Compound Engineering

What it is

An AI-native engineering methodology developed by Kieran Klaassen (GM of Quora, creator of the Compound Engineering plugin) that structures human-AI collaboration as a loop where each iteration makes the system smarter. The core insight is that the "work" phase of software development is increasingly automatable, so human attention should concentrate at the edges: framing problems at the start and applying taste-based polish at the end.

Why it matters

Traditional development stops at Review. Compound Engineering adds a Compound step that feeds learnings back into the system, creating a self-improving loop. Klaassen uses this with a single-engineer team to operate five products. The framework also produces the "AI Sandwich" mental model: humans are the bread (problem-framing and final polish), AI is the filling (execution and iteration).

Key points

• The loop has eight steps, not four. Originally Plan → Work → Review → Compound, it now includes Ideate and Brainstorm at the front and Polish before Compound. The full cycle: Ideate (go wide, generate possibilities) → Brainstorm (go deep, understand a known problem) → Plan (set the execution frame) → Work (largely automated) → Review (catch errors) → Polish (apply taste) → Compound (store learnings) → Repeat.
• The Work phase is essentially solved. Klaassen's assessment, based on building Quora with agents: if you provide a good plan, LLMs can execute faithfully for hours or days. The bottleneck has shifted from execution to planning and taste.
• Compound is the most powerful step. Storing mistakes, solutions, and patterns as retrievable knowledge inside the repository means agents do not repeat errors. This is where single-engineer output scales toward team-level throughput.
• Humans frame; AI optimizes. Dan Shipper's addition: humans are good at "frame shifting" — redefining a problem at a different level of the stack. AI optimizes within a given frame. This asymmetry is why humans remain essential at the beginning of any non-trivial project.
• Taste is the final filter. Without a human polish step, AI output converges toward sameness ("slop"). The final review is not just error-checking; it is elevating output from functional to exceptional.
• Time allocation inverts traditional engineering. Plan + Review consume roughly 80% of attention; Work + Compound take 20%. Most engineers do the reverse.

The Process Becomes the Product（2026-06-13）

来源：Every 2026-06-13 — The Moral of Fable

Dan Shipper 将 Compound Engineering 的循环理念扩展到「知识工作即园艺」的隐喻：

• Gardening and Loops：开发者不再直接「种植」产品，而是创造条件让产品生长。收集用户反馈 → 转化为可执行计划 → AI 执行 → 审查 → 合并变更 → 整合经验 → 反馈到下一轮。每个阶段的输出成为下一个阶段的输入。
• 过程即产品（The process becomes the product）：当工作被循环化后，产品不是一次设计出来的，而是在持续循环中生长的。Cora 的邮件应用就是一个「反向发条玩具」——每转一圈都获得更多能量和动力，而不是逐渐减速。
• 从编码到所有知识工作：Claude Code 最初是开发者工具，现在同样的方法论被用于幻灯片、电子表格、文档等一切知识工作。循环设计正在从工程扩散到所有知识领域。
• 成本与可达性：Fable 5 的极高成本（2× token 价格 + 自动启动数十个子 Agent）意味着 Level 7-8 的 Compound Engineering 需要资本。这创造了新的不对称：能负担持续运行前沿模型的人与不能负担的人之间产生距离。

这与 Kieran Klaassen 的原始框架形成互补：Kieran 强调循环的「复合」步骤（存储知识防止重复错误），Dan Shipper 强调循环的「生长」特性（产品作为持续过程而非静态产物）。两者共同指向同一个结论：在 AI 时代，「设计循环系统」正在取代「设计产品」成为核心能力。

/goal 长时间运行的最佳实践（2026-05-06）

来源：宝玉 — Codex /goal 长时间运行经验

长时间运行的核心不是「时间长」，而是有清晰目标、验收标准、分阶段文档和中间进度追踪。宝玉的 17 小时逆向工程经验可归纳为迭代式规范构建：

1. 制定详细计划并固化：将目标、验收标准和完成定义写成 Agents.md，避免 AI 在模糊目标下空转
2. 初始化规范文件：明确 input/output 格式、约束和质量标准，作为契约而非建议
3. 试运行观察偏差：让 Agent 跑一小段，对照预期输出找 gap
4. 手动写出期望样板：针对偏差，人工写出「理想输出」作为 reference
5. 更新计划再运行：把样板吸收进 Agents.md，再次运行
6. 反复迭代直到稳定：通常 2-3 轮后，AI 就能按预期执行长时间任务

这与 Compound Engineering 的 Plan 步骤高度同构：规范即 Plan，试运行即 Review，样板沉淀即 Compound。区别仅在于时间尺度——/goal 把同一个循环压缩到数小时内完成多次迭代。

