PTC-101 2026 秋 · 现代构建者 Fall 2026 · Modern AI Builder

PTC-101 2026 秋 · 核心工程课 · 对标 Stanford CS146S (Fall 2026) Fall 2026 · Core Engineering Course · Inspired by Stanford CS146S

现代构建者：与智能体一起创造 The Modern AI Builder: Agentic Software Engineering

掌握 System Prompt、Context Engineering（上下文工程）、MCP、Agent Skills 与多智能体协作，是为了让你拥有与智能体顺畅配合、独立把想法构建出来（Independent Building）的能力；而在这之上，我们更看重真正决定上限的上层能力——软件工程的架构品味、设计艺术，以及「我要为谁创造什么真实价值（Value Definition）」的独立思考。 Mastering System Prompts, Context Engineering, MCP, Agent Skills, and Multi-Agent Collaboration equips you to work fluently with coding agents and achieve Independent Building. Above these operational interfaces lies the true upper ceiling we cultivate: software engineering taste, design art, and independent thinking around Value Definition.

核心主线： Core Pillars: System Prompt Context Engineering MCP Agent Skills AGENTS.md & Subagents Architecture Taste
课程周期：Duration: 10 周（四大进阶篇章） 10 Weeks (4 Progressive Phases) 单次节奏：Session: 70 分钟（20 + 25 + 25 分层工作坊） 70 mins (20 + 25 + 25 Tiered Workshop) 先修要求：Prerequisites: 零手搓代码门槛 · 借助 AI 探索与构建 Beginner-friendly · Build alongside AI

课程设计哲学 · 双层能力模型 Course Philosophy · Two-Layer Capability Model

分清「与智能体协作的工程接口」与「决定上限的设计哲学」 Distinguishing Agent Collaboration Interfaces from Upper-Ceiling Design Philosophy

很多教程把提示词模板或配置文件奉为终极秘籍；但在青禾·元野，我们把能力清晰划分为上下两层：底层接口帮你跨过「独立构建」的门槛，上层品味与思考决定你能创造多大的价值。 Many tutorials treat prompt templates or config files as silver bullets. At Proto Commons, we separate capabilities into two distinct layers: operational interfaces unlock Independent Building, while engineering taste and design philosophy define your true ceiling.

Layer 01 · 驾驭接口 → 解锁「独立构建」 Layer 01 · Agent Interfaces → Unlock Independent Building

与智能体顺畅协作的工程接口层 The Operational Engineering Interface Layer

哪怕你现在的纯手写代码能力偏弱，只要系统掌握 System Prompt（系统提示词）、Context Engineering（上下文工程）、MCP（模型上下文协议）、Agent Skills（智能体技能）以及 AGENTS.md 与多智能体协作（Hooks & Subagents），你就能告别盲目试错，真正拥有与智能体顺畅配合、独立把完整作品构建出来（Independent Building）的工程能力。 Even if your manual coding background is still growing, mastering System Prompts, Context Engineering, MCP, Agent Skills, and AGENTS.md & Multi-Agent Collaboration empowers you to collaborate fluently with agents and achieve Independent Building.

  • System Prompt
  • Context Engineering
  • MCP Protocol
  • Agent Skills
  • AGENTS.md & Hooks
  • Subagent Collaboration
Layer 02 · 上层能力 → 迈向「价值定义」 Layer 02 · Upper Ceiling → Advance to Value Definition

软件工程品味、设计艺术与价值定义 Engineering Taste, Design Art & Value Definition

操作接口本身并不是核心资产——它背后凝聚的问题拆解力、高内聚低耦合的软件工程品味、系统提示词与工作流的设计艺术，才是真正拉开差距的上层能力。当你跨过独立构建的门槛后，这些审美与思辨将支撑你回答那个终极问题：「我要为谁创造什么真实价值（Value Definition）？」 Config files alone are not your moat—the true upper-layer capability lies in problem decomposition, software architectural taste, and the design art behind prompts and workflows. Once you can build independently, this taste empowers you to answer the ultimate question: “What real value am I defining, and for whom?”

  • 软件工程品味 Architectural Taste
  • 提示词与系统设计艺术 System Design Art
  • 工程权衡与安全边界 Trade-offs & Safety
  • 价值定义 (Value Definition) Value Definition

课程进阶图谱 · 把断掉的路接起来 Curriculum Ascent · Connecting the Path

从「驾驭协作接口」到「独立构建」，最终迈向「价值定义」 From Mastering Agent Interfaces to Independent Building & Value Definition

山顶锚点：独立构建 → 价值定义 Summit: Independent Building → Value Definition
01

看懂智能体内核 Agent Internals

ReAct 循环 · System Prompt 拆解 ReAct Loop & System Prompts

02

上下文工程与协议 Context & MCP

Context Engineering · MCP 工具人体工学 Context Tax & MCP Ergonomics

03

技能沉淀与多智能体协作 Agent Skills & Collaboration

Agent Skills · AGENTS.md · Hooks 与子智能体 Agent Skills, AGENTS.md & Subagents

04

工程品味与安全护栏 Engineering Taste & Safety

Clean Code · 架构可视化 · 安全与软件工厂 Architecture Taste, Security & Scale

  • ANCHOR 01

    对抗碎片化，建立体系网 Systemic Mental Models over Fragmented Tips

    告别在收藏夹里囤积零散的「神级提示词」。每周吃透一个核心工程范式，把碎片经验串联成扎实的软件工程知识网。 Stop hoarding fragmented tricks. Master one foundational engineering paradigm each week to build a coherent mental model.

  • ANCHOR 02

    锤炼「Token 效能」 Token Efficiency & Leverage

    学会「用更少的 Token 办更大的事」——懂得何时精炼上下文、何时把确定性逻辑交给 Agent Skills 脚本、何时合理路由模型。 Learn to do more with fewer tokens—knowing when to prune context, offload deterministic steps to Agent Skills, and route models wisely.

