一份给国内朋友的实操指南
沿着「问答 → 协作 → 委托」三态，把 AI 从"偶尔聊一句"变成"会自己干活的搭档"
不需要会编程，跟着做就行
A hands-on guide for friends in China
Follow the three modes — Ask → Co-create → Delegate — to turn AI from "occasional chatting" into "a partner that gets work done on its own"
No coding required — just follow along
这本书不按"工具清单"来讲——因为工具一直在变，按工具讲很快过时。它按你和 AI 的关系深度分成三态：
① 问答：你主导每一步，随手问、查、翻译。
② 协作：人机一来一回，共创一份产物。
③ 委托：把多步任务交给 AI 自主跑，你只验收。
从"你问它答"到"你交给它办"，这条线才是 AI 协作真正的进化。工具只是填进三态里的例子，换了工具，框架照样成立。每一态都有：推荐工具、操作步骤、费用、以及无法翻墙时的国内平替。
This book isn't organized as a "tool list" — tools keep changing, and a tool list goes stale fast. It's organized by how deep your relationship with AI goes, in three modes:
① Ask — you drive every step: ask, look up, translate.
② Co-create — back-and-forth, building something together.
③ Delegate — hand a multi-step task to AI to run on its own; you just sign off.
From "it answers you" to "it handles it for you" — that line is the real evolution of AI collaboration. Tools are just examples that fill the modes; swap the tool, the framework still holds. Each mode includes: recommended tools, setup steps, costs, and China-accessible alternatives.
交给 Agent 自主跑多步，你只验收（任务型 + 常驻型综合代理）Hand multi-step tasks to agents that run on their own — you just sign off (task agents + standing agents)
人机一来一回，共创产物——分析、写作、做表、出原型Back-and-forth to build something — analysis, writing, sheets, prototypes
你主导每一步，随手问、查、翻译、头脑风暴You drive every step — ask, look up, translate, brainstorm
越往上，你交出去的越多，回报也越高——但前提是你信得过它。The higher you go, the more you hand over and the greater the payoff — but only once you can trust it.
建议从「问答」开始，熟练后逐态往上。
即使只走到「协作」，你的工作效率也会有质的飞跃。
两条线贯穿全书：模型档位（同一工具，选快答还是深度推理）+ 沉淀复利（用过就记下来，越用越省）。
Start with "Ask" and move up as you get comfortable.
Even reaching "Co-create" dramatically boosts your productivity.
Two throughlines run across the book: model tier (same tool — fast answer vs deep reasoning) + compounding capture (write down what works; it pays off over time).
为什么需要？ChatGPT、Claude、Gemini、Cursor 等主力工具均部署在海外服务器，国内无法直接访问。
费用范围：¥15~50/月（年付通常更优惠），部分支持多设备同时在线。
选择建议：
Why needed? ChatGPT, Claude, Gemini, Cursor etc. are hosted overseas and not directly accessible from mainland China.
Cost range: ¥15~50/month (annual plans usually cheaper). Many support multiple devices.
Tips:
必备：
支付方式（用于订阅付费工具）：
Essential:
Payment (for paid subscriptions):
你主导每一步，随手问、查、翻译、头脑风暴You drive every step — ask, look up, translate, brainstorm
三态里最浅、也是入口的一态。像和一个博学的朋友聊天——问问题、讨论想法、翻译、写邮件草稿、头脑风暴。每一步都由你发起，AI 只回答这一句。不需要任何技术背景，打字就行。 The shallowest mode, and the entry point. Like chatting with a knowledgeable friend — ask questions, discuss ideas, translate, draft emails, brainstorm. You initiate every step; AI just answers this one. No technical background needed — just type.
生态最完善、插件最丰富，旗舰对话模型综合能力强，还有单独的「深度推理」档处理难题。支持联网搜索、图片生成、文件上传分析、语音对话。适合作为"第一个 AI 工具"入门。
（截至 2026-06，对话主力为 GPT-5 系，推理档为 o 系；型号迭代很快，认准"旗舰/推理"两档即可。）
Richest ecosystem and plugins; the flagship chat model is a strong all-rounder, with a separate "deep reasoning" tier for hard problems. Supports web search, image generation, file analysis, and voice. Perfect as your first AI tool.
