◆/_catwu_analysis
path _catwu_analysis.md

Cat Wu (@_catwu) — Persona Analysis

Cat Wu is Head of Product for Claude Code and Cowork at Anthropic, writing at claude.com and speaking on Lenny's Podcast, Every's AI & I, and O'Reilly's AI Codecon. Prior: Dagster, Scale AI, a brief stop in VC at Index, and a two-week departure to Cursor in July 2025 before returning to Anthropic — the one biographical fact about her that's remarkable, and the one she never discusses publicly. This analysis is drawn from 100 posts and 60 replies between October 2025 and April 2026, three high-engagement threads, her canonical essay "Product management on the AI exponential" (Claude blog, Mar 2026), and two long-form interviews.

The analysis is short because her corpus is narrow: she writes almost exclusively about one thing — how Claude Code ships, what's in it, and why the PM role is being rewritten underneath her. That is the real subject of this file.

I. Core Worldview & Mental Models

The platform is moving. Planning long is malpractice.

Wu's single most important belief — the one every other opinion derives from — is that the technology stack underneath a software product is no longer stable on the timescale of a roadmap. This is the thesis of her Mar 20, 2026 thread (2,029 likes) and her co-authored Claude.com essay: "The PM playbook was built on an assumption that the technology underneath your product is roughly stable. With the current pace of model progress, this is no longer true." Everything else she says is a rule extracted from that premise.

She's consistent about it across surfaces. On the Lenny podcast (Apr 23 2026) she calls strategy-first thinking a legacy PM tell and argues that "just do things" is the most important principle for AI-native companies. On the same podcast she goes further than she will in writing: build products that don't fully work yet, so you're in position when the next model closes the gap. That's a more aggressive stance than the blog's "revisit features with new models" framing, which is a retrospective sweep; the interview version is a prospective bet.

Delegate, don't pair-program

Her second load-bearing mental model — newer, first stated this clearly in the Opus 4.7 launch on Apr 16 2026 — is that the frontier model is now better treated as a junior engineer you delegate to than as a pair programmer you guide line by line: "Give Claude Code your full task context upfront: goal, constraints, acceptance criteria in the first turn. This lets Claude Code do its best work." This is a real shift from her Oct 2025 framing, when she was telling users that Sonnet 4.5 was the "strongest all-around coding model" and pushing plan-mode (hold shift+tab twice, verify the plan) as the power user's default. In six months she has moved from "inspect the plan before execution" to "write a clear brief and let it run with auto mode."

Do the simple thing

Principle four of her PM playbook — "With agentic systems, failures compound with system complexity. Find the simplest thing that works" — deserves its own attention because it cuts against her own industry. Most product writing in her space celebrates elaborate orchestration: routers, tool stacks, vector stores, retrieval pipelines. Wu's opinion, stated on Every's AI & I podcast, is that vector embeddings create maintenance burden and security exposure, and that agentic search (grep, read, reason) matches their accuracy without the baggage. She won't say this on Twitter — too many of those companies are her ecosystem — but it's visible in the product choices: Claude Code has no vector RAG layer by design.

Intellectual DNA

The thinkers she actually cites, in public, are thin: Lenny Rachitsky as a peer-mentor in the PM world, Guillermo Rauch on the markdown-for-agents idea ("markdown is way more token efficient than html," Jan 13 2026), and her CEO Dario Amodei, whom she defends emphatically ("Proud to work at Anthropic and stand up for our values," Feb 28 2026 — her highest-hitting tweet at 12,574 likes). What's absent is conspicuous: no book citations, no academic references, no older product thinkers quoted. Her intellectual world is built from her coworkers and the Anthropic-adjacent product podcast circuit. That is either discipline (she doesn't pretend depth she doesn't have) or a genuine narrowing of aperture.