  • ANCHOR 03

    不考手搓代码，重在拆解品味 Design Taste over Rote Boilerplate

    全课配备开箱即用的 Playground 范例工程。我们借助 AI 编写代码，把精力聚焦在看懂机制、调优接口与打磨架构审美上。 Every week ships with a ready-to-run Playground repo. Write code alongside AI while focusing your energy on architecture and design critique.

  • ANCHOR 04

    真实共同体，对抗 AI 迎合 Real Community vs. AI Sycophancy

    大模型天生倾向于顺从与迎合（Sycophancy）。线下课堂与沙龙存在的意义，是在真实同路人的碰撞中探讨工程取舍与技术边界。 LLMs are inherently sycophantic. Our classroom and salon exist to debate real engineering trade-offs and boundaries face-to-face.

课堂结构设计 · 20 + 25 + 25 Session Architecture · 20 + 25 + 25

三段分层工作坊：通识破局 · 动手实验室 · 设计哲学沙龙 Three-Stage Tiered Workshop: Concept · Playground Lab · Design Philosophy Salon

每周课程（约 70 分钟）严格划分为三个递进板块，对应青禾·元野的三种参与深度（旅人 Visitor / 同路人 Fellow / 共建者 Builder），每个人都能按自己的节奏自由生长。 Every 70-minute weekly session is structured into three progressive stages, welcoming casual Visitors, hands-on Fellows, and deep-diving Builders alike.

Stage 01 · Concept & Mechanism 20 min

通识破局：看懂底层机制 Concept & Mechanism

零代码门槛的高密度中文导讲。用直观架构图把本周核心概念（如 ReAct 循环、上下文降智区、MCP 协议、Agent Skills 渐进披露）讲透，听完前 20 分钟即可带走完整认知框架。 Zero-code-barrier visual lecture. Clear architectural diagrams demystifying the week's core mechanism so anyone can walk away with a solid mental model in 20 minutes.

适合人群：For: 全体成员（旅人 Visitor / 同路人 Fellow / 共建者 Builder） All Members (Visitor / Fellow / Builder)
Stage 02 · Hands-on Playground 25 min

动手实验室：Playground 对照实战 Hands-on Playground Lab

不考手搓底层代码。在预置好的 Playground 范例工程上，使用社团课堂 Token 池，亲手拆解 System Prompt、优化 Context、挂载 Agent Skills 或配置 Hooks，同屏对比改动前后的行为与 Token 账单。 No manual boilerplate coding required. Using our pre-built Playground repos and club Token pool, refine System Prompts, curate Context, mount Agent Skills, and compare before-vs-after behavior live.

适合人群：For: 实践探索者（同路人 Fellow / 共建者 Builder） Practitioners (Fellow / Builder)
Stage 03 · Design Philosophy Salon 25 min

设计哲学沙龙：工程品味与思辨 Design Philosophy Salon

关掉课件，像赏析艺术品一样逐句批注生产级 System Prompt 与架构案例，探讨技术背后的软件工程品味、行业能力狂飙与安全护栏的真实权衡，激荡独立思考。 Close the slides for live prompt critique and open debate on architectural taste, industry trade-offs between capability velocity and safety, and value definition.

适合人群：For: 深度思考者（共建者 Builder & 自由留堂研讨） Deep Divers (Builders & Open Roundtable)

十周课程大纲 · 四大进阶篇章 10-Week Syllabus · Four Progressive Phases

从「看清黑箱」到「多智能体软件工厂」的十周工程进阶 Ten Weeks from Agent Internals to Multi-Agent Software Factory

点击任意周次卡片即可展开查看本周的「通识破局 / 动手实验室 / 设计哲学沙龙」三段详情。支持按篇章筛选或一键展开全部。 Click any week card to inspect its Concept, Hands-on Playground, and Design Philosophy Salon details. Filter by phase or expand all weeks at once.

Phase I · Week 01 看清黑箱 Demystify the Black Box ReAct 循环与从 Nano 到生产级的系统提示词架构 ReAct Loop & Nano-to-Production System Prompts
Phase II · Week 02–04 驾驭上下文与工具 Master Context & Skills Context Engineering、MCP 协议、Agent Skills 与多智能体协作 Context Engineering, MCP, Agent Skills & Subagents
Phase III · Week 05–07 锤炼工程品味 Cultivate Engineering Taste Agent 友好架构、视觉化审美与能力/安全双刃剑 Agent-Ready Taste, Visual Architecture & Security
Phase IV · Week 08–10 迈向独立创造 Independent Creation 并行智能体编排、Token 经济学与软件工厂 Demo Day Parallel Agents, Token Economics & Software Factory
WEEK 01 智能体内核解剖：什么是 ReAct Agent？从极简到生产级 System Prompt Internals of Coding Agents: ReAct Loop & Nano-to-Production System Prompts 第一篇 · 看懂它 Phase I · Understand It

不考手搓 200 行 Python。带你拆解大模型如何通过「推理 + 行动 + 观察」变成编程智能体，并从 15 行极简 Nano Agent System Prompt 一路剖析到 Claude Code 级别的生产级系统提示词。 Demystify how an LLM becomes a coding agent via Reasoning + Acting + Observation, and progressively dissect System Prompts from a 15-line Nano Agent to production-grade Claude Code.