(As of 2026-06, the flagship chat line is GPT-5-class and the reasoning tier is o-class; model names change fast — just remember the two tiers: flagship / reasoning.)
如果你已有 Google 账号，Gemini 可以直接读取你的 Gmail、Google Docs、Drive 中的内容来回答问题。多模态能力强——可以分析图片、理解视频。和 ChatGPT 形成互补。 If you already have a Google account, Gemini can read your Gmail, Google Docs, and Drive to answer questions. Excellent multimodal capabilities — analyzes images and videos. Complements ChatGPT well.
不需要翻墙、不需要付费、中文能力出色。适合作为"第零步"——还没搞定 VPN 之前，先用豆包熟悉和 AI 对话的感觉。 No VPN needed, completely free, excellent Chinese capabilities. Perfect as "Step Zero" — get familiar with AI conversation before setting up VPN access.
人机一来一回，共创一份产物Back-and-forth, building something together
从"问一句答一句"进到"一起做一件事"。你给方向和素材，它出结构和初稿，来回几轮，做出一份真正能用的产物——分析报告、文章、表格、代码原型、竞品拆解。关键变了：你不再只是提问者，而是和 AI 一起对一份东西负责。 From "ask once, answer once" to "build one thing together." You give direction and material; it gives structure and a first draft; a few rounds later you have something actually usable — an analysis, an article, a sheet, a code prototype, a competitor breakdown. The shift: you're no longer just asking, you co-own a deliverable with the AI.
我的主力协作工具。超长上下文（旗舰档可达 100 万 token），整个项目文档一次喂给它分析。Artifacts 能在对话里直接生成可预览的代码、文档、图表；Projects 能为一个项目建专属空间，把背景资料固定下来，每次对话都带着上下文。 My primary co-creation tool. Ultra-long context (up to 1M tokens on the flagship tier) — feed it an entire project's docs at once. Artifacts generates previewable code, docs, and charts right in the chat; Projects gives a project its own space, pinning background material so every conversation carries the context.
✦ 上传 50 页行业报告，提炼 5 个核心结论
✦ 贴入竞品 App 截图做竞品分析
✦ 用 Projects 功能建专属项目空间
✦ 直接生成数据可视化 Dashboard
✦ Upload a 50-page industry report, extract 5 key insights
✦ Paste competitor app screenshots for analysis
✦ Use Projects feature for dedicated project spaces
✦ Generate data visualization dashboards directly
基于 VS Code 的 AI 编程编辑器。即使不是程序员，也能用自然语言让它写 Python 脚本处理 Excel、做网页原型、写自动化工具。 AI coding editor built on VS Code. Even non-programmers can describe what they want in plain language and get working Python scripts, web prototypes, or automation tools.
交给 Agent 自主跑多步，你只验收Hand multi-step tasks to agents that run on their own — you just sign off
三态里最深的一态，也是这两年变化最大的地方。前两态 AI 都在"配合你"，到了委托，你是把一件多步骤的事整个交出去——它自己读资料、自己动手、自己跑完，做完了给你看结果，你只负责验收和拍板。委托又分两类：任务型（坐你电脑旁，接一个任务跑完一轮）和常驻型综合代理（不在你电脑里，7×24 常驻、定时自己干活）。 The deepest mode — and where the most has changed in the past two years. In the first two modes AI assists you; in delegation you hand over a whole multi-step job — it reads, acts, and runs to completion on its own, then shows you the result for sign-off. Two kinds: task agents (sit beside you, run one task to completion) and standing agents (not on your machine — running 24/7, doing scheduled work on their own).
像一个坐在你旁边的 AI 工程师/助手。你给它一个任务，它在你的电脑/项目里读文件、改文件、跑命令、调 API，一口气做完。最适合"有明确目标、要动手干一轮"的活。 Like an AI engineer/assistant sitting next to you. Give it a task; it reads and edits files, runs commands, and calls APIs inside your machine/project, finishing in one go. Best for jobs with a clear goal that need a round of hands-on work.
