After these three cards, you'll understand AI better than 90% of people. Not because you know more facts — because you know how to look.
先看地图：领域 × 交互模式
The Map: Domain × Interaction
横轴 = 你跟 AI 的交互方式。纵轴 = AI 在什么领域工作。每个工具都能在里面找到自己的位置。
X-axis = how you interact with AI. Y-axis = what domain AI works in. Every tool finds its position here.
领域↓ 模式→Domain↓ Mode→
💬
Chat
你说，AI回应
You speak, AI replies
🤖
Agent
你说目标，AI自主执行
You set the goal, AI runs it
⚙️
Workflow
拖拽搭自动化
Drag-drop automation
⚡
Vibe Coding
自然语言造东西
Build with plain language
🧠思考 & 文字Thinking & Text
ClaudeChatGPTGeminiPerplexityKimi
CoworkNotion AIManus
n8nMakeZapier
CursorClaude Code
🎨图像 & 设计Image & Design
MidjourneyDALL·EFluxIdeogramNanobanana
Canva AIFigma AIAdobe Firefly
ComfyUIn8n+API
Framer AICursor+Canvas
🎬视频 & 动态Video & Motion
KlingRunwaySeedanceSoraVeoPika
HeyGenDescriptRunway Act-One
RemotionDify+API
Cursor+Remotion
🎵音频 & 语音Audio & Voice
SunoUdioElevenLabsNotebookLM
DescriptAdobe Podcast
n8n+ElevenLabs
正在涌现中
Still emerging
📊数据 & 分析Data & Analytics
ChatGPT DataClaude ArtifactsJulius AI
CoworkRows AI
n8n+DBRetool AI
Cursor+PythonReplit
怎么用这张表：问自己两个问题——我的工作在哪个领域（行），我现在用到了哪个阶段（列）。下一步不一定是往右移一格——Workflow 和 Build（Vibe Coding）是两条平行路径，很多人会直接从 Chat/Agent 跳到 Build。How to use this map:Ask yourself two questions—which domain is my work in (rows), which stage am I using now (columns). The next step isn't always moving right one column—Workflow and Build (Vibe Coding) are parallel paths. Many people jump directly from Chat/Agent to Build.
判断一
Judgment One
Vibe Coding 正在吃掉 Workflow
Vibe Coding Is Eating Workflow
你用自然语言描述需求，AI 写代码实现。你不需要懂代码——你需要懂需求和验收。这不是"程序员的工具"，这是产品经理和甲方的超能力。You describe what you want in plain language, the AI writes the code. You don't need to know code — you need to know what you want and how to review it. This isn't a programmer's tool. It's a superpower for PMs and decision-makers.
能力光谱：Claude Artifacts（对话框里秒出原型）→ Cursor/Windsurf（本地多文件项目）→ Claude Code/Cowork（命令行级全栈开发）。
关键判断：同一个目标——比如"每天自动汇总客户反馈发到 Slack"——你可以用 n8n 拖拽搭，也可以用 Cursor 让 AI 写脚本。Workflow 工具的上限是平台支持什么；Vibe Coding 的上限是你的想象力。越来越多人发现 Vibe Coding 比拖拽更快更灵活。但需要稳定运行+有现成集成的场景，Workflow 仍然有价值。The spectrum: Claude Artifacts (instant prototypes in a chat box) → Cursor/Windsurf (local multi-file projects) → Claude Code/Cowork (command-line full-stack dev).
The key judgment: Same goal — say "auto-summarize customer feedback to Slack daily" — you can drag-and-drop it in n8n, or ask Cursor to write a script. Workflow tools are capped by what the platform supports; Vibe Coding is capped by your imagination. More and more people find Vibe Coding faster and more flexible. But Workflow tools still earn their keep when you need rock-solid stability and ready-made integrations.
