干别的 · 调研、写作、数据都在它里面Do Other Work · Research, Writing, and Data Inside It.
把前面几章拼起来，会得出一个官网不会明说的结论：agent + 终端 + 上下文 + MCP，凑成的不是编辑器，是一台什么活都能接的工作台。一个装满资料的文件夹当项目，它能替你调研出一份带引用的笔记、清洗合并一堆 CSV、批量重命名几百个文件、就着 Canvas 写长文。这一章给三个非编码的真实场景，也诚实说清：哪一步它不如一个专用工具。Stack the earlier chapters together and you reach a conclusion the site won't spell out: agent + terminal + context + MCP add up not to an editor but to a workbench that takes on almost any task. Point it at a folder of material as a 'project' and it can research a cited note, clean and merge a pile of CSVs, batch-rename hundreds of files, draft long-form in Canvas. Three real non-code scenarios — and an honest account of the step where a dedicated tool still wins.
把前面几章拼起来，会得出一个官网不会明说的结论：agent + 终端 + 上下文 + MCP，凑成的不是编辑器，是一台什么活都能接的工作台。Stack the earlier chapters together and you reach a conclusion the site won't spell out: agent + terminal + context + MCP add up not to an editor but to a workbench that takes on almost any task.一个装满资料的文件夹当项目，它能替你调研出一份带引用的笔记、清洗合并一堆 CSV、批量改几百个文件、就着 markdown 写长文。这一章给三个非编码的真实场景，也诚实说清：哪一步它不如一个专用工具。这是整本书的落点 —— 如果元工具论点只在一处兑现，就是这里。Point it at a folder of material as a 'project' and it can research a cited note, clean and merge a pile of CSVs, batch-edit hundreds of files, draft long-form in markdown. This chapter gives three real non-code scenarios — and an honest account of where a dedicated tool wins. This is where the whole book lands — if the meta-tool claim pays off anywhere, it's here.
非代码任务Non-code task
留在 Cursor 的信号Keep in Cursor when
切走的信号Switch away when
技术调研 / 写笔记Technical research / notes
资料能文件化、来源可追溯、输出是 markdownMaterial can be file-based, sources traceable, output is markdown
需要对抗式多源查证和深度网页研究Requires adversarial multi-source verification and deep web research
写作项目Writing project
有资料、风格指南、版本控制，重点是结构和改写Has material, style guide, version control; focus is structure and revision
需要心流、精细排版、协作文档批注Needs flow, fine layout, collaborative document comments
数据清洗Data cleaning
规则清楚、可 dry-run、输出可 diffRules are clear, dry-run is possible, output can be diffed
需要产品管理系统或知识库权限流Needs product-management or knowledge-base permission flows
— I
调研：搜、取、写成带引用的笔记Research: Search, Fetch, Write a Cited Note.
它能查、能取、能写，三步在一个窗口里闭合。It can search, fetch, and write — three steps that close in one window.用 agent 的网页搜索拉来源、用 MCP 取你自己系统里的数据、用 @Docs 把一份文档站索引进来，1注 1Note 1Cursor Docs · Agent —— agent 的工具里含网页搜索、图像生成、终端执行、跨文件编辑、浏览器控制、MCP，这些不绑定「写代码」，构成它做非编码活的基础。截至 2026-07-10。Cursor Docs · Agent — the agent's tools include web search, image generation, terminal execution, cross-file editing, browser control, and MCP — none tied to 'writing code,' and together they're the basis for its non-code work. As of 2026-07-10.然后让它边读边把结论写进一份 markdown 笔记、每条带出处，资料就在你这个文件夹里、随时可复查。2注 2Note 2Cursor Docs · Codebase Indexing —— 打开任何项目（文件夹）即自动建索引、可按语义检索，非代码资料同样适用；@Docs 可把外部文档站索引进来。截至 2026-07-10。Cursor Docs · Codebase Indexing — any project (folder) you open is indexed automatically and searchable by meaning, non-code material included; @Docs can pull in an external docs site. As of 2026-07-10.它和聊天框查资料的区别在于：结论落进了文件，不是停在一段对话里、关掉就没了。同一套机制也让它当「第二大脑」：把多年的笔记建成索引，按意思找回那条你早忘了文件名的记录。Pull sources with the agent's web search, fetch data from your own systems via MCP, index a docs site with @Docs,1注 1Note 1Cursor Docs · Agent —— agent 的工具里含网页搜索、图像生成、终端执行、跨文件编辑、浏览器控制、MCP，这些不绑定「写代码」，构成它做非编码活的基础。截至 2026-07-10。