我整个 AI 工作流的中枢，也是"任务型委托"的代表。它能读取整个项目、自动改文件、跑测试、执行命令、调 API——就像一个坐在你旁边的 AI 工程师。早就不只是写代码：数据分析、内容流水线、知识库维护都能交给它。现在 CLI / 桌面 App / 网页 / IDE 插件都能用，还有更快的 fast 档。 The hub of my entire AI workflow, and the poster child for "task agents." It reads an entire project, auto-edits files, runs tests, executes commands, and calls APIs — like an AI engineer sitting next to you. Long past just writing code: data analysis, content pipelines, knowledge-base upkeep all get handed to it. Now available via CLI / desktop app / web / IDE plugin, with a faster "fast" tier too.
npm install -g @anthropic-ai/claude-codeclaude 启动，登录 Anthropic 账号Run claude to start, login to Anthropic
✦ 内容流水线：选题→生成→质检→多平台改写一条龙，写完自检，留好我要亲笔填的段落，发布我自己点
✦ 投研助理：把"读资料→分析→出报告"整条交给它，我只看结论
✦ 数据分析：用自然语言让它写 SQL、跑查询、出可视化
✦ 知识库维护：自动整理笔记、生成日报、沉淀洞察
✦ Content pipeline: topic→draft→QA→multi-platform rewrite end to end; it self-checks, leaves the passages I must write by hand, and I click publish myself
✦ Research assistant: hand it the whole "read→analyze→report" chain; I just read the conclusions
✦ Data analysis: plain-language → SQL, queries, visualizations
✦ Knowledge upkeep: auto-organize notes, generate daily reports, capture insights
与 Claude Code（本地）互补。Codex 在云端沙箱中运行，适合 GitHub 项目的代码修改和 PR 创建。两者配合：Claude Code 做本地日常开发，Codex 在云端并行跑修改任务。 Complements Claude Code (local). Codex runs in a cloud sandbox, ideal for GitHub code modifications and PR creation. Together: Claude Code for local daily work, Codex for parallel modification tasks in the cloud.
不在你电脑里，7×24 常驻、跨平台、定时自己干活的"AI 管家"Not on your machine — a 24/7, cross-platform "AI butler" that does scheduled work on its own
任务型委托是"你叫它一次、它干一轮"；常驻型则是一个一直在线的代理——它跑在云主机或一台常开的小机器上，连着你的飞书 / Telegram / 邮箱，按你设的时间点自己醒来干活、把结果推给你，做久了还会把经验沉淀成自己的技能。这正好接住了"个人助理"那些活：早间摘要、信息聚合、定时提醒，本来就该由一个常驻的代理来做，而不是你每次手动喊一声。 Task agents run one round when you call them; a standing agent is one that's always on — running on a cloud VPS or an always-on little box, connected to your Feishu / Telegram / email, waking up on your schedule to do work and push results to you, and over time distilling what it learns into its own skills. This is exactly what "personal assistant" work needs: morning briefings, info aggregation, scheduled reminders — jobs a standing agent should own, not ones you trigger by hand each time.
具体用哪个不重要——重要的是认出这一类代理的共同特征：
· 常驻：7×24 在线，不靠某台开着的电脑。
· 跨平台：飞书 / Telegram / Slack / Discord 等多个入口，你在哪它在哪。
· 定时自主：按 cron 自己醒来干活，不用人每次手动触发。
· 自学习：把做过的事固化成技能和记忆，越用越懂你。
OpenClaw、Hermes 都是这一类的代表，挑顺手的用即可。哪天又冒出一个新的，框架照样成立。
Which one you pick matters less than recognizing the traits this category shares:
· Always on: 24/7, not tied to a laptop that has to stay open.
· Cross-platform: Feishu / Telegram / Slack / Discord and more — wherever you are.
· Scheduled & autonomous: wakes on a cron to work, no manual trigger each time.