Compound Engineering in Product Management（2026-05-07）

来源：Marcus Moretti — A Guide to Agent-native Product Management

Marcus Moretti（Spiral 总经理）以"two-slice team"（一人全栈）身份实践 agent-native PM，将 Compound Engineering 核心原则映射到产品管理：

• Plan 占 80%，Ship 占 20%：软件开发已从"20% 计划 + 80% 执行"翻转为"80% 计划 + 20% 执行"。Agent 负责执行，PM 负责 framing 和 taste
• Agent 访谈生成策略文档：/ce-strategy 基于 Richard Rumelt《Good Strategy Bad Strategy》，通过结构化访谈生成 docs/strategy.md。Agent 会追问模糊答案——"具体是谁的情境？他们今天怎么做的，为什么不行？"
• 放弃手写工单：有了 strategy.md 后，用 ce-ideate / ce-brainstorm / ce-plan 决定做什么。Agent 负责写工单、移动状态、保持更新。PM "不再读写工单，只跟 Agent 聊它们"
• Product Pulse 作为产品记忆：每次运行 /ce:product-pulse 保存为 ~/pulse-reports/ 下的 Markdown。单个 Pulse 回答"今天发生了什么"，文件夹回答"这个月发生了什么"
• Compounding 策略：每隔几个月重跑策略访谈，Agent 会基于数周的规划、发布、数据上下文提出更尖锐的问题

这验证了 Compound Engineering 的通用性：不局限于代码，任何需要“计划 → 执行 → 回顾 → 沉淀”的知识工作都适用。

Matt Van Horn's Compound Engineering workflow (2026-06-15)

来源：AI Briefing 2026-06-15 (evening)

Peter Yang 总结了 Matt Van Horn 的 Compound Engineering 工作流：

• 用 Printing Press 将常用网站转换为 agent 可运行的 CLI
• 运行 /ce-plan 和 /ce-work 循环完成规划、编码、审查和发布
• 用针对性问题（会改动哪些文件？最可能在哪里失败？）替代阅读完整计划
• 用 last30days 跨 Reddit、X、YouTube、Hacker News 和 Polymarket 做实时研究

这显示 Compound Engineering 的循环不仅用于代码交付，也被扩展到实时信息收集和外部研究。

AI 职业教练与个人审计（2026-06-28）

来源：Every 2026-06-28

Katie Parrott 用 Codex 构建 career-coach-mode/ 文件夹系统，对第二季度目标与关键结果进行审计：

• 核心是一份 AGENTS.md 文件，由 Codex 通过访谈生成，明确权威信息位置、如何区分证据与解释、以及不同输出该放在哪里。
• Codex 在 10 分钟内审计了 4 个目标和 16 个可衡量结果，生成交互式 OKR 健康仪表盘。
• AI 指出盲区："你的危险在于试图通过数量证明影响力"，并建议她更多分享工作。
• 根据 Every 品牌定位分析，Codex 将她的角色定位为"AI 好奇者的入口匝道"，这改变了她的自我认知和工作方式。
• 最小可行版本：一个项目文件夹、五种上下文（目标、代表性工作文件夹、最近的管理者更新、结果数据文件、职位描述）、四条证据规则、一个结构化 prompt，形成 commitment → work → feedback → cited assessment → next action → new evidence 的闭环。

Austin Tedesco 的写作工作流进一步展示了 Compound Engineering 在创意生产中的应用：

• Proof 文档分三个区域：Ideas bank（粗糙想法）、Outline（正式化）、Draft（成文）。
• OpenClaw 作为收集器，Spiral 训练个人风格，Monologue 用语音突破写作障碍。
• 核心原则：AI 负责收集和组织，人类负责判断和表达。

这验证了 Compound Engineering 的通用性：从代码交付扩展到个人职业管理、创意生产和其他知识工作。

Evidence across sources

Open questions

• Does the framework apply equally to non-engineering knowledge work (design, strategy, writing), or does software's testable nature make it a special case?
• As LLMs improve at planning and even ideation, will the human role compress to only the final polish step?
• How does Compound Engineering scale when multiple human contributors are involved, each with different taste standards?

Prompts for witness

• Where in your current workflow are you still doing execution that an agent could handle? What would it take to move your attention to framing and polish?
• If you had to store one recurring mistake or insight as "compound knowledge" today, what would it be and where would you put it?
• What is the "taste" standard in your work that AI output currently fails to meet? How do you communicate that standard?

Related

• harness-engineering/folder-is-the-agent — Folder-as-agent implementation pattern used with Compound Engineering
• harness-engineering/your-harness-your-memory — Harness as passive memory vs Compound as active accumulation
• product-trends/software-soul-ai-slop — The "slop" problem that the polish step addresses
• product-trends/aaron-levie-agent-deployer-role — The organizational role that manages agent workflows