LLenny's Podcast
← All episodes
Cat Wu (Head of Product, Claude Code)23 April 2026

How Anthropic’s product team moves faster than anyone else

7Frameworks
15Insights

Episode overview

Cat Wu explains how Anthropic compresses product timelines by setting clear goals, reducing launch friction, releasing features in research preview, and empowering engineers to carry ideas through to users. She argues that AI-era product management increasingly depends on product taste, first-principles thinking, hands-on model use, carefully chosen feedback, and lightweight evals. Wu also describes how Claude Code and Co-work divide coding and non-coding work, how stronger models remove old harness interventions, and why useful automations should be pushed beyond prototypes into reliable daily workflows.

Key ideas

  • AI-native teams should shorten the path from an idea to a product in users' hands, often aiming to ship within a week.
  • Clear goals, shared metrics, team principles, research previews, and a repeatable cross-functional launch process let teams move quickly with less coordination overhead.
  • As engineering becomes cheaper, product taste—choosing what to build and what a delightful experience should be—becomes more valuable.
  • Strong AI product managers spend substantial time using models, investigate unexpected behavior, identify trusted evaluators, and build a small number of meaningful evals.
  • Claude Code is positioned for code-producing tasks, while Claude Co-work is positioned for non-code outputs such as research, documents, inbox workflows, and slide decks.
  • New model generations can eliminate product features and prompts that previously acted as capability crutches while unlocking entirely new products such as reliable code review.
  • The long-term interface must help people supervise many remote agents, verify completed work quickly, and carry corrective feedback into future runs.
  • People should automate repetitive work, refine important automations until they are dependable, and build tools they actually use rather than stopping at prototypes.

Transcript available · source text is retained privately and is not published

Frameworks in this episode

Strategy4 steps

Build for the Current Model, Prototype for the Next

Elicit max capability from today's model while pre-building the products the next model will unlock

Read →
Strategy4 steps

Clear-Goal Ambiguity Cut

Because general LLMs can do anything, a sharp key-user + problem + use-case triad is what rules approaches out

Read →
Leadership3 steps

Mission-Above-Product Prioritization

Route every cross-org tradeoff through a single shared mission so decisions are fast and everyone stands behind them

Read →
Innovation4 steps

Model Introspection Harness Repair

When an AI agent misbehaves, ask it why — its explanation reveals the harness gap to fix

Read →
Innovation4 steps

Ship-in-Research-Preview Speed Loop

Cut idea-to-user time from months to a week by branding launches as previews and pre-wiring the launch chain

Read →
Productivity4 steps

The 100%-or-Nothing Automation Rule

An automation that works 95% of the time isn't an automation — push it to 100% or don't rely on it

Read →
Productivity3 steps

Three-Legged Model Evaluation

Judge a model-harness combo with heavy usage, a trusted five-person taste panel, and ~10 sharp evals

Read →

People & resources mentioned

Attributed to the moment in the episode. Timestamps are approximate.

People · 15

  • Cat WuMentionsToday my guest is Kat Wu, head of product for Cloud Coding who work at Anthropic.

    “Today my guest is Kat Wu, head of product for Cloud Coding who work at Anthropic.”

    ▶ Watch at ~1:00approx.
  • Boris ChernyCoinedShe oversees both Boris Cherney and Kat Wu, both of whom who have been on the podcast

    “He uh created Cloud Code. He leads the eng team”

    ▶ Watch at ~1:30approx.
  • Cat WuMentionsToday my guest is Kat Wu, head of product for Cloud Coding who work at Anthropic.

    “Kat, welcome to the podcast.”

    ▶ Watch at ~1:30approx.
  • Lenny RachitskyMentionstoday we've got another very special compilation episode something I've been pulling on more and more with the podcast and the newsletter

    “Kat, welcome to the podcast.”