· Self-learning: distills what it's done into skills and memory, getting to know you over time.
OpenClaw and Hermes are both representatives — use whichever fits. When a new one appears, the framework still holds.
✦ 早间摘要：每天早上自动把昨夜的消息、要点聚合好，推到我飞书
✦ 盯盘提醒：定时扫一遍我关注的标的，有异动主动告警
✦ 定时报告：到点把某份日/周报生成好，直接推到指定的群
✦ 群动态值守：每隔一段时间看一眼群里的进展，达到条件就提醒我
✦ Morning briefing: auto-aggregates overnight messages and key points, pushed to my Feishu each morning
✦ Watch alerts: periodically scans what I'm tracking and flags anything unusual
✦ Scheduled reports: generates a daily/weekly report on time and pushes it to the right group
✦ Group watch: checks progress in a group at intervals and pings me when conditions are met
👥 想给整个团队配一套共用的？👥 Want to set this up for a whole team?
上面讲的是给自己配。如果你想给公司团队部署一套共用的综合代理（VPN、账号、一键部署、连飞书），我们另有一本《企业部署指引》——团队版即将更新，敬请期待。 The above is for setting it up for yourself. To deploy a shared standing agent for a company team (VPN, accounts, one-click deploy, Feishu integration), there's a companion Enterprise Deployment Guide — the team edition is coming soon.
把每一态产生的东西记下来，越用越省——这不是单独一态，而是贯穿三态的习惯Write down what each mode produces; it pays off over time — not a separate mode, but a habit running through all three
问答里问出的好结论、协作里磨出的好模板、委托里跑顺的好流程——用完就忘，下次从零开始；记下来，就成了越滚越大的资产。这是最容易被忽略、却最有长期价值的习惯。落点很简单：一个能被 AI 直接读写的本地知识库。 A good conclusion from Ask, a good template from Co-create, a working flow from Delegate — forget them and you restart from zero; write them down and they become a snowballing asset. The most overlooked yet most valuable habit. The anchor is simple: a local knowledge base that AI can read and write directly.
数据完全属于你。所有笔记都是本地 Markdown 文件。更重要的是：Obsidian 的文件结构可以被 Claude Code 直接读取和操作——这是 AI + 知识管理的最佳组合。 Your data is fully yours. All notes are local Markdown files. Better yet: Obsidian's file structure can be directly read and manipulated by Claude Code — this is the ultimate AI + knowledge management combo.
✦ 日报系统：每天自动生成模板
✦ 知识 Wiki：AI 对话洞察整理成 Wiki 条目
✦ Agent 知识库：为不同 AI Agent 维护专属知识
✦ 写作素材库：收集灵感、引用、案例
✦ 所有通过 Claude Code 自动维护
✦ Daily Reports: Auto-generated templates every day
✦ Knowledge Wiki: AI conversation insights organized as wiki entries
✦ Agent Knowledge Base: Dedicated knowledge for each AI Agent
✦ Writing Library: Ideas, quotes, case studies
✦ All auto-maintained via Claude Code
"工具用完就忘 = 每次从零开始
用完就沉淀 = 持续复利增长"
"Use and forget = starting from zero every time
Use and capture = compound growth"
三态会变、工具会换，唯有沉淀随时间升值。 Modes shift and tools change; only capture appreciates over time.
