● ai-maker-summit-talk — Cursor · main

⌘K · ⌘L · ⌘I

◆分享简介

开场问题

上下文工程

日常用法

总结

AI Maker Summit

大模型技术与 AI Coding

Cursor 的核心竞争力
Context Engineering
与 Coding Agent 新范式

Model is the engine · Context is the road

SpeakerMai Yang · 杨文

RoleCursor Ambassador

Format45 min · Q&A

mainUTF-8Markdown

OPENING HOOK多个第三方观察，指向同一个问题

先看两份外部证据

同一个模型，
不等于同一个工程能力

从早期网文观察，到 Composer 2.5 出来后的第三方 benchmark，大家开始测的不只是模型，而是完整 Coding Agent。

早期观察same model · different harness

Cursor Opus vs Claude Opus performance comparison

最新测评Artificial Analysis · Coding Agent Index

Artificial Analysis Coding Agent Index, time per task, and cost per task charts

今天的重点不是“谁第一”，而是：模型之外的上下文、工具、成本、时间和反馈链路，正在一起决定结果。

按 → 依次展示第三方证据 · then thesis

THESISModel is the engine · Context is the road

Main Claim

Model is the engine
Context is the road

决定 AI Coding 上限的，不只是模型——还有它如何看到问题、如何使用工具、如何被验证。

ENGINEModel

ROADContext → Harness → Workflow

01Context组织相关信息，压掉噪声。

02Harness（工作系统）模型外面的工作系统——把工具、提示、错误反馈组装成可迭代闭环。

03Workflow从写代码，转向定义目标与验收。

THESIS · engine and road

分享大纲 — Cursor

⌘K · ⌘L · ⌘I

◆分享大纲

123456

// 今天讲什么

分享大纲

▸talk/45 min · 30 slides

📄走进 Cursor文化 · Dogfooding · 上下文系统

📄Context Engineering索引 · 检索 · 上下文治理

📄日常用法经验 · 示例 · 沉淀

📄收尾与问答Takeaways · Q&A

main45 min · 4 parts

№01— 第一幕

INSIDE CURSOR

第一幕 · INSIDE CURSOR

走进 Cursor 的
工作方式

从 文化、Dogfooding 到上下文系统——看 Cursor 把这三层连成产品能力的方式。

▮▮▮ 第一幕FROM CULTURE → PRODUCT CAPABILITY

SOURCE · INSIDE CURSOR@zebriez · 190万+ views

@zebriez on X · Inside Cursor

深度观察 60 天

01Dogfooding（吃自己的狗粮）团队每天用 Cursor 做真实工作；产品能力从真实工作流里长出来，功能先被内部压力测试再 ship。

02Talent Density招聘像迭代产品；顶级 IC 密度本身，就是交付速度的一部分。

03Raise the ceiling先抬高上限，再降低门槛——优先满足高水平 builder 的能力诉求。

SOURCE · x.com/zebriez/status/1986510506199556395 · colossus.com/article/inside-cursor

