Most important take away

The defining skill of the modern AI PM is shrinking the time from idea to user-facing feature — Anthropic has compressed timelines from six months to one week (sometimes one day) by ruthlessly removing process, shipping in branded “research previews,” and hiring engineers and designers with strong product taste so roles overlap. Product taste — knowing what to build and how to expose model strengths while patching weaknesses — is the durable, scarce skill as code becomes cheap to write.

Summary

Key Themes

• Speed is the product strategy. Anthropic deliberately removes every barrier to shipping. Most Claude Code features ship as clearly branded “research previews” so the team isn’t locked into long-term commitments and can iterate on real user feedback within days.
• Roles are merging. Engineers do PM work, PMs ship code, designers were front-end engineers. Cat’s team prefers hiring engineers with strong product taste over adding more PMs, because it cuts coordination overhead. Product taste is the rarest, most durable skill.
• Mission over org. Anthropic’s safety-AGI mission sits above any individual product line, which makes hard trade-off decisions (e.g., the OpenClaude/third-party reseller decision, deprioritizing Claude Code work for higher-priority Anthropic goals) fast and unifying.
• Build for the model that’s coming, then strip back as it arrives. Many Claude Code “features” (like the original to-do list) were crutches for weaker models. With each new model, the team reads the full system prompt and removes scaffolding the model no longer needs. Conversely, they prototype features that don’t quite work yet (e.g., code review) so they’re ready to swap in the new model the moment it’s capable.
• Action-based AI is the inflection. The 2024 chat-based products underwhelmed many users; agent products that do things (Claude Code, the Chrome extension) are what convert skeptics.
• Calmness as a hiring filter. The team explicitly hires people who lean into chaos with a smile, brutally prioritize, and accept shipping imperfect features knowing fast feedback will fix them.

Actionable Insights

• Set crisp goals to compress decision-making. A clear statement like “help professional developers safely reach zero permission prompts” rules out entire approach categories and unblocks the team.
• Build a repeatable shipping pipeline. Anthropic uses an “evergreen launch room” Slack channel where engineers post ready features; docs and PMM turn around marketing the next day. The PM’s job is setting up this scaffolding so engineers don’t get blocked.
• Run weekly metrics readouts so everyone deeply understands the business; pair with a written list of team principles (key users, what you’ll trade off) so people can make decisions without escalating.
• Skip PRDs by default, but write one-pagers for ambiguous features and full PRDs for multi-month infrastructure work.
• Talk to the model. Ask it to introspect on why it made a wrong decision — it often surfaces the system-prompt or harness gap directly.
• Find your ~5 trusted human evaluators. Most user feedback is noise; a small group of articulate testers gives the fastest signal on a new model or harness.
• Build 10 great evals, not 100. Evals turn fuzzy quality goals into measurable progress and are an underused PM skill.
• Get automations to 100%, not 95%. A 95% automation isn’t an automation; the last 5% is where real leverage lives. Put in the elbow grease to teach Claude your preferences.
• Build the prototype that doesn’t work yet so you know exactly what’s missing — then swap in the next model.

Career Advice

• Develop product taste above all. As writing code gets cheap, deciding what to write and how to shape the UX becomes the scarce, valuable skill. It can come from any background.
• An engineering background is currently a multiplier because it helps you estimate cost-of-build during prioritization — but Cat expects which adjacent skill matters most to keep shifting every few months.
• Be low-ego about role boundaries. “Just do things” — first-principles thinking plus cross-functional execution is what wins; jobs are largely fake constructs.
• Build apps you actually use daily. Prototype tinkering teaches little; real usage exposes real gaps and gives real leverage.
• Don’t over-customize your setup. There’s a camp that obsesses over skills/MCPs/workflows at the expense of shipping. Simple setups often work better.
• Lean into chaos with calm. Brutal prioritization, willingness to ship imperfect work, and protecting sleep are how people sustain pace at Anthropic.
• Spend time at a sub-50-person startup once. It builds the muscle of operating without role boundaries.
• For automating your work: find the repetitive tedious parts of your job, automate them with Claude, then redeploy the freed 20% on the pet projects no one had bandwidth for.

Product Strategy Notes

• Product surface map: Claude Code CLI gets new features first and is most powerful; Claude Desktop is best for front-end work (live preview pane) and gives non-technical users a friendlier surface plus a control pane for all sessions; web/mobile is for kicking off tasks on the go; Co-work is for non-code outputs (decks, docs, inbox-zero, customer briefings).
• Co-work onboarding: connect Slack, Gmail, Google Calendar, Google Drive, Figma MCP — quality scales with context access.
• Internal tooling explosion: Claude Code has lowered the bar so much that teams build personalized internal apps (e.g., a sales deck generator pulling from Salesforce + Gong + customer notes) instead of forcing fit on generic tools.
• Roadmap arc: task success → multi-task in parallel → 50–100 remote tasks at a time → infra and interfaces for verifying agent work and self-improving from feedback.
• What you sacrifice for speed: product consistency. Overlapping features and a heavier user-education burden are accepted costs of shipping fast.
• Claude’s character matters. Low-ego, positive, action-biased personality (Amanda’s work) is core to why users prefer Claude — it makes Claude a great co-worker, not just a competent model.