正确姿势：① 从小做起——先做一个有用的小工具 ② 描述清楚验收标准，不要说"做一个CRM" ③ 迭代，不要重写 ④ 不懂代码也能验收——打开用、看结果对不对、告诉AI哪里不对How to do it right: ① Start small — one useful tool first ② Spell out acceptance criteria, don't say "build me a CRM" ③ Iterate, don't rewrite ④ You can review without knowing code — open it, use it, say what's wrong
判断二
Judgment Two
Agent 不是新产品，是一种设计模式
Agent Isn't a New Product — It's a Design Pattern
Chat = 你指挥每一步。Agent = 你只说终点，它自己开车。用的还是同一个 AI 模型，区别在于怎么用。Chat = you direct every step. Agent = you name the destination, it drives. Same underlying model. The difference is how you use it.
四种模式：
① Chat——你问一句它答一句。最基础。
② Copilot——AI 嵌入你正在用的工具，实时辅助。你主导。
③ Agent——你给目标，它自己规划步骤、调用工具、执行到底。
④ Agentic Workflow——多个 Agent 串联，自动运行的系统。
"Agentic"是形容词，描述一种设计模式——不是具体产品。任何"能自主规划+调用工具+执行到底"的系统都叫 Agentic。传统自动化是"如果A则B"的死规则，Agentic 系统能根据情况灵活决策——这是 AI 带来的本质变化。Four modes:
① Chat — you ask, it answers. The most basic.
② Copilot — AI embedded in the tool you're using, assisting in real time. You lead.
③ Agent — you state a goal, it plans steps, calls tools, executes end-to-end.
④ Agentic Workflow — multiple Agents chained into an autonomous system.
"Agentic" is an adjective — it describes a design pattern, not a specific product. Any system that "plans autonomously + calls tools + executes through to the end" is agentic. Traditional automation is rigid "if A then B." Agentic systems make flexible decisions based on the situation — that's the fundamental shift.
Agent 的三个核心能力：规划（拆目标为步骤）· 工具调用（靠 MCP 和 Skills）· 自我纠错（发现不对就换策略）。缺任何一个，Agent 就是"很贵的自动回复机器"。Three core capabilities: Planning (breaking goals into steps) · Tool Use (via MCP and Skills) · Reflection (recognizing errors, switching strategy). Take any one away and the Agent becomes "an expensive auto-responder."
判断三
Judgment Three
选 AI 平台，看它能连接什么
When Picking an AI Platform, Ask What It Connects To
一个只能打字的 AI 和一个能读你 Slack、查你 Drive、帮你发邮件的 AI，是完全不同的物种。前者是聊天工具，后者是工作搭档。让 AI 从"聊天"变成"干活"的关键 = MCP + Skills + Plugins。An AI that can only type and one that reads your Slack, searches your Drive, and sends your emails are completely different species. The first is a chat tool. The second is a work partner. What makes the leap: MCP + Skills + Plugins.
Skills = AI 的专业能力包。MCP 给了 AI "手"，Skills 给了 AI "专业知识"。
Plugins = MCP + Skills 的一键安装包。装一个 Slack 插件，AI 同时获得连接能力和写好消息的知识。MCP (Model Context Protocol) = an open standard that lets AI connect to external tools. Think USB-C — one universal interface, many tools can plug in. With MCP, AI can read Slack, search Notion, query Drive, manipulate Figma.
Skills = packaged expertise for the AI. MCP gives it "hands"; Skills give it "know-how."
Plugins = MCP + Skills bundled. Install a Slack plugin and the AI gets both the connection and the craft.
选平台的逻辑：不要只看"聊天有多聪明"——要看它能连接什么。优先选 MCP 生态丰富的平台。就像选手机看 App Store——平台只是载体，插件决定它能帮你做多少事。The logic for picking a platform: Don't just ask "how smart is the chat?" — ask what can it connect to? Pick the platform with the richest MCP ecosystem. Like choosing a phone for its App Store — the device is just a vessel; the plugins determine how much it can do.
评估新工具的 5 个问题
5 Questions to Evaluate Any New AI Tool
① 它在坐标系哪？① Where on the map?
那个格子你有趁手的了吗？Do you already have something solid there?
② 能嵌进现有工作流吗？② Can it slot into your workflow?
有 API？支持 MCP？API? MCP? Standard exports?
③ 上手到出效果多久？③ Time to first value?
10分钟见效 = 试。学3天 = 观望10 min = try. 3 days = wait
④ 护城河在哪？④ Where's the moat?