Cursor Docs · Agent — the agent's tools include web search, image generation, terminal execution, cross-file editing, browser control, and MCP — none tied to 'writing code,' and together they're the basis for its non-code work. As of 2026-07-10. then have it write conclusions into a markdown note as it reads, each line sourced, the material living in your folder and re-checkable anytime.2注 2Note 2Cursor Docs · Codebase Indexing —— 打开任何项目（文件夹）即自动建索引、可按语义检索，非代码资料同样适用；@Docs 可把外部文档站索引进来。截至 2026-07-10。Cursor Docs · Codebase Indexing — any project (folder) you open is indexed automatically and searchable by meaning, non-code material included; @Docs can pull in an external docs site. As of 2026-07-10. The difference from looking things up in a chat box: the conclusions land in a file, not in a conversation that vanishes when you close it. The same mechanism makes it a 'second brain' too: index years of notes and recall by meaning the entry whose filename you long forgot.诚实的边界：要十几个来源交叉核对、对抗式查证的深度报告，一个专门的 deep-research 工具走得更远。Cursor 的甜区是「就着你已有的资料 + 几次搜索」产出可用、可复查的笔记，不是替代严肃的多源调研。要它出报告，先把来源喂准，别指望它替你决定该信谁。The honest boundary: for a deep report that cross-checks a dozen sources and verifies adversarially, a dedicated deep-research tool goes further. Cursor's sweet spot is 'your existing material plus a few searches' into a usable, re-checkable note, not a replacement for serious multi-source research. To have it produce a report, feed the sources precisely — don't expect it to decide whom to trust for you.
提示词PromptCursor Agent
就<主题>查证并记笔记：
- 先用网页搜索找 3–5 个一手来源
- 每条结论后用方括号标出处链接
- 拿不准的标「待核」，别编
写进 notes/<主题>.md。Research and note <topic>:
- find 3–5 primary sources via web search first
- tag each conclusion with its source link in brackets
- mark anything uncertain as 'unverified'; don't fabricate
Write it to notes/<topic>.md.
就「欧盟 AI 法案对开源模型的豁免」查证并记笔记：
- 先用网页搜索找 3–5 个一手来源（法案原文、官方问答优先）
- 每条结论后用方括号标出处链接
- 拿不准的标「待核」，别编
写进 notes/eu-ai-act-open-source.md。Research and note 'the EU AI Act's exemptions for open-source models':
- find 3–5 primary sources via web search first (the Act's text and official Q&As preferred)
- tag each conclusion with its source link in brackets
- mark anything uncertain as 'unverified'; don't fabricate
Write it to notes/eu-ai-act-open-source.md.
— II
数据与文件：在终端里清、合、批处理Data and Files: Clean, Merge, Batch in the Terminal.
你描述要什么结果，它写脚本、跑脚本、读输出再修。You describe the outcome; it writes the script, runs it, reads the output, and fixes.合并去重一堆 CSV、按规则批量重命名或转换几百个文件、从一批文档里抽字段做成一张表 —— 这些都在集成终端里用 python / pandas / shell 跑，你不必会这些命令。第 4 章讲过这条入口；这里是它最省事的用法：你说「把这二十个表按日期列合并、去掉重复行、空值填 0」，它写好脚本、跑给你看、有问题接着改，你审最后那张表就行。Merge and dedupe a pile of CSVs, batch-rename or convert hundreds of files by a rule, extract fields from a batch of documents into one table — all run in the integrated terminal with python / pandas / shell, no commands needed from you. Chapter 4 covered this on-ramp; here's its most labor-saving use: you say 'merge these twenty sheets on the date column, drop duplicate rows, fill blanks with 0,' it writes the script, runs it, fixes what's off, and you just review the final table.诚实的边界：要反复看图、调参的交互式数据探索，一个 notebook 更直接 —— Cursor 适合「一次性、规则清楚」的批处理，不适合「边看边想下一步」的探索。还有一条：它跑的是脚本，处理得了结构化数据和文件，处理不了「只能在某个 app 的界面里点」的数据 —— 那又回到 Computer Use 那条线。The honest boundary: for interactive data exploration with repeated plotting and tuning, a notebook is more direct — Cursor suits 'one-off, clearly-ruled' batch work, not 'look-then-decide-next-step' exploration. And one more: it runs scripts, handling structured data and files, but not data that 'can only be clicked through some app's interface' — which returns to the Computer Use line.