    ▶ Watch at ~1:30approx.
  • Boris ChernyMentionsShe oversees both Boris Cherney and Kat Wu, both of whom who have been on the podcast

    “I feel very lucky to work with Boris. He's been an amazing thought partner. He's our tech lead, he's very much the product visionary”

    ▶ Watch at ~2:00approx.
  • LydiaMentionsTarek and Lydia on DevRel

    “Sarah, who leads our docs, and Alex, who leads PMM, and Tarek and Lydia on DevRel just like jump in”

    ▶ Watch at ~8:30approx.
  • TarekMentionsTarek and Lydia on DevRel

    “Sarah, who leads our docs, and Alex, who leads PMM, and Tarek and Lydia on DevRel just like jump in”

    ▶ Watch at ~8:30approx.
  • AlexMentionsAlex, who leads PMM

    “Sarah, who leads our docs, and Alex, who leads PMM, and Tarek and Lydia on DevRel just like jump in”

    ▶ Watch at ~8:30approx.
  • SarahMentionsSarah, who leads our docs

    “Sarah, who leads our docs, and Alex, who leads PMM, and Tarek and Lydia on DevRel just like jump in”

    ▶ Watch at ~8:30approx.
  • DianeMentionsthe research PM team uh who Diane leads

    “we have the research PM team uh who Diane leads”

    ▶ Watch at ~14:30approx.
  • AmolMentionsAmol, the head of growth at growth at Anthropic

    “So Amol was just on the podcast. He had this really interesting insight that most people haven't been sharing.”

    ▶ Watch at ~15:30approx.
  • AmandaMentionsAmanda who def who molds Claude's character

    “one Amanda who def who molds Claude's character. It's just like such a hard role because the task is so ambiguous.”

    ▶ Watch at ~57:00approx.
  • Ben MannMentionsSo, I talked to Ben Mann, co-founder of Anthropic.

    “I had Ben Mann on the podcast co-founder and he talked about this just like the character, the constitution of Claude”

    ▶ Watch at ~58:30approx.
  • Andrej KarpathyMentionsI remember Karpathy tweeted that he just like has never seen a model like this.

    “There's this Karpathy tweet that just uh came out on yesterday where he talked about this divide that's interesting”

    ▶ Watch at ~1:14:00approx.
  • Alex HonnoldMentionsabout this guy Dean Potter who he was like Alex Honnold before Alex Honnold was Alex Honnold.

    “I also really love Free Solo, which is about Alex Honnold um climbing El Capitan without a harness.”

    ▶ Watch at ~1:16:30approx.

Resources · 48

  • Lenny's NewsletterCoinednewsletter · Lenny Rachitsky

    “available exclusively to Lenny's newsletter subscribers.”

    ▶ Watch at ~1:00approx.
  • Lenny's Product PassRecommendswebsite · Lenny Rachitsky

    “don't forget to check out Lenny's product pass.com for an insane set of deals available exclusively to Lenny's newsletter subscribers.”

    ▶ Watch at ~1:00approx.
  • WorkOSRecommendscompany · WorkOS, Inc.

    “Visit workos.com to get started or just hit up their Slack where they have actual engineers waiting to answer your questions.”

    ▶ Watch at ~4:00approx.
  • AnthropicMentionscompany

    “I've never seen anything like the pace folks at Anthropic are shipping at.”

    ▶ Watch at ~10:30approx.
  • MythosUsessoftware · Anthropic

    “Mythos is an incredibly powerful model. We do use the models internally”

    ▶ Watch at ~11:00approx.
  • OpenClawMentionssoftware · Peter Steinberger

    “Uh another question I had is open claw.”

    ▶ Watch at ~12:30approx.
  • Managed AgentsMentionsproduct · Anthropic

    “They also release things like managed agents, which is a way for you to build your agents and we can host it”

    ▶ Watch at ~14:30approx.
  • Claude Developer PlatformMentionsproduct · Anthropic

    “there is the Claude developer platform team that maintains the APIs that Claude Code is built on top of.”

    ▶ Watch at ~14:30approx.
  • OpenAIMentionscompany

    “Open AI was way ahead. It was just like no way Anthropic has any chance to compete significantly long term.”

