it and forget it.

8. Information and finances

Scattered notes. Three steps: choose one home database and consolidate everything into it; point your other AI tools at it; then let an agent tidy it overnight — deleting what is not needed, connecting what should be connected, filing things where other agents can reach them. He keeps a personal hub in Obsidian while his team uses Notion.

Financials. The obstacle is not the tool but the data: most accounting software is not AI-friendly, and most operators do not want to organise their numbers, so garbage in, garbage out. The fix is to connect bank statements and payment systems to the AI, then ask directly — is there enough cash for this hire, what is the current burn rate, how many months of inventory remain — and insist on graphs and dashboards.

Saved tutorials that never land. He calls collecting tutorials "organised procrastination". The fix: paste the link into the AI and ask whether it is useful for you — it already knows your goals and systems — then implement it immediately if it is; better still, turn that judgement into an automated skill so future links are triaged automatically.

9. Assistants and debugging

Building an AI assistant. Three layers: connect the APIs of all your systems; run an agent layer that keeps them working together — watching the calendar, the inbox, Slack; then give the assistant an identity, with behaviour rules and security boundaries. He separates two forms: single-player mode for yourself, and multiplayer mode for a team, where company data is integrated, correlated and cleaned into a single source of truth that dashboards and task execution sit on.

Code errors. The core rule is that the AI has to see the problem to solve it. If the error is not visible, have the program generate an error window you can screenshot during development; if the fix still fails, give the screenshot to a different model and let the two check each other.

A note on the tool recommendations

Several tools recommended in the video — Apex, Precision and HelloFrank — are Martell's own companies, which he discloses, and Buyback Your Time is his book. Read those recommendations as a sample of what he actually runs rather than a neutral review; the methods stand apart from the specific tools.

Comprehension check

  1. Why is AI's default answer usually generic? Answer: It is trained on the full human corpus and condenses the best answer into an average, without the specifics of your situation — so the prompt has to supply mission, parameters and output shape.

  2. What does each letter of MAPS require? Answer: Mission is the goal behind the task; Ask is the concrete deliverable; Parameters is material about your situation, up to several books' worth; Shape is the form of the output.

  3. How do you stop re-introducing yourself in every new chat? Answer: Connect email, cloud drive and notes, have the AI generate a detailed master prompt about you, correct it, save it into settings, and paste it into your other AI tools.

  4. What do red, yellow and green mean on the daily dashboard? Answer: Red means below 10% of the required activity; yellow means within 10% of it; green means on track.

  5. How does the DRY rule apply to AI workflows? Answer: The second time a task comes up, stop doing it manually — record and narrate it, have the AI convert the recording into a skill, then schedule that skill.

  6. Why is financial data hard to use with AI? Answer: Most accounting tools lack usable AI interfaces and operators often will not organise the data, so input quality is poor. The fix is to connect bank and payment systems first, then ask questions and visualise.

  7. What is the first step when code throws an error? Answer: Make the error visible — add a development-time error window you can screenshot, or capture the error message and hand it to the AI; if that fails, give the screenshot to a second model.

Relation to the rest of this book

The first chapter is about distilling an idea into a shape others can follow; this chapter applies that to working with machines — supply context (the equivalent of parameters), then define the output shape. The Jensen Huang chapter is about judgement at the systems level; this one brings it down to daily practice: identify repetitive work and hand it to a system that can run it again.

来源