| 场景Use Case | 海外工具Overseas | 国内平替CN Alternative | 平替度Parity |
|---|---|---|---|
| 日常对话Daily Chat | ChatGPT | 豆包 / Kimi / 通义千问 / DeepSeekDoubao / Kimi / Qwen / DeepSeek | ⭐⭐⭐⭐ |
| 深度分析Deep Analysis | Claude | Kimi / DeepSeek | ⭐⭐⭐ |
| AI 搜索AI Search | Perplexity | 秘塔 AI 搜索 / 天工Metaso AI / Tiangong | ⭐⭐⭐⭐ |
| AI 编程AI Coding | Cursor | Trae / MarsCode | ⭐⭐⭐ |
| 任务型委托Task Agent | Claude Code / Codex | Trae Agent / DeepSeek API | ⭐⭐⭐ |
| 常驻型综合代理Standing Agent | OpenClaw / Hermes | 扣子 Coze / Dify（自部署）Coze / Dify (self-hosted) | ⭐⭐⭐ |
| 知识库Notes | Obsidian | Obsidian（无需翻墙）/ 思源Obsidian (no VPN) / SiYuan | ⭐⭐⭐⭐⭐ |
纯国产零成本方案：Full China-only stack (¥0/month):
对话 → DeepSeek + 豆包（免费，两个换着用）
分析 → Kimi（上传长文档分析）
搜索 → 秘塔 AI 搜索
编程 → Trae（字节出品的 AI IDE）
工作流 → 扣子 Coze
笔记 → Obsidian（本身不需要翻墙）
Chat → DeepSeek + Doubao (free, alternate between both)
Analysis → Kimi (upload long documents)
Search → Metaso AI Search
Coding → Trae (ByteDance's AI IDE)
Workflow → Coze
Notes → Obsidian (no VPN needed)
起床前，常驻代理已经把昨夜的消息和今日要点聚合好，推到我飞书——我不用喊它，它按点自己干完了。Before I'm up, the standing agent has aggregated overnight messages and today's priorities to my Feishu — I didn't ask; it ran on schedule and finished.
③ 委托·常驻Delegate·Standing投研 loop 定时跑完，市场态势和持仓告警都更新好了，有异动推我飞书。我只看结论、拍板。The research loop has run on schedule — market state and position alerts refreshed, anomalies pushed to me. I just read the conclusions and decide.
③ 委托·任务Delegate·Task一份数据分析，我给方向和数据，它出 SQL、解读、可视化，来回几轮磨出结论。A data analysis — I give direction and data; it writes SQL, interprets, visualizes; a few rounds get us to a conclusion.
② 协作Co-create把"拉数据→处理→出周报"整条交给任务型 Agent，它自己跑完，我去忙别的。Hand the whole "pull data → process → weekly report" chain to a task agent; it runs to completion while I do something else.
③ 委托·任务Delegate·Task同一个问题问几个工具，交叉对比，防止单一来源带偏。Ask the same question across a few tools and cross-check, so no single source leads me astray.
① 问答Ask选题→生成→质检→多平台改写一条龙，它写完自检、留好我亲笔填的段落，发布我自己点。Topic→draft→QA→multi-platform rewrite end to end; it self-checks, leaves the passages I write by hand, and I click publish myself.
③ 委托·任务Delegate·Task自动生成当天工作摘要，把有价值的结论沉进 Obsidian 知识库——明天 AI 更懂我。Auto-generate today's work summary and capture the valuable conclusions into the Obsidian knowledge base — tomorrow the AI knows me better.
🧭 沉淀复利Capture| 工具Tool | 费用Cost | VPN | 难度Difficulty | 核心能力Core Strength |
|---|---|---|---|---|
| ① 问答 AskAsk | ||||
| 豆包Doubao | 免费Free | No | ⭐ | 中文对话Chinese chat |
| ChatGPT | 免费/$20月Free/$20mo | Yes | ⭐ | 全能对话All-purpose chat |
| ② 协作 Co-createCo-create | ||||
| Claude | 免费/$20月Free/$20mo | Yes | ⭐⭐ | 长文本深度分析Long-doc analysis |
| Cursor | 免费/$20月Free/$20mo | 部分Partial | ⭐⭐⭐ | AI 辅助编程AI coding |
| ③ 委托 DelegateDelegate | ||||
| Claude Code / Codex | $5~200+/mo | Yes | ⭐⭐⭐⭐ | 任务型 AgentTask agent |
| OpenClaw / HermesOpenClaw / Hermes | 自部署Self-host | 部分Partial | ⭐⭐⭐⭐ | 常驻型综合代理Standing agent |
| 🧭 贯穿 · 沉淀Throughline · Capture | ||||
| Obsidian | 免费Free | No | ⭐⭐ | 知识管理Knowledge mgmt |