Blind spots

The corpus has one real one: she almost never engages with failure cases of delegation-style coding. When users push back in reply — "I sit down 5 min and already blew 50% of my session usage wtf" under the Opus 4.7 launch, "for pro/max users, after you run claude update..." as a one-line triage — she deflects to pricing/plan answers and does not engage the underlying critique that more-autonomous agents burn more tokens for the same intent. The Twitter channel is tuned for promotion; the real discussion lives in her team's Slack and evals, which is consistent with her public philosophy but means her feed is a weak place to find her thinking about what the product still gets wrong.

II. Product Management on the AI Exponential

This is Wu's canonical domain. Her Claude.com essay (co-authored with Bihan Jiang at Decagon and Kai Xin Tai at Datadog) is the piece every other product artifact she makes refers back to. Strip the essay down and it has four rules, each restated in the Mar 20 thread:

  1. Plan in short sprints. Long roadmaps lose to high-velocity small experiments. She cites Claude Code on Desktop, the AskUserQuestion tool, and todo lists as features that started as team ideas, got prototyped, got internal users, then shipped. No PRD cycle.

  2. Encourage demos and evals over docs. "It's faster and higher fidelity than trying to convey the idea in a doc." The subtext: docs are a tax you pay when your platform is stable enough for the doc to outlast the idea. It isn't.

  3. Revisit features with new models. Every model release, audit the backlog for features the previous generation couldn't support, and strip the scaffolding from features that now don't need it. The audience made the same point back to her in the thread replies — one reader called it keeping a "model-blocked backlog" — and it's the point in her essay that landed hardest on PMs outside Anthropic.

  4. Do the simple thing. Failures compound in agent systems. Engineering workarounds that make sense today become dead weight six months from now. She echoes this directly in the Feb 11, 2026 tweet about the Claude Code guide agent doing customization setup — "don't want to read docs" is the interface.

What's missing from the essay is more interesting than what's in it. There is no discussion of what to build — only of how to build it once you've decided. The "what" problem shows up in the reply thread ("identifying what to build that will eventually become a competitive advantage... the most critical decision remains what to build in the first place") and she does not answer it. For Wu, the "what" is settled: Claude Code, shipped faster. The essay is a PM-process essay, not a product-strategy essay, and she is disciplined about staying in her lane.

The Lenny interview adds the off-timeline piece she won't put in writing: Anthropic's speed is not only a process choice, it's a mission-alignment dividend. She argues the mission clarity eliminates the friction that slows larger orgs. That's an uncomfortable claim to make publicly about competitors, so it stays in the podcast.

III. Building Claude Code: Principles from the Inside

Her posts about Claude Code itself distill into four operating principles that read as consistent over six months:

The CLI is the source of truth

"We are very focused on making sure the CLI is the most intelligent that we can make, and that's as customizable as possible" (Every podcast). Every other surface — Desktop, iOS, web, Slack, Excel, PowerPoint, Cowork — is a delivery vehicle for the CLI's capability model, not a divergent product. This is visible in the ship order: the terminal tool launched first in Feb 2025, on-web in Oct 2025, desktop in Nov 2025, mobile in Jan 2026, remote-control in Feb 2026. She explicitly teleports users back to CLI when they're on web, and she credits the CLI's extensibility model as the moat.

Dogfood aggressively, in public

Every launch announcement includes a sentence about how she personally uses it. "I've become a daily user of Claude Code on desktop" (Apr 15 2026). "I love using Claude Code on Claude mobile app to fire off ideas on the go" (Apr 1 2026). "I've been delegating a lot of my tasks to cowork. today, cowork filed a workplace ticket for a coffee spill, prepped me for upcoming meetings, looked up salesforce data via chrome tool, and helped me buy a new down jacket" (Jan 13 2026). The buy-a- down-jacket detail is the most revealing tweet in the corpus — it is how she signals that Cowork is not a coding tool, which is the entire strategic premise of the product line split.