CULTURE THREAD · 1/2从文化到工作方式

Steve Jobs at Apple — product culture reference

REFERENCE · Apple product culture

Culture becomes workflow

你怎么判断，
会变成团队默认

如何描述问题、选择证据、如何验收——强产品团队把这些写进日常协作。

BRIDGE · culture → dogfooding → product capability

CULTURE THREAD · 2/2文化即工作状态

Ben 把它做成网站

No-shoes.fun：a culture, not a policy

无拘无束、像在家一样、多元文化——No-shoes 说的是工作状态，不是 dress code。这种环境支撑敢试错，才撑得起 Dogfooding。

SOURCE · no-shoes.fun · Inside Cursor

01-culture/dogfooding-loop.md — cursor

⌘K · ⌘L · ⌘I

◆dogfooding-loop.md

12345

## 文化最后必须落到产品

Dogfooding 不是段子，是能力生产循环

自己高频使用

团队成员每天用 Cursor 完成真实工作。

发现真实摩擦

上下文不够、流程太慢、Agent 不会找信息。

内部先尝鲜

功能先进内部版本，被最挑剔的用户试用。

快速 Fuzz

大家一起找 bug、挑界面细节、找边界。

沉淀给用户

真正有效的能力，才进入公开产品。

maindogfooding

№02— 第二幕

WHAT CURSOR CAN ACTUALLY DO

第二幕 · REAL CAPABILITIES

Cursor 到底
帮我们做什么

Context OS、Model + Harness（工作系统）、AI-first Workflow：从具体能力看 AI Coding 工具如何帮助工程师完成真实工作。

▮▮▮ 第二幕INDEX · RETRIEVE · DISCOVER · ACT

THE PROBLEM一句话需求的困境

研发都懂

「一句话需求」
AI 也绕不开

典型一句

跟 xxx 一样，加个小功能。

接下来：Cursor 怎么做？

CONTEXT · 静态 → 动态行业共识 · 按需发现

两种组织上下文的方式

上下文不是一次性塞满窗口

静态上下文 · 预先塞进去

把规则、摘要、文件片段、工具描述都提前放进窗口，希望模型自己筛选。

简单直接，但 token 很容易膨胀。
无关信息互相干扰，造成 context rot。
摘要是有损压缩，关键细节可能被抹掉。

动态发现 · 按需拉回来

把长响应、历史、工具、终端、skills 变成可检索对象，由 agent 在工作中拉取。

只把必要信息带进上下文窗口。
文件路径提供可逆引用，不必一次性压缩。
适用于长任务、多工具、多 agent 协作。

BRIDGE · → 看 Cursor 怎么落地

CODEBASE INDEXING落地第一步

价值先说清楚

让 Agent 知道相关代码在哪里

Codebase Indexing 的核心价值：Agent 能定位「这个项目里，和问题相关的代码在哪」——不必靠泛化经验瞎猜。大仓库还要快，靠的是只同步变化的部分。

01Chunk把代码切成语义片段。

02Index为片段建立可检索表示。

03Retrieve按问题找到相关上下文。

CAPABILITY · understand the codebase

MERKLE TREEEngineer's Codex

大仓库也要快

不用每次重扫整个代码库

索引的秘密很简单：只同步变化的部分——改过的文件、改过的代码块，不必全量重扫。Merkle Tree 是 Cursor 用来做到这一点的机制。

Merkle tree explaining Cursor codebase indexing

SOURCE · read.engineerscodex.com/p/how-cursor-indexes-codebases-fast

POINT · incremental sync, not brute force

SCREENSHOTCodebase Indexing

从工程能力落到产品界面

项目被索引后，Agent 才能按项目语境工作

代码结构、相似实现和历史变更进入可检索环境，模型才不必靠泛化经验猜测项目意图。

VISUAL · show the real product

SCREENSHOTInstant Grep

第二层能力：快速找事实

好 Agent 会先查代码，再动手

不是一上来就改文件——而是在任务推进中，用搜索、读文件、对照文档，验证理解后再改。

CAPABILITY · retrieve facts quickly

02-capabilities/agent-in-action.md — Cursor · LIVE

⌘K · ⌘L · ⌘I

1234567

// 不要一次性堆全部上下文

好的 Agent 会边做边找

# 用户请求：给项目加 i18n
Agent: 先理解项目结构 → glob package.json
Agent: 找入口和路由 → read App.tsx / routes
Agent: 确认是否已有方案 → grep i18n|locale
Agent: 读取文档和约束 → @docs / web
Agent: context loaded → start editing