ReAct LoopSystem PromptTool CallingNano Agent
§01.1 · 通识破局 §01.1 · Concept & Mechanism 20 min

ReAct 循环本质与 System Prompt 渐进式解剖 The ReAct Loop & Progressive System Prompt Anatomy

  • 大模型本身只会「文字接龙」，是如何变成能读文件、改代码、跑终端的智能体的？拆解 ReAct（Reasoning + Acting + Observation）事件循环与 Tool Use 的 JSON 报文往返。
  • 从自制的 15 行极简 Nano Agent System Prompt 出发，逐层对照生产级智能体（如 Claude Code）完整提示词架构：行为准则、专用工具签名（Tool Schemas）、Plan Mode、Compaction 上下文压缩与 Memory。
  • How does a text-completion model become an agent that reads files, edits code, and runs terminal commands? Trace the JSON message round-trip of the ReAct (Reasoning + Acting + Observation) loop.
  • Start from our 15-line Nano Agent System Prompt and progressively compare it with production prompts (Claude Code): behavioral rules, Tool Schemas, Plan Mode, Compaction, and Memory.
§01.2 · 动手实验室 §01.2 · Hands-on Playground 25 min · Nano Agent Playground

报文流透视与 System Prompt 调优挑战 Live Message Inspector & System Prompt Tuning

  • 透视报文流：在带实时 Token 计价器与 JSON 报文透视仪的 Playground 中运行排错任务，亲眼观察大模型与本地工具之间的每一次调用往返。
  • System Prompt 拆解与重构实验：无需手写底层 Python 代码，仅通过修改与借助 AI 调优 Nano Agent 的系统提示词与工具边界描述，直观对比「粗糙提示词」与「生产级结构化提示词」在排错成功率与 Token 消耗上的巨大差距。
  • Inspect the live wire: run debugging tasks inside our Playground with a real-time Token meter and JSON message inspector to watch every LLM-to-tool round-trip.
  • System Prompt refactoring challenge: without writing low-level Python boilerplate, refine the Nano Agent System Prompt and tool descriptions to compare naive prompts against structured production prompts.
§01.3 · 设计哲学沙龙 §01.3 · Design Philosophy Salon 25 min

逐句批注生产级 System Prompt 的设计艺术 Close-Reading Production System Prompts as Design Art

「为什么顶尖智能体的系统提示词里，会有那么多看似『反直觉』的硬性禁令？」 “Why do top-tier coding agent system prompts contain so many seemingly counter-intuitive prohibitions?”

为什么生产级 System Prompt 严令「优先使用专用工具而非通用 bash」？为什么强制规定「工具被拒绝后绝不能原样重试」？为什么「工具越少、越正交，智能体反而越聪明」？我们将像赏析建筑图纸一样，拆解一段优秀 System Prompt 背后的软件工程哲学。 Why mandate dedicated file tools over raw bash? Why forbid retrying an identical tool call after user rejection? Why do fewer, orthogonal tools make an agent smarter? We critique production system prompts like architectural blueprints.

实验产出： Lab Deliverable: 调优后的 Nano Agent System Prompt 配置文件与前后 Token 消耗对比记录 Refactored Nano Agent System Prompt config & before/after Token spend log
WEEK 02 上下文工程（Context Engineering）与 MCP 协议 Context Engineering & Model Context Protocol (MCP) 第二篇 · 教会它 Phase II · Teach It

理解「上下文税」与「降智区（Dumb Zone）」，掌握智能体连接外部世界与开发工具的标准协议 MCP，学会面向大模型设计高信噪比的工具接口。 Understand the Context Tax and the "Dumb Zone", master the Model Context Protocol (MCP), and design high-signal, ergonomic tools tailored for LLMs.

Context EngineeringDumb ZoneMCPAgent Ergonomics
§02.1 · 通识破局 §02.1 · Concept & Mechanism 20 min

上下文税（Context Tax）、降智区与 MCP 工具人体工学 The Context Tax, Dumb Zone & MCP Tool Ergonomics

  • 为什么对话越长、塞入的参考资料越多，智能体反而越笨、越慢、越贵？剖析注意力稀释与「降智区（Dumb Zone）」，掌握何时压缩（Compact）、何时果断开启新会话。
  • MCP（Model Context Protocol）核心机制：智能体连接外部数据源与工具的统一接口，以及什么是「面向智能体的工具人体工学（Agent Ergonomics）」。
  • Why do longer conversations and bloated context windows make agents slower, pricier, and dumber? Analyze attention dilution and the "Dumb Zone", and learn when to compact versus start fresh.
  • Model Context Protocol (MCP) fundamentals: how agents connect to external data and tools, and what "Agent Ergonomics" means when designing tool schemas.
§02.2 · 动手实验室 §02.2 · Hands-on Playground 25 min · Context & MCP Lab

上下文信噪比实验与 MCP 工具封装对比 Context Signal-to-Noise Experiment & MCP Tool Wrapping

  • 上下文信噪比对照：对比「把整份文档/全量日志一股脑塞入对话」与「投喂结构化高信噪比上下文」在排错准确率与 Token 账单上的显著差异。
  • MCP 外设调优：为智能体挂载自定义 MCP 工具，对比「直接暴露细碎的原始 REST API」与「封装为意图清晰的高内聚工具」时的调用成功率。
  • Signal-to-noise benchmark: compare dumping raw logs/docs into context versus curating structured context on task accuracy and token spend.
  • MCP tool ergonomics: connect a custom MCP server and compare exposing chatty REST endpoints directly versus wrapping them into high-cohesion, agent-ergonomic tools.
§02.3 · 设计哲学沙龙 §02.3 · Design Philosophy Salon 25 min

上下文管理的取舍哲学与人机接口差异 Philosophy of Context Curation & Human vs. Agent Interfaces

「为什么给人类程序员设计的好 API，直接拿给智能体用往往是一场灾难？」 “Why is a great API for human developers often a terrible tool interface for an LLM agent?”

人类擅长在文档里组合几十个原子接口，而大模型每次工具往返都在燃烧上下文窗口与注意力预算。探讨在人机协作中，哪些知识应该沉淀在代码库，哪些应该按需检索，哪些应该用完即焚。 Humans excel at composing dozens of fine-grained primitives; LLMs pay a steep context and latency tax on every round-trip. We discuss what knowledge belongs in the repo, what should be fetched on demand, and what should be ephemeral.

实验产出： Lab Deliverable: 高信噪比上下文改造报告与高内聚 MCP 工具封装对比实验 High-signal context benchmark & ergonomic MCP tool wrapper comparison
WEEK 03 Agent Skills 与 CLI：用 SkillHub 沉淀可复用能力 Agent Skills & CLI: Building Reusable Capabilities with SkillHub 第二篇 · 教会它 Phase II · Teach It

深入理解 Agent Skills 的工作原理与渐进式披露机制，直观对比「有 Skill」与「无 Skill」的天壤之别，并开启青禾·元野 SkillHub 共建。 Master how Agent Skills and progressive disclosure work, witness the stark contrast between running with and without Skills, and contribute to the Proto Commons SkillHub.