Chapter Summaries

Cat’s role and partnership with Boris. Cat handles the path from today to the 3–6 month vision Boris sets, plus all cross-functional alignment (marketing, sales, finance, capacity) and unblocking shipment. ~80% mind-meld with Boris; the remaining 20% they each drive based on who cares more.

Why most PM candidates are approaching AI wrong. Old PM skill was multi-quarter roadmap coordination. New skill is compressing idea-to-user time to days. Stop optimizing for cross-team alignment over multi-quarter plans; start building “concept corner” workflows where a feature ships within the week.

How Claude Code ships so fast. Three pillars: clear goals that rule out approaches, a repeatable research-preview shipping process branded so users know what they’re getting, and an engineering→docs→PMM “evergreen launch room” Slack pipeline. PRDs only for ambiguous or infrastructure-heavy work; otherwise rigorous weekly metrics + written team principles.

The leak and the OpenClaude decision. The Claude Code source leak was human error in a release-package PR that passed two layers of review; processes have been hardened. OpenClaude restrictions came from prioritizing first-party products and the API given compute scarcity — a mission-driven trade-off.

Inside the PM org. Roughly 30–40 PMs across teams: Research PM (Diane), Claude Developer Platform, Claude Code (Code + Co-work), Enterprise (cost controls, RBAC, security), and Growth.

The future of the PM role. Roles are blurring — engineers, PMs, designers all overlap. Anthropic favors engineers with product taste because it minimizes coordination overhead. Cat and most Claude Code PMs were engineers first. Product taste is the durable scarce skill; engineering background is a near-term advantage for cost estimation. Be low-ego, swap hats, learn first-principles thinking.

Staying sane at Anthropic’s pace. Hire people who smile at chaos, ship imperfect things knowing feedback fixes them, brutally prioritize, protect sleep. The “Pirates of the Caribbean staircase” energy is the cultural norm.

What you sacrifice for speed. Product consistency suffers — overlapping features, education debt, users feel a fast treadmill. Mitigation: built-in onboarding like /power-up, even though originally the team resisted any tutorial.

Why Anthropic is winning from behind. Two ingredients: (1) a unifying mission that makes cross-org trade-offs fast and accepted, and (2) team-level willingness to sacrifice their own KRs for Anthropic’s mission.

Claude Code vs. Desktop vs. Web/Mobile vs. Co-work. CLI = power user / first to get features. Desktop = front-end work + control pane + non-technical friendly. Web/mobile = on-the-go kickoff. Co-work = non-code outputs (decks, docs, inbox, customer briefings). Cat walked through how she generated a 20-page conference deck overnight by feeding Co-work the Slack/Drive/PMM context and letting it research and assemble using Anthropic’s internal design template.

PM tool stack at Anthropic. Heavily Claude Code + Co-work + Slack (“the OS of Anthropic”). 30% of Cat’s time is pushing the model’s limits to feel its weaknesses. Internal teams build their own custom apps (e.g., the sales deck generator) rather than adopting generic tools.

Token spend and the Applied AI team. Applied AI is the second-biggest token consumer after engineering — they prototype for customers and use Co-work for pre-meeting customer dossiers. Per-engineer token spend rises with every model jump but is still well below average engineer salary.

Emerging skills for AI PMs. Define what the product should look like in a month under model uncertainty. Be “the right amount of AGI-pilled” — not building only for the super-strong future model, but eliciting maximum capability from today’s model. Build skill via: heavy model usage + asking the model to introspect on mistakes + finding ~5 trusted human evaluators + writing ~10 great evals.

Claude’s character and the value of personality. Amanda Askell’s work crafting Claude’s low-ego, positive, earnest, action-oriented character is core to why users love it — personality is a product feature, not decoration.

Removing scaffolding as models improve. With every new model, the team rereads the full system prompt and removes prompts the model no longer needs (e.g., the to-do list nudge is now obsolete). New capabilities (like reliable code review on Opus 4.5/4.6 + Sonnet 4.6) unlock features that previously sat as failed prototypes.

Vision for Claude Code and Co-work. Building blocks: reliable single tasks → many parallel tasks → 50–100 remote tasks. Required: remote infra, interfaces showing humans which tasks need attention, agents that fully self-verify, and a self-improving feedback loop.

Career advice for thriving in AI. Automate your tedious work to free time for creative work and pet projects. Push automations to 100% reliability — 95% isn’t automation. Build apps you actually use daily; don’t over-customize at the expense of shipping. The chat-vs-action divide is why some users dismiss AI: action-based products like Claude Code are the eye-opener.

Lightning round. Books: How Asia Works, The Technology Trap, Paper Menagerie. Shows: Drive to Survive, Free Solo. Product: Waymo (twice daily commute). Motto: “just do things” / first-principles thinking / “jobs are fake.” Favorite Claude thinking word: “manifesting.” Post-AGI plan: rock climb in Fontainebleau, read 1–2 books a week. Best way to help her team: report reproducible edge cases and failures of Claude Code / Co-work on Twitter (@sirenacatwoo).