mainstep-by-step

CONTEXT OS上下文不是一坨文本，是分层环境

上下文分层

常驻的少而稳，按需的多而准

CodebaseIndex / PR history代码结构、历史变更、相似实现。

RulesProject / User团队习惯、架构约束、工作流模板。

SkillsDomain manuals只在相关任务中加载的专业手册。

ToolsSearch / Edit / Shell把行动转化成可观察结果。

TerminalLogs as files构建、测试、报错、运行输出。

MCPExternal context连接文档、需求单（Ticket）、设计、生产系统。

WorktreeIsolated execution并行任务与独立验证环境。

EvalKeep rate / satisfaction好不好要回到真实用户和代码库。

POINT · Cursor 把代码库、历史、规则、工具和外部系统变成 agent 可读环境

REAL SCREENSHOTyour actual context panel

抽象图之后，给真实证据

这就是一次真实会话的上下文账单

22% Full，约 61K / 272K tokens。Conversation 占 41.1K，Tools 占 10.5K。上下文窗口不是无限仓库，而是需要被管理的工作预算。

Cursor context window breakdown screenshot

REAL SCREENSHOT · Context Window Breakdown

POINT · context is budget

THREE PRINCIPLES这比「多塞上下文」更重要

Context Governance

把 Cursor 用好，本质是在做上下文治理

01静态上下文要少而稳Rules、架构约束、工作流模板要稳定、简短、可复用。

02动态上下文要可发现代码、历史、终端、文档、MCP 工具要能被 Agent 按需找到。

03工具结果要可追溯搜索、编辑、测试、日志和 PR 评论都要能回流到下一轮判断。

TAKEAWAY · 少而稳、可发现、可追溯

EVIDENCE修好路 · 体感会变

三个方向 · 一组公开数据

上下文工程，日常也能感知

首次 Query8s → 0.5s

中位仓库：团队索引复用后，新人打开项目不必从零等待。

Semantic Search+12.5%

语义搜索平均提升问答准确率。

MCP Context-46.9%

按需发现工具描述，相关运行中减少 token 占用。

EVIDENCE · index · retrieve · govern

№03— 第三幕

CONTEXT ENGINEERING IN PRACTICE

第三幕 · DAILY PRACTICE

日常怎么
用起来

前面讲清了 Cursor 如何组织上下文——这一幕回到 你在日常工作中如何运用这些原则。

▮▮▮ 第三幕REQUEST · SEARCH · ACT · VERIFY

WORKING LOOP好的上下文，是被工作过程不断生成的

先有一个工作框架

别命令每一步，定义目标、边界、验收

01 · REQUEST

提出目标

不是命令每一步，而是定义完成标准、范围、限制和不做什么。

02 · SEARCH

发现上下文

semantic search、grep、history、skills、terminal、MCP 都是上下文入口。

03 · ACT

编辑与执行

代码修改、工具调用、测试运行和日志观察形成可追踪轨迹。

04 · VERIFY

结果回流

失败测试、diff、PR 评论和用户反馈，都会变成下一轮上下文。

FRAME · Request → Search → Act → Verify

DAILY PRACTICE上下文习惯 · 协作习惯

回去就能用的 7 条

日常经验 · 7 条

01同一 chat 保持同一模型缓存按模型隔离，切换会 cache miss——更慢更贵，还可能丢细节。

02一个 chat 一个任务上下文预算有限；混任务会让无关信息互相干扰。

03复杂事先 Plan，再 Agent 执行先对齐范围和步骤，比边做边纠偏省很多轮。

04用 @ 精确引用@file / @folder / @docs 比「帮我看看项目」准得多。

05Prompt 里写验收标准「完成 = 测试通过 + 不影响现有 API」比「加个功能」有效。

06跑偏 → 新开 chat 重述约束长 thread 里反复纠错会污染上下文，越纠越偏。

07Rules 短而稳，Skills 从 workflow 里长静态上下文 3–5 条就够；重复流程做熟再沉淀成 skill。

PRACTICE · 7 habits

BEFORE / AFTER一个 Prompt 的差别

常见写法

一句话丢给 Agent

「跟 xxx 一样，加个小功能。」
Agent 只能猜范围、猜参考、猜验收标准。
容易改错文件、漏测边界、多轮纠偏。

更好写法

目标 + 约束 + 验收

在 src/features/ 下新增 Y，参考 @src/features/X/ 的实现模式。
约束：不改现有 API；测试框架用项目现有的 vitest。
完成标准：yarn test 通过；改动文件不超过 5 个。