    ▶ Watch at ~29:00approx.
  • Claude CodeUsessoftware · Anthropic

    “So, I tend to use uh Claude Code in the terminal when I'm just kicking off like a one-off coding task”

    ▶ Watch at ~32:30approx.
  • Claude CodeRecommendssoftware · Anthropic

    “I would highly recommend checking out Claude Code on desktop.”

    ▶ Watch at ~33:30approx.
  • Claude Co-workUsessoftware · Anthropic

    “And if the output is anything that's not code, I'll use co-work for it.”

    ▶ Watch at ~35:30approx.
  • GmailUsessoftware · Google

    “I connect it to my Google Calendar, I connected to my Slack, to my Gmail, to my Google Drive”

    ▶ Watch at ~36:00approx.
  • Claude Co-workRecommendssoftware · Anthropic

    “If you're getting started on co-work, the first thing that you really need to do is connect all the data sources”

    ▶ Watch at ~36:00approx.
  • Google CalendarUsessoftware · Google

    “I connect it to my Google Calendar, I connected to my Slack, to my Gmail, to my Google Drive”

    ▶ Watch at ~36:00approx.
  • Google DriveUsessoftware · Google

    “I connect it to my Google Calendar, I connected to my Slack, to my Gmail, to my Google Drive”

    ▶ Watch at ~36:00approx.
  • Code with ClaudeMentionsevent · Anthropic

    “we have this Code with Cloud conference coming up and there's a few talks that I'm giving there.”

    ▶ Watch at ~36:30approx.
  • X (formerly Twitter)Useswebsite

    “It walked through Twitter to see what we launched, it looked through our evergreen launch room”

    ▶ Watch at ~37:30approx.
  • SlackRecommendssoftware · Slack Technologies

    “Slack, Google Calendar, Gmail, G Drive. You should connect your communications tools and where you store your source of truth data”

    ▶ Watch at ~39:00approx.
  • FigmaUsessoftware · Figma

    “You can also connect like your Figma MCP if you if you have your slide format um saved there”

    ▶ Watch at ~41:30approx.
  • SlackUsessoftware · Slack Technologies

    “So, my stack is pretty heavily Claude Code, co-work, and Slack. Anthropic largely runs on Slack.”

    ▶ Watch at ~42:00approx.
  • SalesforceUsescompany · Salesforce

    “he has a way to input specific customer context that pulls from Salesforce, that pulls from Gong, that pulls from other notes.”

    ▶ Watch at ~43:30approx.
  • GongUsessoftware · Gong

    “he has a way to input specific customer context that pulls from Salesforce, that pulls from Gong, that pulls from other notes.”

    ▶ Watch at ~43:30approx.
  • Amazon BedrockMentionssoftware · Amazon Web Services

    “this customer is using like Bedrock or Cogram for Enterprise or Console.”

    ▶ Watch at ~43:30approx.
  • Google Cloud Vertex AIMentionssoftware · Google Cloud

    “if this is a customer that's on Vertex or Bedrock and doesn't want to use Cloud for Enterprise”

    ▶ Watch at ~44:00approx.
  • VantaRecommendssoftware · Christina Cacioppo

    “Learn more at vanta.com/lenny. And as a listener of this podcast, you get $1,000 off Vanta.”

    ▶ Watch at ~46:30approx.
  • ClaudeUsessoftware · Anthropic

    “Like it's kind of very easy to build the product for the super AGI uh strong model.”

    ▶ Watch at ~52:30approx.
  • ClaudeMentionssoftware · Anthropic

    “People really like that Claude's low ego.”

    ▶ Watch at ~1:00:00approx.
  • VS CodeMentionssoftware · Microsoft

    “Similar to how in VS Code you would look up all the call sites and it'll be a list on the left side”

    ▶ Watch at ~1:01:30approx.
  • Claude Opus 4Mentionssoftware · Anthropic

    “But then with Opus 4 and later models, we realized that we didn't need to force it to use this to-do list.”