提示词PromptCursor Agent + Terminal
数据清洗 dry-run：
目标：<清洗后要得到什么>
输入：<raw files>
输出：先写到 data/working，不覆盖 raw
步骤：
1. 先列文件和字段概览
2. 写脚本但先 dry-run，输出将改动的行数 / 样例
3. 我确认后再生成 data/output/<file>
4. 输出校验：行数、空值、重复值、字段类型
5. 最后给回滚办法。Data cleaning dry-run:
Goal: <what cleaned output should be>
Input: <raw files>
Output: write to data/working first; do not overwrite raw
Steps:
1. List files and field overview first
2. Write the script but dry-run, showing changed row counts / samples
3. After confirmation, generate data/output/<file>
4. Validate row count, nulls, duplicates, field types
5. Finish with rollback instructions.
数据清洗 dry-run：
目标：把 12 个月的销售表合并成一张，按订单号去重，金额列空值填 0
输入：data/raw/sales-2025-*.csv
输出：先写到 data/working，不覆盖 raw
步骤：
1. 先列文件和字段概览
2. 写脚本但先 dry-run，输出将改动的行数 / 样例
3. 我确认后再生成 data/output/sales-2025-merged.csv
4. 输出校验：行数、空值、重复值、字段类型
5. 最后给回滚办法。Data cleaning dry-run:
Goal: merge 12 monthly sales sheets into one, dedupe by order ID, fill blanks in the amount column with 0
Input: data/raw/sales-2025-*.csv
Output: write to data/working first; do not overwrite raw
Steps:
1. List files and field overview first
2. Write the script but dry-run, showing changed row counts / samples
3. After confirmation, generate data/output/sales-2025-merged.csv
4. Validate row count, nulls, duplicates, field types
5. Finish with rollback instructions.
— III
写作：在它里面写，但知道何时切走Writing: Write in It, but Know When to Leave.
它适合「就着资料结构化地写」，不适合「需要心流的纯散文」。It suits 'structured writing alongside material,' not 'flowing prose that needs a writing trance.'一套真实可跑的写作流：文稿用 git 管的 markdown，把你的风格指南和几篇旧文用 @Files 喂进去定调，初稿让 agent 起、你用 Cmd+K 逐段收（「这段压缩 15%、换成主动语态、一句一个意思」），写完 git push 发布；Canvas 给你一个画布式的排布和分享面。3注 3Note 3Cursor Docs · Canvases / Design Mode —— Canvas 让 agent 生成可交互的 artifact（仪表盘、分析、报告）在聊天旁渲染、可复开可迭代；Design Mode 在浏览器里点选元素、画标注、语音描述来改 UI。截至 2026-07-10。Cursor Docs · Canvases / Design Mode — Canvas has the agent produce interactive artifacts (dashboards, analyses, reports) rendered beside the chat, reopenable and iterable; Design Mode edits UI from the browser by clicking elements, drawing, or describing changes by voice. As of 2026-07-10.它强在「资料、索引、改写都在同一处」—— 你不必在浏览器、文档、笔记之间来回搬。A real, runnable writing flow: keep the draft as git-managed markdown, feed your style guide and a few past pieces via @Files to set the voice, have the agent draft, tighten paragraph by paragraph with Cmd+K ('cut this 15%, active voice, one idea per sentence'), and git push to publish; Canvas gives you a canvas-style layout and sharing surface.3注 3Note 3Cursor Docs · Canvases / Design Mode —— Canvas 让 agent 生成可交互的 artifact（仪表盘、分析、报告）在聊天旁渲染、可复开可迭代；Design Mode 在浏览器里点选元素、画标注、语音描述来改 UI。截至 2026-07-10。Cursor Docs · Canvases / Design Mode — Canvas has the agent produce interactive artifacts (dashboards, analyses, reports) rendered beside the chat, reopenable and iterable; Design Mode edits UI from the browser by clicking elements, drawing, or describing changes by voice. As of 2026-07-10. Its strength is that material, index, and rewriting sit in one place — you stop shuttling between browser, doc, and notes.
提示词PromptCmd+K
就选中这段，按我的风格改：
- 压到原长度的 85%
- 主动语态，一句一个意思
- 删填充词和套话
- 术语和事实别动
只给改后的文本。Edit just the selected passage in my style:
- cut to 85% of the length
- active voice, one idea per sentence
- drop filler and clichés
- leave terms and facts unchanged
Return only the revised text.
就选中这段产品发布说明，按我的风格改：
- 压到原长度的 85%
- 主动语态，一句一个意思
- 删「赋能」「打造闭环」这类套话
- 版本号和日期别动
只给改后的文本。Edit just the selected product release note in my style:
- cut to 85% of the length
- active voice, one idea per sentence
- drop buzzwords like 'empower' and 'leverage synergies'
- leave version numbers and dates unchanged
Return only the revised text.