Agent SkillsProgressive DisclosureCLI ScriptsSkillHub
§03.1 · 通识破局 §03.1 · Concept & Mechanism 20 min

什么是 Agent Skills？渐进式披露与能力边界 What Are Agent Skills? Progressive Disclosure & Boundaries

  • Agent Skills 核心机制：如何通过按需加载的领域指令（如 `SKILL.md`）配合确定性的本地 CLI 脚本，把个人摸索出的优秀工作流转化为团队可复用的智能体能力。
  • 渐进式披露（Progressive Disclosure）：为什么给智能体安装几十个 Agent Skills 也不会撑爆上下文？厘清一次性 Prompt、Agent Skills 与 MCP 三者的职责边界。
  • How Agent Skills combine on-demand domain instructions with deterministic local CLI scripts to turn personal workflows into reusable team capabilities.
  • Progressive Disclosure: why mounting dozens of Agent Skills never bloats the initial context window, and how to draw clean boundaries between Prompts, Agent Skills, and MCP.
§03.2 · 动手实验室 §03.2 · Hands-on Playground 25 min · SkillHub Arena

同题对决（有 Skill vs. 无 Skill）与 SkillHub 封装实战 With-Skill vs. Without-Skill Showdown & SkillHub Packaging

  • 同题直观对决：在同一个多步骤工程构建任务上，同屏对比「无 Skill 裸模型反复试错」与「挂载 Agent Skill 后一次成型」在产出质量、耗时与 Token 消耗上的悬殊差距。
  • 共建元野 SkillHub：借助 AI 将自己常用的一个高频开发或科研工作流打包为标准 Agent Skill，现场验证并发布至社团 SkillHub。
  • Side-by-side showdown: run the same multi-step engineering task with a bare model versus an agent equipped with a curated Agent Skill, comparing output quality, retries, and token cost.
  • Contribute to Proto SkillHub: use AI to package a high-frequency workflow into a standardized Agent Skill and publish it to our community SkillHub.
§03.3 · 设计哲学沙龙 §03.3 · Design Philosophy Salon 25 min

把确定性的事交给代码，把不确定性的事交给模型 Leave Determinism to Code, Leave Ambiguity to the Model

「既然大模型什么代码都能现写，为什么顶尖构建者还要在 Agent Skill 里放一堆现成的 CLI 脚本？」 “If LLMs can write code on the fly, why do top builders bundle deterministic CLI scripts inside Agent Skills?”

让大模型每次现场手搓数据转换或校验脚本，既昂贵又充满随机幻觉。探讨如何识别工作流中的「确定性骨架」与「创造性判断」，以及为什么开源社区的竞争正从单纯拼模型转向拼 Agent Skills 生态。 Having an LLM re-invent parsing or verification scripts on every run wastes tokens and invites hallucinations. We explore how to separate deterministic scaffolding from creative judgment in agent workflows.

实验产出： Lab Deliverable: 1 个经过对照验证并提交至青禾·元野 SkillHub 的标准 Agent Skill One verified Agent Skill package contributed to Proto Commons SkillHub
WEEK 04 智能体协作与仓库定制：AGENTS.md、Hooks 与子智能体编排 Agentic Collaboration & Repo Harness: AGENTS.md, Hooks & Subagents 第二篇 · 教会它 Phase II · Teach It

学会编写高信噪比的仓库级宪法 `AGENTS.md`，用生命周期 Hooks 建立确定性护栏，并通过子智能体（Subagents）分工协作隔离脏上下文。 Author high-signal `AGENTS.md` repository constitutions, enforce deterministic guardrails with lifecycle Hooks, and orchestrate Subagents to isolate noisy exploration.

AGENTS.mdLifecycle HooksSubagentsMulti-Agent Collaboration
§04.1 · 通识破局 §04.1 · Concept & Mechanism 20 min

AGENTS.md 编写法则、Hooks 确定性拦截与子智能体协作 AGENTS.md Principles, Deterministic Hooks & Subagent Collaboration

  • `AGENTS.md` / `CLAUDE.md` 黄金法则：什么该写（构建与测试命令、核心架构边界、非显而易见的陷阱），什么绝不该写（冗长的目录树或泛泛而谈的废话）。
  • Hooks 与多智能体协作（Subagents）：在工具执行前后挂载确定性检查钩子（自动格式化、跑单测、高危命令拦截），并用 `Researcher / Planner / Implementer` 子智能体切分上下文。
  • Golden rules for `AGENTS.md` / `CLAUDE.md`: what to include (build/test commands, architectural invariants, non-obvious pitfalls) and what bloat to delete.
  • Lifecycle Hooks & Subagent Collaboration: attach deterministic `PreToolUse` / `PostToolUse` hooks and orchestrate `Researcher / Planner / Implementer` subagents to keep the main context clean.
§04.2 · 动手实验室 §04.2 · Hands-on Playground 25 min · Harness & Subagents Lab

驯服失控工程：AGENTS.md + Hooks + 子智能体实战 Taming an Unruly Repo with AGENTS.md, Hooks & Subagents

  • 仓库护栏改造：接手一个智能体一改就破、风格混乱的示例工程，为其编写高信噪比 `AGENTS.md` 并配置自动校验 Hook，直观感受工程稳定性的跃升。
  • 多智能体协作演练：将一项涉及跨模块调研与重构的任务拆解给只读探索子智能体与执行子智能体协同完成。
  • Repo harness makeover: take a fragile sample project where agents repeatedly break builds, add a crisp `AGENTS.md` and verification Hooks, and measure the jump in reliability.
  • Multi-agent collaboration drill: delegate codebase research and implementation across specialized subagents without polluting the primary conversation.
§04.3 · 设计哲学沙龙 §04.3 · Design Philosophy Salon 25 min

约束与自由的张力：如何防止项目沦为 AI Slop？ Constraint vs. Autonomy: Preventing AI Slop through Engineering Guardrails

「规则写得太死会让智能体寸步难行，不写规则又会堆出满屏低质代码（AI Slop）——护栏的黄金分割点在哪里？」 “Over-constraining an agent paralyzes it; under-constraining it produces AI slop. Where is the golden ratio of guardrails?”

