資料來源#
摘要#
Anthropic 旗下 Claude Code 與 Cowork 的產品負責人。多年擔任工程師，後短暫進入 VC 領域，再加入 Anthropic。訪談過「數百位試圖進入 AI 領域的 PM」；將這份觀察轉化為一個強烈觀點：PM 角色正處於重構之中（參見 Engineer PM Convergence）。
重要主張與立場#
- 節奏轉變。 Anthropic 的功能交付週期從 6 個月 → 1 個月 → 有時 1 天。透過移除流程阻力、將多數發布冠上 research preview 標籤以降低承諾、先在內部上線再對外推出來達成。
- 使命 > 協調。「如果有兩個互相競爭的優先順序，我們會討論哪一個對 Anthropic 的使命更重要。」這才是在規模化下消除摩擦的關鍵，而不是人頭數或流程。
- 就去做事。 個人生活信條。「工作是假的。如果你理解了限制條件，就能想清楚自己能做什麼，然後盡快去做。」她認為主動性（agency）是新創公司應該招募的稀缺特質。
- 為當下的模型打造。 為超級 AGI 稻草人打造產品很容易；要從今天的模型中引出最大能力卻很難。最困難的 PM 技能，是定義一個月後產品該長什麼樣。
- 請模型自我檢視。 被低估的除錯技巧：當 Claude 做出意料之外的事，問它為什麼。「很多時候，只要保持高度好奇心去探究模型為何做出那個決定，你就會看到是什麼把它誤導了，於是你就能修正 harness。」（參見 Model Introspection Feedback）
- 建立 evals。 十個優秀的 evals 勝過一百個平庸的；撰寫 eval 是「被低估的工作」，應該有更多 PM／工程師動手做。
- 95% 自動化不算自動化。 不衝到 100% 就別費心。最後那 5% 占了大部分工作量，卻是讓工作流值得信賴的關鍵。
- 打造你每天會用的應用，而不是原型。 客製化設定超過某個程度就會變成干擾——「我覺得簡單的設定其實效果更好。」
- 性格就是產品。 Claude 的個性（低自我、正向、輕鬆、偏向行動）是產品成功的核心——Amanda 塑造模型性格的工作「比寫程式更難」，因為這項任務本身極度模糊。
訪談中的運作細節#
- Anthropic 內部約有 30–40 位 PM，分布於 research-PM、Claude Developer Platform、Claude Code、Enterprise、Growth 等團隊。
- 寧可招募具備強烈產品品味的工程師，也不要工程能力薄弱的 PM——團隊中許多工程師能直接把 Twitter 上的反饋一路打造成上線產品，全程無需 PM 介入。
- 團隊中的設計師都具備前端工程背景。
- 每晚使用 Cowork，從 Slack／Drive／Twitter 的脈絡草擬 20 頁簡報（已預載 anthropic design system）。
- 內部技術堆疊：大量使用 Claude Code + Cowork、Slack（「公司的作業系統」）、由團隊為個人化工作流自製的內部應用。
- 「Applied AI」團隊是 token 消耗量第二大的單位（僅次於工程團隊）——這是個技術型 go-to-market 的角色，為客戶製作原型。
重要引述#
- 「為超級 AGI 強模型打造產品非常容易。困難的是針對當下的模型，弄清楚該如何引出最大能力？」
- 「當寫程式的成本變得便宜許多時，真正變得更有價值的事情，是決定要寫什麼。」
- 「即使某個產品不成功，只要它沒有擋住核心使用場景，那就無妨。」
- 「每當模型變得更聰明，我們就能移除許多 prompting 介入。事實上，我們每次發布新模型時都會這麼做。」
相關連結#
- Claude Code — 共同領導產品
- Cowork — 共同領導產品
- Boris Cherny — 技術主管搭檔
- Anthropic — 雇主
- Engineer PM Convergence — 主要倡議者
- Harness Shrinkage as Models Improve — Opus 4 上移除待辦清單拐杖是她的經典範例
- Model Introspection Feedback — 她命名的除錯技巧
- AI Native Product Cadence — 體現了相關實踐
- Claude Character as Product — 闡述為何性格是承重元素
- Fiona Fung — 同樣產品上的互補資深領導者（Fung 統整工程組織；Wu 擔任產品負責人）
- Dogfooding as Product Discipline — 她午餐時的 vibe-check，是 Fiona Fung 稱之為「螞蟻食」的那種 dogfooding 在 eval 形式上的對應
- Verification as the New Bottleneck — 她「十個優秀的 evals」／「衝到 100%」的立場，是「驗證成為瓶頸」這個現象在產品側的呈現
資料來源#
- How Anthropic's product team moves faster than anyone else | Cat Wu (Head of Product, Claude Code) — Lenny's Podcast, 2026-04-23
Cited by 40
- Learning to Co-Work with AI: A Software Engineer's Field Guide×6
The most underrated technique Cat Wu names: when the agent does something wrong, ask it why (Model…
- Opinions on Using AI Tools & the Future of the Software Engineering Role×4
The harness should shrink, not grow. Harness Shrinkage As Models Improve: every model release lets…
- Build for the Next Model×4
A product-strategy corollary of Harness Shrinkage As Models Improve, now stated independently by…
- Harness Shrinkage as Models Improve×4
The harness — prompts, skills, scaffolding, mechanical verification — exists to compensate for what…
- Dogfooding as Product Discipline×3
The cross-source claim that product sense is built, not innate — and the way you build it is…
- Engineer PM Convergence×3
Both Boris Cherny (Sequoia AI Ascent 2026) and Cat Wu (Lenny's Podcast, April 2026) report the same…
- Evals as Product Spec×3
Cat Wu's articulation of why writing evals is the emerging core PM skill for AI products — not a QA…
- Fiona Fung×3
> Note: both Fiona Fung and Cat Wu are described as leading product for Claude Code + Cowork. They…
- AI Native Product Cadence×2
Cat Wu's account of how Anthropic ships at a pace that surprises observers. Cycle time per product…
- Anthropic×2
AI safety company / vendor of Claude; mission-as-tiebreaker culture; ~30–40 PMs across teams; Mike Krieger leads Labs r…
- Claude Character as Product×2
Cat Wu argues that Claude's character — low-ego, lighthearted, positive, bias-toward-action,…
- Compounding Loop Optimization×2
Where is the line between worthwhile internal tooling and yak-shaving? Carey's "afternoon" bar is…
- Cowork×2
Anthropic's knowledge-work agent product, sibling to Claude Code. Where Claude Code targets work…
- HTML as the New Markdown×2
At first glance this contradicts the wiki's running Harness Shrinkage As Models Improve thesis (Cat…