EXAMPLE · 目标 + 约束 + 验收

EXAMPLE回扣 Working Loop

回到刚才的例子

好 Prompt，怎么跑完一圈

上一页的好写法，放进 Request → Search → Act → Verify。

01REQUEST · 目标 + 约束 + 验收新增 Y 功能，参考 X 模式；不改 API；yarn test 通过。

02SEARCH · Agent 先找参考@src/features/X/ → grep 相似实现 → read 关键文件，再动手。

03ACT · 对齐模式再编辑按现有结构改代码，控制 diff 范围——不是从零发明。

04VERIFY · 测试通过才算完Agent 跑 test、读 terminal；失败日志回流，进入下一轮判断。

EXAMPLE · loop in action

CONTEXT ASSETS回扣上下文治理

从一次协作到可复用资产

把经验沉淀成上下文

Rules 和 Skills 不是额外负担——是「静态上下文要少而稳」的落地方式。

01Rules 承载团队共识架构约束、测试框架、命名习惯——短、稳、可复用，Agent 每次都能读到。

02Skills 承载重复 workflowcommit 规范、code review、分层提交——做熟三次，再让 Agent 帮你整理成 skill。

ASSETS · rules · skills · tests

03-practice/agent-evolution.md — Cursor

⌘K · ⌘L · ⌘I

12345

// cursor.com/changelog

能力演进：从 Chat 到 Cloud Agent

①

2023Chat问答、补全

②

2024Compose多文件编辑

③

2024Agent工具调用、自主改代码

④

2025Plan先计划再执行

⑤

2025Multi-Agent并行 Agent · Sub-Agent 分工

★

2026Cloud本地交给云端继续跑

maincursor.com/changelog

TAKEAWAYS回到开场

今天带走三件事

Model is the engine
Context is the road

01不要只问模型强不强要问你的上下文、工具和验证系统是否足够好。

02把个人经验沉淀成系统资产规则、脚本、测试、skill 和检查清单，会让下一次协作更可靠。

03工程师正在成为 Builder定义目标、组织上下文、调度 Agent、验收结果——Context is the road。

TAKEAWAYS · engine and road

Cursor@Summit:~$ echo "context"

Context is the
new code.

上下文是路。在 AI 时代，它承接了原来只有「代码库」才承载的工程认知——结构、约束、历史与验收，都变成 Agent 可读、可按需发现的环境。

● MAIN ENDING

ONE MORE THING活动转场

如果你想继续面对面聊

One
More
Thing.

如果你也在探索 AI Coding、Agent 工作流和上下文治理，欢迎继续面对面交流。

TRANSITION · Cafe Cursor

CAFE CURSOR SHENZHEN2026.06.26

JOIN US · 一起聊聊

Cafe Cursor
Shenzhen

2026.06.26 · Friday

如果今天分享让你产生共鸣，欢迎 6/26 来深圳 Cafe Cursor，面对面继续聊 AI Coding、Agent 工作流和 Builder 的新能力。

luma.com/cursorcommunity

COMMUNITY CTA

Cursor@summit:~$ open questions.md

Q & A

Mai Yang · 杨文 · Cursor Ambassador · Shenzhen

● END OF TALK

● ai-maker-summit-talk — Cursor · main

⌘K · ⌘L · ⌘I

◆Overview

Opening hook

Context engineering

Daily practice

Wrap-up

AI Maker Summit

LLMs & AI Coding

Cursor's Core Edge
Context Engineering
& the Coding Agent Paradigm

Model is the engine · Context is the road

SpeakerMai Yang

RoleCursor Ambassador

Format45 min · Q&A

mainUTF-8Markdown

OPENING HOOKThird-party observations, same question

Two external signals first

Same model,
not the sameengineering outcome

From early blog takes to post–Composer 2.5 benchmarks—people measure the full Coding Agent, not the model alone.