真正的团队工程规范不是靠人在对话里一遍遍叮嘱，而是把软性约定沉淀进 `AGENTS.md`，把硬性底线变成自动化 Hooks。探讨人机协作与多智能体协作中的「立法权」与「执行权」。 Sustainable engineering standards do not rely on nagging the model in chat—they encode architectural intent in `AGENTS.md` and enforce hard invariants via Hooks.

实验产出： Lab Deliverable: 为示例工程定制的 `AGENTS.md`、自动化校验 Hook 配置与子智能体协作记录 Custom `AGENTS.md`, automated verification Hook config & subagent workflow trace
WEEK 05 Agent-Ready 代码库与软件工程品味：模块化、测试闭环与规格驱动 Agent-Ready Codebases & Software Engineering Taste: Modularity, Verification & Specs 第三篇 · 信得过 Phase III · Trust It

为什么有的项目 AI 越改越乱，有的项目却能自主完成复杂迭代？探讨 AI 时代的 Clean Code、自验证测试闭环与规格驱动开发（Spec-Driven Development）。 Why do agents spiral into chaos in some repos yet ship complex features autonomously in others? Explore Clean Code for agents, self-verification loops, and Spec-Driven Development.

Agent-Ready RepoSoftware TasteVerification LoopSpec-Driven
§05.1 · 通识破局 §05.1 · Concept & Mechanism 20 min

什么是 Agent-Ready 代码库？自验证闭环与规格驱动 What Makes a Codebase Agent-Ready? Verification Loops & Specs

  • AI 时代的软件工程品味：为什么强类型签名、清晰的模块边界、单文件职责内聚与快速反馈的单元测试，比以往任何时候都更重要？
  • 自验证闭环（Self-Verification Loop）与规格驱动（Spec-Driven Development）：如何给智能体提供可执行的「验收标尺」，让它在交付前自我纠错。
  • Software engineering taste in the AI era: why strict types, explicit module boundaries, cohesive files, and fast unit tests matter more than ever.
  • Self-Verification Loops & Spec-Driven Development: giving agents executable acceptance criteria so they catch and fix their own bugs before handoff.
§05.2 · 动手实验室 §05.2 · Hands-on Playground 25 min · Architecture & Spec Lab

面条代码库 vs. Agent-Ready 模块化架构对比实验 Spaghetti Repo vs. Agent-Ready Modular Architecture

  • 架构友好度对照：让同一个智能体分别在「耦合严重的单文件面条工程」与「契约清晰、带快速测试套件的模块化工程」上实现同一新特性，观察两者的成功率差异。
  • 规格先行演练：先与智能体对齐一份简明可验证的 `SPEC.md` 与验收测试，再交由智能体自主实现并跑通闭环。
  • Architecture benchmark: task the same agent with adding a feature to a tangled monolithic repo versus a modular, typed repo with fast tests, and compare success rates.
  • Spec-first workflow: co-author a concise `SPEC.md` and acceptance test suite before letting the agent implement and self-verify.
§05.3 · 设计哲学沙龙 §05.3 · Design Philosophy Salon 25 min

当写代码不再昂贵，软件工程的核心壁垒转移到了哪里？ When Writing Code Is Cheap, Where Does Engineering Value Live?

「我们不会手写每一行底层代码，但我们凭什么判断 AI 写出的架构是优雅还是腐烂的？」 “Even if we don't hand-type every line of code, how do we cultivate the taste to tell elegant architecture from rotting complexity?”

对标 CS146S 嘉宾 Martin Fowler 与 Dan Terhorst-North 的核心洞见：过去几十年软件工程沉淀的「高内聚、低耦合、契约先行」不仅没有过时，反而成为了决定你能否指挥智能体构建大型系统的天花板。 Drawing on insights from Martin Fowler and Dan Terhorst-North: decades of software engineering wisdom—high cohesion, loose coupling, and clear contracts—have become the very ceiling of what you can build with agents.

实验产出： Lab Deliverable: 模块化重构前后智能体任务成功率对比与可执行验收测试套件 Before/after modular refactoring benchmark & executable acceptance test suite
WEEK 06 「话不可说」：走出纯文本对话，用架构可视化掌控全局 Beyond the Chat Box: Visual Architecture & Cognitive Command 第三篇 · 信得过 Phase III · Trust It

人类擅长空间与拓扑感知，不擅长阅读万行陌生代码。结合开源项目 Archify 等架构可视化工具，学会把智能体产出的代码还原为可交互的系统蓝图。 Humans excel at spatial and topological reasoning, not reading 10,000 lines of AI-generated diffs. Use open-source tools like Archify to turn codebases into interactive architectural maps.

Visual ArchitectureArchifyCode ReviewMental Model
§06.1 · 通识破局 §06.1 · Concept & Mechanism 20 min

认知带宽瓶颈与代码架构可视化（如 Archify） Cognitive Bandwidth & Interactive Architecture Visualization (Archify)

  • 「生成速度远超人类阅读速度」的困境：当智能体在几分钟内改动十几个文件，纯文本聊天框与传统逐行 Diff 为什么会让人彻底丧失对系统的掌控感？
  • 走出纯文本界面：介绍开源项目 Archify 等代码库可视化与依赖拓扑生成工具，如何将抽象代码实时映射为模块依赖图、数据流图与状态机视图。
  • The review bottleneck: when an agent touches a dozen files in two minutes, why do plain-text chat boxes and raw line diffs cause humans to lose mental command of the system?
  • Stepping beyond text: how open-source tools like Archify transform codebases into interactive module dependency graphs, data-flow diagrams, and state machines.
§06.2 · 动手实验室 §06.2 · Hands-on Playground 25 min · Visual Review Lab