- Instruction Compounding×2
Harness Shrinkage As Models Improve says scaffolding becomes unnecessary as capability migrates…
- MCP and Computer Use×2
Cat Wu's nightly slide-deck workflow (Cowork) explicitly uses MCP — Figma MCP, Slack MCP, Drive MCP…
- Model Introspection Feedback×2
Cat Wu names this as her single most underrated AI debugging technique: when Claude does something…
- Thariq Shihipar×2
Unhobbling. (July 2026 context-engineering post.) The Claude Code team was over-constraining the…
- The Three Loops of AI-Native Building×2
This is Engineer Pm Convergence from a third independent vantage (after Cat Wu and Boris Cherny),…
- Andrew Ambrosino
Cat Wu / Boris Cherny — Anthropic counterparts on role convergence; Ambrosino is the OpenAI-side…
- Andrew Ng
Engineer Pm Convergence — third independent vantage on the same role merge, after Cat Wu and Boris…
- Anthropic Labs
Ai Native Product Cadence — the bet-factory cadence is the same ship-fast/research-preview pattern…
- Authority and Audit Survive Abundance
What would change this answer. Cat Wu's prediction that "all the safety mechanisms today — prompt…
- Boris Cherny
Creator of Claude Code at Anthropic; phone-driven workflow with hundreds of agents; primary advocate of `/loop` primiti…
- Claude Code
Anthropic's agentic coding product; created by Boris Cherny late 2024; TypeScript/React on Bun (itself Claude-rewritten…
- Compute Allocator
Product taste as the bottleneck skill (Engineer Pm Convergence, Cat Wu: "as code becomes much…
- Cost-per-Task Over Cost-per-Token
simonwillison judgement — Simon Willison, "Fable's judgement," 2026-07-03 (practitioner-opinion,…
- Dan Carey
Cat Wu / Boris Cherny / Fiona Fung — fellow Anthropic product/eng leaders whose AI-native-cadence…
- How Do You Write Evals for Taste? Character as the Limit Case
Cat Wu holds two claims that look contradictory. First, character is the hardest thing to evaluate…
- Excellence as an Operating System
Selflessness. The tiebreaker is "Netflix members matter," not personal success — the same…
- Harness Build-vs-Buy
Churn is not size, and shrinkage produces churn. Cat Wu's read-the-whole-prompt-at-every-launch…
- Jagged Intelligence (Ghosts, Not Animals)
Model Introspection Feedback — Cat Wu's "ask the model why it failed" presumes a ghost whose…
- Entities — People, Orgs, Tools & Projects
Cat Wu — Head of Product for Claude Code and Cowork at Anthropic; primary articulator of AI-native…
- Mythos Model
Cat Wu: "Mythos is an incredibly powerful model. But we do use the models internally and I think…
- The PRD-Replacement Spectrum at AI-Native Speed
Lighter PRD · Ai Native Product Cadence (Cat Wu) · Calibrated to ambiguity — 1-pager ↔ full PRD ·…
- Prototype Over PRD
Lighter PRDs · Ai Native Product Cadence (Cat Wu) · 1-pagers for ambiguous features; full PRD only…
- Shared Harness, Differentiated Surfaces
Anthropic answered two: Claude Code for work whose output is code, Cowork for work whose output…
- Telemetry vs. Survey Measurement
Evals As Product Spec — Cat Wu's evals encode the spec; telemetry encodes what actually shipped —…
- The Verifiability Thesis
Evals As Product Spec — Cat Wu's "ten great evals" is the product-side mirror: encoding what…
- Verification as the New Bottleneck
Evals As Product Spec — Cat Wu's evals are verification encoded as product spec; the PM-side…
Related articles
- Harness Shrinkage as Models Improve
Prompt scaffolding shrinks each model release; Cat Wu's pruning discipline; Boris Cherny "100 lines of code a year from…
- Boris Cherny
Creator of Claude Code at Anthropic; phone-driven workflow with hundreds of agents; primary advocate of `/loop` primiti…
- Claude Code
Anthropic's agentic coding product; created by Boris Cherny late 2024; TypeScript/React on Bun (itself Claude-rewritten…
- Anthropic
AI safety company / vendor of Claude; mission-as-tiebreaker culture; ~30–40 PMs across teams; Mike Krieger leads Labs r…
- Evals as Product Spec
Cat Wu's framing of evals as the emerging core PM skill: ten great evals beats a hundred mediocre; encode what done loo…