Early signalsame model · different harness

Latest benchmarkArtificial Analysis · Coding Agent Index

Today's point isn't "who's #1"—context, tools, cost, time, and feedback loops together shape the result.

Press → reveal evidence · then thesis

THESISModel is the engine · Context is the road

Main Claim

Model is the engine
Context is the road

What caps AI Coding isn't only the model—it's how the system sees the problem, uses tools, and gets verified.

ENGINEModel

ROADContext → Harness → Workflow

01ContextOrganize what matters; cut the noise.

02HarnessThe work system around the model—tools, prompts, and error feedback in a tight loop.

03WorkflowFrom writing code to defining goals and acceptance criteria.

THESIS · engine and road

outline — Cursor

⌘K · ⌘L · ⌘I

◆outline.md

123456

// today's agenda

Talk outline

▸talk/45 min · 30 slides

📄Inside Cursorculture · dogfooding · context system

📄Context Engineeringindex · retrieve · govern

📄Daily practicehabits · examples · assets

📄Close & Q&Atakeaways · Q&A

main45 min · 4 parts

№01— Act I

INSIDE CURSOR

Act I · INSIDE CURSOR

How Cursor
actually works

From culture and dogfooding to the context system—how Cursor wires three layers into product capability.

▮▮▮ ACT IFROM CULTURE → PRODUCT CAPABILITY

SOURCE · INSIDE CURSOR@zebriez · 1.9M+ views

@zebriez on X · Inside Cursor

60-day deep dive

01DogfoodingThe team uses Cursor for real work daily; capabilities grow from real workflows and get pressure-tested before they ship.

02Talent DensityHiring is iterated like the product; top IC density is part of delivery speed.

03Raise the ceilingRaise the ceiling first, then lower the floor—optimize for high-skill builders first.

SOURCE · x.com/zebriez/status/1986510506199556395 · colossus.com/article/inside-cursor

CULTURE THREAD · 1/2From culture to workflow

REFERENCE · Apple product culture

Culture becomes workflow

How you judge
becomes the team default

How you frame problems, pick evidence, and define done—strong product teams bake this into daily collaboration.

BRIDGE · culture → dogfooding → product capability

CULTURE THREAD · 2/2Culture as working mode

Ben made it a website

No-shoes.fun：a culture, not a policy

Feeling at home, multicultural openness, working without artificial constraints—not playfulness as policy. That environment enables experimentation—and dogfooding.

SOURCE · no-shoes.fun · Inside Cursor

01-culture/dogfooding-loop.md — cursor

⌘K · ⌘L · ⌘I

◆dogfooding-loop.md

12345

## culture must land in the product

Dogfooding is a capability loop, not a slogan

Use it daily

Everyone ships real work with Cursor.

Find real friction

Missing context, slow flows, agents that can't find information.

Ship internally first

Features hit internal builds and the most demanding users first.

Fuzz fast

Bugs, UI edges, and boundary cases get hammered together.

Ship what works

Only durable capabilities reach the public product.

maindogfooding

№02— Act II

WHAT CURSOR CAN ACTUALLY DO

Act II · REAL CAPABILITIES

What Cursor
does for you

Context OS, model + harness, AI-first workflow—concrete capabilities for real engineering work.

▮▮▮ ACT IIINDEX · RETRIEVE · DISCOVER · ACT

THE PROBLEMThe one-liner trap

Every engineer knows this

The "one-liner"
request

Typical line

Like xxx, add a small feature.

Next · how Cursor responds

CONTEXT · static → dynamicIndustry consensus · discover on demand

Two ways to organize context

Context isn't a one-shot window fill

Static context · preload everything

Rules, summaries, file chunks, and tool descriptions stuffed into the window up front—hope the model filters well.