借助架构可视化图谱揪出「看似跑通实则越界」的隐蔽缺陷 Catching Architectural Drift with Visual Dependency Maps

  • 可视化架构体检：使用架构可视化工具（如 Archify / 交互式依赖图生成器）扫描一段智能体生成的复杂功能分支，在不通读全量源码的情况下定位循环依赖与跨层违规调用。
  • 图谱驱动重构：基于可视化架构图向智能体下达精准的模块解耦指令，验证重构前后的拓扑改善。
  • Visual architecture audit: scan an agent-generated feature branch with an interactive architecture mapper (such as Archify) to spot circular dependencies and layer violations without reading every line.
  • Graph-guided refactoring: use the visual topology to instruct the agent to decouple tangled modules and verify the cleaned-up graph.
§06.3 · 设计哲学沙龙 §06.3 · Design Philosophy Salon 25 min

永远不要外包你的系统心智模型（Mental Model） Never Outsource Your System Mental Model

「如果一个项目虽然能跑起来，但你已经完全画不出它的核心数据流向图，它还算是你的作品吗？」 “If your app runs, but you can no longer sketch its core data flow on a whiteboard, are you still its builder?”

智能体可以代劳实现细节，但绝不能替代构建者脑中的系统全局图景。探讨「话不可说」背后的交互哲学——未来的人机协作界面应当如何超越聊天框，服务于人类的直觉与审美判断。 Agents can own implementation details, but they must never replace the builder's mental model. We discuss how human-AI interfaces must evolve beyond text chat to amplify human spatial intuition.

实验产出： Lab Deliverable: 示例工程的可视化架构诊断图谱与跨层解耦重构记录 Visual architecture dependency audit & layer-decoupling refactoring log
WEEK 07 智能体安全攻防：致命三角、提示词注入与纵深防御 Agent Security & Defense: The Lethal Trifecta, Prompt Injection & Guardrails 第三篇 · 信得过 Phase III · Trust It

系统梳理现有智能体安全研究与常见攻击面（间接提示词注入、MCP 投毒、越权执行），探讨为什么当前行业漏洞频发以及如何构建沙箱纵深防御。 Survey state-of-the-art agent security research and attack surfaces (indirect prompt injection, MCP poisoning, privilege escalation), and examine why vulnerabilities proliferate in today's race for capability.

Lethal TrifectaPrompt InjectionSandboxingSecurity Philosophy
§07.1 · 通识破局 §07.1 · Concept & Mechanism 20 min

智能体「致命三角（Lethal Trifecta）」与既有攻防研究 The Agent "Lethal Trifecta" & Existing Security Frameworks

  • 梳理业界核心安全研究：什么是智能体安全的「致命三角」（同时具备：接触不可信外部输入 + 访问私有敏感数据 + 对外通信或执行命令的能力）？
  • 常见攻击面与现有防御手段：间接提示词注入（Indirect Prompt Injection）、恶意仓库配置与 MCP 工具投毒，以及文件系统/网络沙箱（Seatbelt / Bubblewrap / 容器隔离）与最小权限原则。
  • Core security frameworks: what is Simon Willison's "Lethal Trifecta" (combining untrusted input + private data access + external communication/execution)?
  • Attack vectors & existing defenses: indirect prompt injection, malicious repo configs, MCP tool poisoning, and defense-in-depth via OS sandboxes and least privilege.
§07.2 · 动手实验室 §07.2 · Hands-on Playground 25 min · Security Sandbox CTF

隔离靶场中的间接注入复现与沙箱加固演练 Indirect Prompt Injection & Sandbox Hardening in an Isolated Range

  • 攻防复现：在完全隔离的本地教学靶场中，观察一段藏在外部网页或文档中的隐蔽指令如何诱导未设防的智能体泄露环境变量。
  • 纵深防御加固：通过拆分工具权限、开启只读沙箱策略与敏感操作人工确认（Human-in-the-loop），彻底阻断攻击链并验证防御效果。
  • Red-team replay: in an isolated local training range, observe how a hidden instruction inside an untrusted document tricks an unguarded agent into exfiltrating mock secrets.
  • Blue-team hardening: break the kill chain by separating tool privileges, enforcing read-only sandboxes, and adding human-in-the-loop confirmation gates.
§07.3 · 设计哲学沙龙 §07.3 · Design Philosophy Salon 25 min

为什么现在的智能体漏洞这么多？能力狂飙与安全护栏的行业困局 Why Are Agent Vulnerabilities Everywhere? The Capability Rush vs. Safety Reality

「今天市面上层出不穷的智能体安全漏洞，是因为工程师们不知道怎么做安全吗？」 “Are today's rampant agent security flaws happening because engineers don't know how to build secure systems?”

探讨技术背后的真实产业逻辑：安全到底重不重要？为什么明知存在致命三角，各大厂商与开源项目依然在默认放开权限？因为所有人都在争分夺秒推进能力边界、抢占市场心智。作为清醒的独立构建者，我们该如何在「极致便利」与「安全底线」之间做出负责任的设计抉择？ We examine the industry reality: vulnerabilities proliferate not from ignorance, but because labs and startups are racing to ship capabilities and capture market share. How should thoughtful builders balance frictionless autonomy with responsible safety boundaries?

实验产出： Lab Deliverable: 靶场攻击链阻断配置与最小权限沙箱策略清单 Kill-chain mitigation config & least-privilege sandbox policy checklist
WEEK 08 超越本地终端：Git Worktree 并行流与云端异步智能体 Beyond the Local Terminal: Git Worktree Parallelism & Cloud Background Agents 第四篇 · 放大它 Phase IV · Scale It

从单线程「守着屏幕等 AI 敲代码」升级为多任务并行编排，理解 Git Worktree 本地隔离与云端后台智能体（Background Agents）的工程设计。 Graduate from babysitting a single terminal to orchestrating parallel workflows using Git Worktrees and cloud Background Agents.