带引用的学术 / 长报告写作也吃这套：它长于改写、重排、套期刊或基金的格式，配合 Zotero、Pandoc 把引用和导出接上。一个真实缺口要先说 —— 它对 PDF 的原生支持有限，源文献多是 PDF 时，得先转成文本或 markdown 再 @Docs 拉进来，这也是社区最常提的不足。Cited academic / long-report writing rides the same setup: it's strong at rewriting, reordering, and conforming to a journal's or grant's format, paired with Zotero and Pandoc for citations and export. One real gap to state up front — its native PDF support is limited, so when your sources are mostly PDFs you convert them to text or markdown first and pull them in with @Docs; this is the shortfall the community names most.诚实的边界：要心流的纯写作，一个专门的写作软件更顺 —— Cursor 是编辑器底子，长在「改」不长在「流」。而且最关键的一条：它写完不能替你「点」发布到你的 CMS、不能替你登录后台贴上去 —— 那是 Computer Use（驱动你真实的 app），是另一类工具（Claude in Chrome / Operator）的事，下一章那条线。4注 4Note 4Cursor Blog · Agents can now control their own computers（2026-02-24）—— Cursor 的 computer use 限于云端 agent 的沙箱，划出「项目内 / 沙箱」与「驱动你真实 app」之间那条边界。Cursor Blog · Agents can now control their own computers (2026-02-24) — Cursor's computer use is confined to the cloud agent's sandbox, marking the boundary between 'inside a project / sandbox' and 'driving your real apps.'它能把资料写成稿，替你点「发布」不归它管。The honest boundary: for flowing prose that needs a trance, a dedicated writing app is smoother — Cursor is editor-grade, strong at revising, not at flow. And the crucial one: once written, it can't 'click' publish to your CMS or log into your backend and paste it in — that's Computer Use (driving your real apps), a different class of tool (Claude in Chrome / Operator), the line in the next chapter.4注 4Note 4Cursor Blog · Agents can now control their own computers（2026-02-24）—— Cursor 的 computer use 限于云端 agent 的沙箱，划出「项目内 / 沙箱」与「驱动你真实 app」之间那条边界。Cursor Blog · Agents can now control their own computers (2026-02-24) — Cursor's computer use is confined to the cloud agent's sandbox, marking the boundary between 'inside a project / sandbox' and 'driving your real apps.' It turns material into a draft; clicking 'publish' for you isn't its job.
专用 Research 工具 + 人工核查Dedicated research tool + human verification
数据大到不能用文件夹脚本舒服处理Data is too large for folder scripts
Notebook / database / BI
要精细版式、评论修订、多人实时协作Needs fine layout, comments, realtime collaboration
文档 / 设计 / 排版工具Docs / design / publishing tools
长期 shell 流程、跨项目命令、系统层操作Long-lived shell workflows, cross-project commands, system ops
Warp / cron / scripts
这一题是全书的试金石 —— 挑一件你平时绝不会想到用编辑器做的事：This exercise is the book's litmus test — pick something you'd never think to use an editor for:
01
选一件非编码的真实活Pick one real non-code task
整理一批文件、清洗合并一堆 CSV、或调研一个主题并产出带引用的笔记 —— 选你这周真要做的那件。Organize a batch of files, clean and merge a pile of CSVs, or research a topic into a cited note — pick the one you actually need to do this week.
02
全程在 Cursor 的 agent + 终端 + MCP 里完成，不开别的 appDo all of it in Cursor's agent + terminal + MCP, no other app
把资料文件夹当项目打开，描述结果，让它搜、取、跑、写。Open the material folder as a project, describe the outcome, and let it search, fetch, run, and write.
03
记下三件事：哪些顺手、哪一步你不得不切走、为什么Log three things: what was smooth, the step you had to switch for, and why
那「不得不切走」的一步，就是 Cursor 这台元工具的真实边界 —— 也是你下一章要对照的那条线。That 'had to switch' step is the real edge of Cursor as a meta-tool — and the line you'll measure against in the next chapter.
它能干的远不止代码，
但不是替你点别的 app.
It does far more than code,
but not click your other apps.
Aklman Library
— 讨论Discussion
讨论Discussion.
评论区初始化中…Initializing comments…
01 / 01
没有匹配结果No matches.
换个关键词，或按 Esc 回到页面Try another keyword, or press Esc to return