Simple, but tokens balloon fast.
Irrelevant info interferes—context rot.
Summaries are lossy; details disappear.

Dynamic discovery · pull when needed

Long responses, history, tools, terminal, and skills become retrievable objects the agent pulls during work.

Only necessary info enters the window.
File paths are reversible references—no one-shot compression.
Fits long tasks, many tools, multi-agent work.

BRIDGE · → Cursor's implementation

CODEBASE INDEXINGFirst implementation step

Value first

Help the agent find relevant code

Core value—locate "where in this project does code related to my question live" without guessing from general priors. Large repos stay fast by syncing only what changed.

01ChunkSplit code into semantic segments.

02IndexBuild retrievable representations.

03RetrievePull relevant context per question.

CAPABILITY · understand the codebase

MERKLE TREEEngineer's Codex

Large repos, still fast

No full-repo rescan every time

The trick is incremental sync—changed files and chunks only. Merkle trees are how Cursor does it.

SOURCE · read.engineerscodex.com/p/how-cursor-indexes-codebases-fast

POINT · incremental sync, not brute force

SCREENSHOTCodebase Indexing

From capability to product UI

Once indexed, the agent works in project context

Structure, similar implementations, and history enter a searchable environment—the model doesn't have to guess intent.

VISUAL · show the real product

SCREENSHOTInstant Grep

Second layer · find facts fast

Good agents search first, then edit

Not "open files and hope"—search, read, check docs, then change with verified understanding.

CAPABILITY · retrieve facts quickly

02-capabilities/agent-in-action.md — Cursor · LIVE

⌘K · ⌘L · ⌘I

1234567

// don't dump the whole context at once

Strong agents discover as they go

# user request: add i18n to the project
Agent: understand structure → glob package.json
Agent: find entry & routes → read App.tsx / routes
Agent: check existing approach → grep i18n|locale
Agent: read docs & constraints → @docs / web
Agent: context loaded → start editing

mainstep-by-step

CONTEXT OScontext is a layered environment

Context layers

Small stable core, precise on demand

CodebaseIndex / PR historyStructure, history, similar implementations.

RulesProject / UserTeam habits, architecture constraints, workflow templates.

SkillsDomain manualsSpecialist playbooks loaded only for relevant tasks.

ToolsSearch / Edit / ShellActions become observable outcomes.

TerminalLogs as filesBuild, test, errors, runtime output.

MCPExternal contextDocs, tickets, design, production systems.

WorktreeIsolated executionParallel tasks and isolated verification.

EvalKeep rate / satisfactionQuality ties back to real users and repos.

POINT · codebase, history, rules, tools, external systems

REAL SCREENSHOTyour actual context panel

After the diagram, real evidence

A real session's context bill

22% full, ~61K / 272K tokens. Conversation 41.1K, Tools 10.5K. The window isn't infinite storage—it's a budget to manage.

REAL SCREENSHOT · Context Window Breakdown

POINT · context is budget

THREE PRINCIPLESMore than "stuff more context"

Context Governance

Using Cursor well is context governance

01Static context · small and stableRules, architecture constraints, workflow templates—short, stable, reusable.

02Dynamic context · discoverableCode, history, terminal, docs, MCP—found on demand by the agent.

03Tool results · traceableSearch, edits, tests, logs, PR comments feed the next decision.

TAKEAWAY · small & stable · discoverable · traceable

EVIDENCEFix the road · feel the difference

Three directions · public data

Context engineering you can feel day to day

First query8s → 0.5s

Median repo—team index reuse; new contributors don't wait from zero.

Semantic Search+12.5%

Average lift in answer accuracy.

MCP Context-46.9%

On-demand tool descriptions cut token use in relevant runs.

EVIDENCE · index · retrieve · govern

№03— Act III

CONTEXT ENGINEERING IN PRACTICE

Act III · DAILY PRACTICE

How to use it
every day

We've seen how Cursor organizes context—this act is how you apply those principles at work.