Git WorktreeBackground AgentsAsync OrchestrationPR Automation
§08.1 · 通识破局 §08.1 · Concept & Mechanism 20 min

同步结对 vs. 异步委派：并行智能体的工作空间隔离 Synchronous Pairing vs. Async Delegation: Workspace Isolation

  • 为什么多个智能体不能在同一个工作目录里同时改代码？解析 `git worktree` 如何零成本共享 Git 历史、同时提供互不干扰的并行工作区。
  • 云端后台智能体（Cloud / Background Agents）架构导览：从 Issue 触发、容器环境冷启动、自动自测到提交 Pull Request 的异步流水线原理（理解架构思想，不要求重度基建落地）。
  • Why can't multiple agents edit the same working tree simultaneously? How `git worktree` shares repository history while isolating parallel agent workspaces.
  • Cloud Background Agents architecture: trace how an issue triggers an ephemeral container, runs automated verification, and opens a clean Pull Request.
§08.2 · 动手实验室 §08.2 · Hands-on Playground 25 min · Parallel Workflow Lab

双路 Worktree 并行开发与多方案同题对比 Dual-Worktree Parallel Development & Design Exploration

  • 双线并行演练：利用 `git worktree` 同时拉起两个独立工作区，让两个智能体并行推进不同模块（或对同一需求探索两种不同架构实现），最后以架构师视角择优合并。
  • 异步任务拆解练习：演练如何编写一份「无需中途追问即可独立交付」的高质量异步任务工单。
  • Dual-worktree drill: spin up two isolated `git worktree` checkouts and run two agents in parallel—either building orthogonal features or exploring two competing designs—then review and merge the winner.
  • Async ticket specification: practice writing self-contained task briefs that an asynchronous agent can execute without mid-flight hand-holding.
§08.3 · 设计哲学沙龙 §08.3 · Design Philosophy Salon 25 min

当一个人就是一支工程队，构建者的注意力该如何分配？ When One Builder Commands a Squad of Agents, What Becomes the Bottleneck?

「同时开 5 个智能体看似酷炫，为什么很多人最后反而陷入了更深的焦虑与返工？」 “Running five agents at once looks impressive—so why does it often lead to cognitive overload and messy rework?”

并行化的真正瓶颈从来不是算力，而是构建者自己的需求清晰度与验收判断力。在没有想清楚问题边界之前，盲目并行只会以 5 倍速度制造技术债。 The bottleneck in parallel agent workflows is never compute—it is the human builder's clarity of thought and review bandwidth. Without clear specs, parallelism just multiplies technical debt by 5x.

实验产出： Lab Deliverable: 双分支并行探索对比记录与择优合并 PR Dual-worktree parallel implementation comparison & merged PR
WEEK 09 善假于物：多模型路由、Token 经济学与线上可观测闭环 Smart Leverage: Multi-Model Routing, Token Economics & Production Observability 第四篇 · 放大它 Phase IV · Scale It

打破「唯单一最贵模型论」，建立前沿推理模型与轻量高速模型的组合拳思维，并理解代码上线后的日志监控与故障闭环。 Move beyond relying on a single pricey frontier model. Master multi-model routing, token ROI, and post-deployment observability.

Token EconomicsModel RoutingEval & ObservabilityROI
§09.1 · 通识破局 §09.1 · Concept & Mechanism 20 min

模型能力象限、Token 经济学与生命周期后半程 Model Capability Quadrants, Token Economics & Post-Deploy Loops

  • 没有「通吃一切」的模型：拆解高推理模型（深度架构规划、疑难杂症排错）与高吞吐/低延迟轻量模型（高频探索、定式转换、子任务执行）的协同路由策略。
  • 软件生命周期的后半程：作品跑起来只是开始，理解线上结构化日志、可观测性（Observability）与智能体辅助故障定位（Incident Triage）。
  • No single model rules them all: orchestrating frontier reasoning models (architecture, gnarly bugs) alongside fast, low-cost models (high-volume exploration, routine transforms).
  • The second half of software engineering: structured logging, production observability, and agent-assisted incident triage after code ships.
§09.2 · 动手实验室 §09.2 · Hands-on Playground 25 min · Token ROI Challenge

固定 Token 预算下的「规划 + 执行」高低配组合挑战 Frontier-Planner + Fast-Executor Routing under a Strict Token Budget

  • Token 预算挑战赛：在给定的严格 Token 额度内完成一项复合工程任务，对比「全程无脑调用最贵旗舰模型」与「旗舰模型做架构规划 + 轻量模型批量执行 + 本地脚本校验」的成本效益比（ROI）。
  • Token budget challenge: complete a multi-part build under a strict token cap, comparing naive all-frontier usage against a routed pipeline (frontier planner + fast executor + local verification scripts).
§09.3 · 设计哲学沙龙 §09.3 · Design Philosophy Salon 25 min

克制是一种高级的工程美学：算力丰裕时代的成本意识 Restraint as Engineering Elegance: Cost Consciousness in an Era of Abundant Compute

「为什么我们反复强调『用更少的 Token 办更大的事』，这只是为了省钱吗？」 “Why do we relentlessly emphasize doing more with fewer tokens—is it merely about saving money?”

高昂的 Token 消耗往往是思维懒惰与架构混乱的症状。当你能用精准的上下文、恰当的 Agent Skill 和合理的模型分层在十分之一的消耗下漂亮地解决问题时，这本身就是对系统理解透彻的证明。 Runaway token burn is usually a symptom of muddled thinking and sloppy context architecture. Solving a problem cleanly at one-tenth the token cost is proof of deep architectural mastery.

实验产出： Lab Deliverable: 严格 Token 预算下的多模型协同构建方案与成本账单分析 Multi-model routed build pipeline & token cost-efficiency breakdown
WEEK 10 软件工厂与「价值定义」：青禾·元野独立构建者结业展（Demo Day） The Software Factory & Value Definition: Proto Commons Builder Demo Day 第四篇 · 放大它 Phase IV · Scale It

展望「软件工厂（Software Factory）」前沿趋势，迎来十周课程的终章：展示你独立构建的完整作品或核心工作流，在共同体中探讨什么才是值得被创造的真实价值。 Explore the frontier of Software Factories and culminate our 10-week journey at Demo Day: showcase what you independently built and reflect on what is truly worth creating.