Evals before features

"Building great evals is the key to fast iteration on AI products" (Feb 18 2026, amplifying a job posting). This is the least commented on but most consequential principle — everything she says about shipping speed is predicated on being able to measure quality mechanically. Without evals, short sprints are Russian roulette. She does not often explain this out loud, which is a minor tell: the thing she thinks is most important is the thing she promotes least, because it's not audience-facing.

Customer numbers over narrative

The Mar 2, 2026 thread (1,527 likes) summarizing how Claude Code changed engineering at Ramp, Rakuten, Brex, Wiz, Shopify, and Spotify is the cleanest example of her default persuasive mode. "Wiz migrated a 50,000-line codebase in 1 day compared to an estimated timeline of 2-3 months human months" — cite with URL, let the customer speak. Separately: "At Anthropic, we've seen a 67% increase in PRs per dev from using Claude Code" (Feb 11 2026). "Claude Code's p99 memory usage dropped by 40x in the last two weeks" (Feb 25 2026, amplifying Jarred Sumner). She prefers a verifiable number with a source over any rhetorical flourish. This is maybe her most distinctive instinct — for someone with 74k followers, she does remarkably little hand-waving.

IV. Rhetorical Style: The Launch-Thread PM

Wu's tweets are narrowly stylized. The dominant template is: headline tweet (launch or tip) → numbered threaded reply → docs link → "let us know your feedback!" She uses it for Claude Code on web, Opus 4.6, Opus 4.7, the $50 credit, fast mode, plugins, cowork, sandbox. It is the PM-launch Twitter archetype executed with discipline. The tip-format is a shorter version: "my three favorite claude code shortcuts" (Mar 13 2026, 1,206 likes) — three numbered items, a crisp verb each, no preamble.

What the audience actually rewards, ranked:

  1. Values signaling — her top tweet at 12,574 likes is the Feb 28 2026 amplification of Dario Amodei's statement on Department of War engagement. Pure team-loyalty, not product-craft.
  2. Free stuff — the $50 extra credit on Feb 7 2026 hit 5,398 likes.
  3. Major product launches — Claude Code on web (3,849), remote-control (2,616), free usage credits for web (2,367), Haiku 4.5 (2,323), the frontend-design plugin (2,211), the Bun acquisition (1,962).
  4. Intellectual content — her PM-playbook thread, the one piece of real thought leadership in the corpus, lands at 2,029 likes. It's the deepest thing she wrote and it trails the free-credit announcement by 2.6x.

There is a reading-between-the-lines take here: the audience likes her as the face of Claude Code more than they like her as a thinker. She seems to know this and has calibrated accordingly — the feed is mostly face- of-product content, with one or two thought-leadership posts per quarter.

Reply voice vs. post voice. Her reply corpus (60 replies over the same six months) is dramatically sparer than her posts. Most replies are triage one-liners: "looking into it!", "pinged the team!", "coming soon!", "we're working on it!", "ack sorry we're looking into it!". There are almost no replies where she argues, concedes, or reveals anything beyond the operational state of a feature. The one real exception is her Feb 8, 2026 reply to @SIGKITTEN explaining the fast-mode / explore-agent / Haiku interaction — a genuinely informative technical reply. Most of the time, replies are her doing visible customer-service work in public, which is a PM technique, not a conversation.

What she doesn't do. No polemic, no jokes at competitors' expense (the "Thank you for playing, Replit and Loveable" snark in her mentions is a fan's, not hers), no hot takes on industry gossip, no posts about politics outside the Anthropic-values lane, no personal life, no disagreement with a colleague in public. The register is "brand-safe product leader" and she holds it tightly.

V. Contrarian & Hidden Takes / Evolution

What she believes that cuts against her own industry

Evolution over six months

The rate of movement in her own framing is faster than most PMs ever change their mind. It's a lived example of the thesis in her essay.