▮▮▮ ACT IIIREQUEST · SEARCH · ACT · VERIFY

WORKING LOOPgood context is generated by the work

A working frame first

Don't micromanage steps—define goals, bounds, acceptance

01 · REQUEST

State the goal

Not every micro-step—completion criteria, scope, limits, and explicit non-goals.

02 · SEARCH

Discover context

Semantic search, grep, history, skills, terminal, MCP—all entry points.

03 · ACT

Edit & execute

Edits, tool calls, tests, and logs form a traceable trail.

04 · VERIFY

Feed results back

Failed tests, diffs, PR comments, and user feedback become next-turn context.

FRAME · Request → Search → Act → Verify

DAILY PRACTICEcontext habits · collaboration habits

seven habits to try tomorrow

Daily practice · 7 habits

01One model per chatCache is per-model—switching causes cache miss, slower costlier runs, lost detail.

02One task per chatContext budget is finite; mixed tasks create interference.

03Plan complex work, then agentAlign scope and steps first—cheaper than correcting mid-flight.

04Use @ references precisely@file / @folder / @docs" beats "look at the project.

05Put acceptance in the promptDone = tests pass + no API break" beats "add a feature.

06Drift → new chat, restate constraintsLong threads of corrections pollute context.

07Short stable rules; skills from workflows3–5 static rules; turn repeated flows into skills once they're muscle memory.

PRACTICE · 7 habits

BEFORE / AFTEROne prompt, two outcomes

Common

One line to the agent

Like xxx, add a small feature.
Agent guesses scope, reference, and done-ness.
Wrong files, missed tests, many correction rounds.

Better

Goal + constraints + acceptance

Add Y under src/features/, mirror @src/features/X/ patterns.
[object Object]
[object Object]

EXAMPLE · goal + constraints + acceptance

EXAMPLEWorking loop recap

Back to the example

A good prompt, one full loop

The "better" prompt from the last slide, through Request → Search → Act → Verify.

01REQUEST · goal + constraints + acceptanceAdd feature Y like X; no API change; yarn test passes.

02SEARCH · agent finds references@src/features/X/ → grep similar → read key files, then edit.

03ACT · match patterns, then editFollow existing structure; keep diff bounded—not invent from scratch.

04VERIFY · tests define doneAgent runs tests, reads terminal; failures feed the next turn.

EXAMPLE · loop in action

CONTEXT ASSETSback to governance

from one session to reusable assets

Turn experience into context

Rules and skills aren't overhead—they're how "static context stays small and stable."

01Rules hold team consensusArchitecture, test stack, naming—short, stable, always readable by the agent.

02Skills hold repeat workflowsCommit style, review, layered commits—after three runs, let the agent help codify a skill.

ASSETS · rules · skills · tests

03-practice/agent-evolution.md — Cursor

⌘K · ⌘L · ⌘I

12345

// cursor.com/changelog

From Chat to Cloud Agent

①

2023ChatQ&A, completion

②

2024ComposeMulti-file edits

③

2024AgentTools, autonomous edits

④

2025PlanPlan before execute

⑤

2025Multi-AgentParallel agents · sub-agent split

★

2026CloudHand off local work to the cloud

maincursor.com/changelog

TAKEAWAYSback to the opening

three things to take home

Model is the engine
Context is the road

01Don't only ask "how strong is the model?"Ask whether context, tools, and verification are good enough.

02Turn personal experience into system assetsRules, scripts, tests, skills, and checklists make the next session more reliable.

03Engineers are becoming buildersDefine goals, organize context, orchestrate agents, verify outcomes—Context is the road.

TAKEAWAYS · engine and road

Cursor@Summit:~$ echo "context"

Context is the
new code.

Context is the road. In the AI era it carries what only a codebase used to—structure, constraints, history, and acceptance—all in an environment agents can read and discover on demand.

● MAIN ENDING

ONE MORE THINGCommunity transition

If you want to keep talking in person

One
More
Thing.