Claude/ChatGPT 明天上线同功能？Will Claude/ChatGPT ship this tomorrow?
⑤ 数据去了哪？⑤ Where does your data go?
免费版通常训练你的数据。API版不会。企业版有合同隔离。处理机密 → 硬标准。Free tiers usually train on your data. API tiers don't. Enterprise tiers are contractually isolated. Handling confidential data → non-negotiable.
💬 Large Language Models (LLM) — the engine behind Chat
Core: given what came before, predict the most plausible next word. It has read enormous amounts of text and learned the patterns of language.
→ This is why it sometimes "confidently makes things up." It optimizes for sounding plausible, not being correct.
🎨 Image generation — Midjourney / DALL·E / Flux
Powered by diffusion models: starts from pure noise and "denoises" step by step into a clear image. Your prompt is the direction.
→ A sculptor facing a block of marble. Your prompt is the design sketch.
🎬 Video generation — Kling / Sora / Seedance
An upgrade of image diffusion — generates multiple frames coherent across time. Requires a sense of physics and motion.
→ Stills are stunning. Complex motion is still weak — but improving fast.
💻 AI Coding — Cursor / Claude Code
Still an LLM — because code is also a language, and one with more regular patterns. AI Coding tools add filesystem access, terminal, and editor integration.
→ Vibe Coding: you're the PM, the AI is the engineer.
🤖 Agent — not a new model, a new way of using one
Chat = ask and answer. Agent = state a goal, it plans steps, calls tools, executes end-to-end.
→ "Agentic" is an adjective — a design pattern, not a new technology.
① 数据会被用来训练吗？免费版通常会，API版不会，企业版有合同保护。 ② 存储在哪？大多数服务器在美国。受 GDPR/数据安全法约束的业务要确认合规。 ③ 员工怎么管？最大风险是员工把机密粘贴进免费 ChatGPT。一页 A4 纸的使用准则就够。
☁️ Cloud vs 🏠 Local
Cloud: how most AI tools run. No powerful machine needed, always available. Data passes through someone else's servers. Local: download and run on your machine. Data never leaves. Requires a GPU, usually behind the frontier.
🔓 Open source vs 🔒 Closed source
Open source (Llama, Stable Diffusion, Mistral): free, customizable. Requires technical skill. Closed source (Claude, GPT-4, Gemini): strongest performance, constantly updated. Vendor lock-in. Many companies use both.
🛡️ Three data-security questions
① Training? Free tiers usually yes, API tiers no, enterprise contractually protected. ② Where stored? Most servers in the US. Check compliance for GDPR-bound businesses. ③ Employee management? Biggest risk: staff pasting confidential material into free ChatGPT. A single A4 page of usage rules is enough.
提高员工工作效率Boost team efficiency做客户提案Create client proposals批量生成内容Batch content creation分析数据出报告Data analysis & reports生成图片视频Generate images & video自动化重复工作Automate repetitive work用AI造一个小工具Build a tool with AI快速学习新领域Learn a new field fast
Perplexity 搜最新市场数据 → 喂给 Claude 做结构化分析 → NotebookLM 多文档综合。
PerplexityClaudeNotebookLM
1-2小时 · 更扎实
🚀 系统化
定制决策 Agent + 数据接入
连接公司内部数据+外部市场数据的分析 Agent。每周自动推行业变化简报。
Cowork+MCP内部数据源
2-4周 · 长期战略工具
Sample PromptI'm a [role] in [industry]. Evaluating whether to [decision]. Please: 1. List 5 critical dimensions 2. Pro/Con for each, backed by data 3. Final recommendation, confidence 1–10, biggest unknown
⚡ Do it today
Just ask Claude
Describe the decision in detail, ask it to build a framework and list Pros/Cons.
Claude
10 min · free
🔧 Get serious
Perplexity + Claude + NotebookLM
Perplexity for data → Claude for structured analysis → NotebookLM for multi-doc synthesis.
PerplexityClaudeNotebookLM
1–2 hours · solid
🚀 Systematize
Custom decision Agent + data
Analysis Agent connected to internal + external data. Weekly auto-briefings on industry shifts.