The thing she will not discuss

The July 2025 departure to Cursor. Two weeks at Anysphere with Boris Cherny for "more senior roles," then back to Anthropic. Reported by Natasha Mascarenhas at The Information, widely covered at the time, and entirely absent from her Twitter output before, during, and since. Understandable — no upside to relitigating it — but worth naming, because in a corpus this focused on shipping the thing she left for two weeks, the silence is loud. It is also the only evidence in the record that she has been weighed against an outside offer, which affects how to read the team-loyalty posts ("Proud to work at Anthropic and stand up for our values"): she has in fact considered leaving, and chose to stay.

After three drinks

What Cat Wu would say at 2am that she won't say on-timeline, inferred from the interviews: she thinks most of her industry is too slow, most product-management training is obsolete, most of the PM lifecycle (research → PRD → spec → build → ship) is waste, and most of the carefully-engineered AI scaffolding people are shipping will be made redundant by the next two model releases. She will say the first of these in writing. The rest shows up only when she's on a podcast with a friendly host.

VI. Network Graph

Wu's network is unusually legible because it's bimodal: a small dense cluster inside Anthropic, and a thin ring of high-profile external amplifiers. Almost nobody between.

Inner circle — the Claude Code team she repeatedly amplifies:

She retweets and tags this group constantly. The pattern is distinctive: every external launch tweet has a teammate name attached. It's both good manners and a deliberate org-chart signal — this is a team, not a cult of Boris or me.

Outside Anthropic, warm:

External amplifiers she appreciates but doesn't converse with:

Competitors who don't appear: Cursor, Codex, Replit, Lovable, Windsurf. The one exception is the two-week Cursor episode, which she has not acknowledged publicly. This is unusual restraint given how often her team ships features that directly substitute for those products (remote-control, on-web, plugins, desktop). The silence is strategic.

What she ignores: AI-safety Twitter discourse, general tech gossip, AI-industry drama around valuations or leadership moves. She ignored the $30B Anthropic raise with a single-line thank-you (Feb 13 2026, 370 likes) — the lowest-effort possible engagement with what is, for most employees, the biggest news of the year. That restraint is a style, not an oversight.

VII. The One Essay She Keeps Rewriting

If you trace every substantive thing Cat Wu has written or said in the scrape window, they converge on one sentence: "The PM playbook was built on an assumption that the technology underneath your product is roughly stable. It no longer is."

It appears, with light variation, in:

She is, in other words, a one-idea thinker in the good sense — she has picked the right idea and is driving it into the ground before someone else does. The idea is that capability drift is now the primary constraint on product planning, and every corollary (sprint short, demo over doc, simplify, dogfood, ship-then-revisit) is a logical consequence.

The test for whether she has more to say is ahead of her: when model progress plateaus — which her framing implicitly predicts will not happen soon — does she have a second idea? Her corpus does not yet show one. She does not write about retention, monetization strategy, organizational design, pricing, or hiring in any depth. What she writes about, repeatedly, is how to ship AI products when the floor is moving. On the evidence, that is because that is where she has conviction, and she is disciplined about not faking depth she doesn't have. It is also the single most useful thing a product leader in her position could spend her public voice on right now.

Reading curriculum implied by the corpus

If you wanted to put Cat Wu's views in a reader's hands in a single afternoon:

  1. "Product management on the AI exponential" (Claude blog, Mar 2026) — the essay, co-authored with Jiang and Tai. The thesis document.
  2. Her Mar 20 2026 Twitter thread — the compressed version, with the replies under it as the distributed commentary.
  3. The Lenny's Podcast episode (Apr 23 2026) — the interview where she says things she won't put in writing, including the "build ahead of the model" stance and the mission-alignment claim.
  4. The Every AI & I episode with Boris Cherny and Dan Shipper — the technical and product-philosophy deep end, including her contrarian take on vector RAG.
  5. The Mar 2 2026 customer-adoption thread (Ramp, Rakuten, Brex, Wiz, Shopify, Spotify) — the numbers-over-narrative performance that is her signature public-facing mode.

That's roughly two hours. It is more or less the entire persona. The rest of the corpus is shipping notes.