If you're exploring AI coding, agent workflows, and context governance—let's continue face to face.

TRANSITION · Cafe Cursor

CAFE CURSOR SHENZHEN2026.06.26

JOIN US

Cafe Cursor
Shenzhen

2026.06.26 · Friday

If today's talk resonated, join us June 26 in Shenzhen for AI coding, agent workflows, and the builder skill set—in person.

luma.com/cursorcommunity

COMMUNITY CTA

Cursor@summit:~$ open questions.md

Q & A

Mai Yang · Cursor Ambassador · Shenzhen

● END OF TALK

Cursor 的核心竞争力 Context Engineering 与 Coding Agent 新范式

同一个模型， 不等于同一个工程能力

Model is the engine Context is the road

分享大纲

走进 Cursor 的工作方式

你怎么判断， 会变成团队默认

No-shoes.fun：a culture, not a policy

Dogfooding 不是段子，是 能力生产循环

自己高频使用

发现真实摩擦

内部先尝鲜

快速 Fuzz

沉淀给用户

Cursor 到底帮我们做什么

「一句话需求」 AI 也绕不开

上下文不是一次性塞满窗口

静态上下文 · 预先塞进去

动态发现 · 按需拉回来

让 Agent 知道相关代码在哪里

不用每次重扫整个代码库

项目被索引后，Agent 才能按项目语境工作

好 Agent 会先查代码，再动手

好的 Agent 会边做边找

常驻的少而稳，按需的多而准

这就是一次真实会话的上下文账单

把 Cursor 用好，本质是在做上下文治理

上下文工程，日常也能感知

日常怎么用起来

别命令每一步，定义目标、边界、验收

提出目标

发现上下文

编辑与执行

结果回流

日常经验 · 7 条

一句话丢给 Agent

目标 + 约束 + 验收

好 Prompt，怎么跑完一圈

把经验沉淀成上下文

能力演进：从 Chat 到 Cloud Agent

Model is the engine Context is the road

Context is thenew code.

Cafe CursorShenzhen

Q & A

Cursor's Core Edge Context Engineering & the Coding Agent Paradigm

Same model, not the sameengineering outcome

Model is the engine Context is the road

Talk outline

How Cursoractually works

How you judge becomes the team default

No-shoes.fun：a culture, not a policy

Dogfooding is a capability loop, not a slogan

Use it daily

Find real friction

Ship internally first

Fuzz fast

Ship what works

What Cursordoes for you

The "one-liner" request

Context isn't a one-shot window fill

Static context · preload everything

Dynamic discovery · pull when needed

Help the agent find relevant code

No full-repo rescan every time

Once indexed, the agent works in project context

Good agents search first, then edit

Strong agents discover as they go

Small stable core, precise on demand

A real session's context bill

Using Cursor well is context governance

Context engineering you can feel day to day

How to use itevery day

Don't micromanage steps—define goals, bounds, acceptance

State the goal

Discover context

Edit & execute

Feed results back

Daily practice · 7 habits

One line to the agent

Goal + constraints + acceptance

A good prompt, one full loop

Cursor 的核心竞争力
Context Engineering
与 Coding Agent 新范式

同一个模型，
不等于同一个工程能力

Model is the engine
Context is the road

走进 Cursor 的
工作方式

你怎么判断，
会变成团队默认

Dogfooding 不是段子，是能力生产循环

Cursor 到底
帮我们做什么

「一句话需求」
AI 也绕不开

日常怎么
用起来

Model is the engine
Context is the road

Context is the
new code.

Cafe Cursor
Shenzhen

Cursor's Core Edge
Context Engineering
& the Coding Agent Paradigm

Same model,
not the sameengineering outcome

Model is the engine
Context is the road

How Cursor
actually works

How you judge
becomes the team default

What Cursor
does for you

The "one-liner"
request

How to use it
every day

Model is the engine
Context is the road

Context is the
new code.

Cafe Cursor
Shenzhen