Software FactoryIndependent BuildingValue DefinitionDemo Day
§10.1 · 通识破局 §10.1 · Concept & Mechanism 20 min

从手工作坊到软件工厂（Software Factory）与人的不可替代性 From Craft Workshop to Software Factory & Human Irreplaceability

  • 前沿眺望：当规格（Spec）、上下文（Context）、技能（Agent Skills）、子智能体协作与自动化护栏（Hooks & CI）完全打通，现代软件工程正在如何演进为高自治的「软件工厂」？
  • 回到青禾·元野的初心：当「把东西做出来（Independent Building）」的门槛被你彻底跨过，你如何迈向更高维度的「价值定义（Value Definition）」？
  • Looking ahead: when Specs, Context Engineering, Agent Skills, Multi-Agent Collaboration, and automated Hooks converge, how does software development evolve into an autonomous Software Factory?
  • Returning to Proto Commons' core ethos: once you have crossed the threshold of Independent Building, how do you step up to Value Definition?
§10.2 · 动手实验室 §10.2 · Hands-on Playground 25 min · PTC-101 Demo Day

独立构建作品与核心工程范式结业路演 Independent Builder Showcase & Workflow Critique

  • 双轨结业展示：无论是你在十周内独立构建出的完整工具/应用/科研流水线，还是你打磨沉淀出的高价值 Agent Skill 与 System Prompt 工程范例，都将走上讲台进行现场演示与架构复盘。
  • 优秀作品与 Agent Skills 将正式收录进青禾·元野公共资产库与 `protocom.org` 产品名录。
  • Dual-track showcase: present either an end-to-end tool/application/research pipeline you independently built, or a battle-tested Agent Skill and prompt engineering kit, complete with live architecture walkthroughs.
  • Outstanding projects and Agent Skills are inducted into the Proto Commons public registry on `protocom.org`.
§10.3 · 设计哲学沙龙 §10.3 · Design Philosophy Salon 25 min

终章圆桌：当构建不再是瓶颈，什么才是值得我们去创造的？ Closing Roundtable: When Building Is No Longer the Bottleneck, What Is Worth Creating?

「当你已经拥有独立把脑海中的想法变成现实的能力，你准备用它去回应哪个真实的具体问题？」 “Now that you have the power to turn ideas into reality independently, what real human problem will you choose to answer?”

技术与工具终究是通往山顶的阶梯。在 `PTC-101` 的最后一课，我们一起复盘这十周从看懂黑箱到独立构建的蜕变，探讨在 AI 浪潮中如何保持清醒的审美、真诚的问题意识与共同体精神——各有其根，蔚成原野。 Tools and protocols are ultimately the ladder up the mountain. In our closing salon, we reflect on our 10-week ascent from opening the black box to independent building, and ask how to preserve taste, authenticity, and community spirit—own roots, common ground.

实验产出： Lab Deliverable: 独立构建作品演示 / 元野 SkillHub 核心资产收录 Live Demo Day presentation & induction into Proto Commons Registry

参与方式与双轨结业 Participation & Dual-Track Completion

无论你来自哪个专业，都能在这里找到自己的步调 Whatever Your Major, Find Your Own Pace Here

Q1 · 我的代码基础比较弱，平时基本靠 AI 写代码，能跟上吗？ Q1 · My manual coding background is weak and I rely on AI to write code. Can I keep up?

完全可以，这门课正是为你设计的。我们不考手搓算法或默写语法。在动手实验室中，我们直接基于预置好的 Playground 工程，教你如何看懂 ReAct 循环、如何用 AI 重构 System Prompt、如何管理 Context 与挂载 Agent Skills，让你从「盲目抽卡」进化为真正有掌控力的构建者。 Yes—this course is built specifically for you. We never test rote syntax memorization. Using pre-built Playgrounds, we teach you how to inspect ReAct loops, refine System Prompts with AI, curate Context, and mount Agent Skills so you move from blind prompting to confident building.

Q2 · 平时学业很忙，必须每周待满 70 分钟并写完大型作业吗？ Q2 · My schedule is packed. Do I have to attend all 70 minutes and build a massive project every week?

完全自由，丰俭由人。你可以作为「旅人（Visitor）」只听前 20 分钟通识导讲；也可以作为「同路人（Fellow）」留下来花 25 分钟在 Playground 上亲手验证对比；如果意犹未尽，再进入最后 25 分钟的设计哲学沙龙探讨工程品味。 Join at your own pace. Drop in as a Visitor for the 20-minute concept lecture, stay as a Fellow for the 25-minute Playground Lab, or dive into the final 25-minute Design Philosophy Salon as a Builder.

Q3 · 课堂实验需要自己充值昂贵的海外大模型 API 吗？ Q3 · Do I need to pay for expensive LLM API keys out of pocket?

不需要。青禾·元野为每次课堂的动手实验室提供预置的课堂临时 Token 额度池与本地 Playground 环境，你只需带上电脑（甚至与身旁同学结对）即可直接上手。 No. Proto Commons provides a shared classroom Token pool and pre-configured Playground environments for every lab session—just bring your laptop.

Q4 · 怎样获得 PTC-101 结业认证与社区共建者席位？ Q4 · How do I earn PTC-101 completion recognition and Builder status?

我们提供双轨结业通道：轨道 A（Playground 探索者）完成至少 6 周课堂 Playground 实验或向元野 SkillHub 贡献 1 个优质 Agent Skill；轨道 B（独立构建者）在 Week 10 Demo Day 展示一个用课程所学独立构建出来的真实工具、应用或工作流。 We offer two completion tracks: Track A (Playground Explorer)—complete at least 6 weekly Playground labs or contribute a reusable Agent Skill to SkillHub; Track B (Independent Builder)—showcase an independently built tool, app, or workflow at Week 10 Demo Day.