    ▶ Watch at ~1:02:00approx.
  • Claude Code Code ReviewUsessoftware · Anthropic

    “our engineering team relies on this code review to pass before we merge PRs.”

    ▶ Watch at ~1:03:30approx.
  • Claude Opus 4.6Usessoftware · Anthropic

    “And it was only with like Opus 4 5 and 4 6 that we”

    ▶ Watch at ~1:04:00approx.
  • Claude Opus 4.5Usessoftware · Anthropic

    “And it was only with like Opus 4 5 and 4 6 that we”

    ▶ Watch at ~1:04:00approx.
  • Claude Sonnet 4.6Usessoftware · Anthropic

    “and uh Sonic 4 6 that we felt like, "Okay, we are now able to like run multiple code review agents simultaneously”

    ▶ Watch at ~1:04:00approx.
  • ChatGPTMentionssoftware · OpenAI

    “this divide that's interesting between people that tried chat GPT cloud back in the day”

    ▶ Watch at ~1:14:00approx.
  • Claude for ChromeRecommendssoftware · Anthropic

    “Shout out Chrome extension, the Claude Chrome extension which you can just watch it doing stuff.”

    ▶ Watch at ~1:15:00approx.
  • How Asia WorksRecommendsbook

    “I really like how Asia works. It's a story about economic development”

    ▶ Watch at ~1:15:30approx.
  • The Technology TrapRecommendsbook

    “The other books that I'm really into are the technology trap.”

    ▶ Watch at ~1:15:30approx.
  • The Paper Menagerie and Other StoriesRecommendsbook

    “And maybe on like a fun note, I really like Paper Menagerie.”

    ▶ Watch at ~1:16:00approx.
  • Formula 1: Drive to SurviveRecommendstv show · Netflix

    “I really like Drive to Survive.”

    ▶ Watch at ~1:16:30approx.
  • Free SoloRecommendsdocumentary

    “I also really love Free Solo, which is about Alex Honnold um climbing El Capitan without a harness.”

    ▶ Watch at ~1:16:30approx.
  • El CapitanMentionsplace

    “Free Solo, which is about Alex Honnold um climbing El Capitan without a harness.”

    ▶ Watch at ~1:16:30approx.
  • WaymoUsescompany · Alphabet

    “Um use it twice a day, get to and from work.”

    ▶ Watch at ~1:18:00approx.
  • WaymoRecommendscompany · Alphabet

    “The product that has like most changed my life outside of cloud products is probably Waymo. Like I'm a die-hard Waymo user.”

    ▶ Watch at ~1:18:00approx.
  • UberMentionscompany · Uber

    “I always thought Waymo needed to be priced lower than Uber and Lyft to succeed”

    ▶ Watch at ~1:19:00approx.
  • LyftMentionscompany

    “I always thought Waymo needed to be priced lower than Uber and Lyft to succeed”

    ▶ Watch at ~1:19:00approx.
  • FontainebleauMentionsplace

    “I'll probably move to like Fontainebleau and just like live amongst 10,000 boulders and climb for a bit.”

    ▶ Watch at ~1:22:30approx.
  • Lenny's PodcastCoinedpodcast · Lenny Rachitsky

    “You can find all past episodes or learn more about the show at lennyspodcast.com.”

    ▶ Watch at ~1:25:00approx.

Spot an error or want something removed? Request a correction or removal.

Insights & moments

The myth-busts, hot takes, explainers, and tools worth keeping.

Myth Buster· 1

Myth Buster10:30

It's Not a Secret Model Making Anthropic Fast

Asked whether a powerful unreleased internal model explains Anthropic's shipping pace, Kat pushes back. The team has moved fast for several quarters, and while internal models help a little, the real driver is a low-process culture that removes every barrier to shipping and empowers anyone to go from idea to launch in under a week.

  • The team has been moving fast for several quarters, so a new internal model doesn't explain the bulk of it.
  • Internal models increased shipping rate only a little.
  • The real driver is process and team expectation: very low process, remove every barrier to shipping.
  • Every person is empowered to take an idea to the world in less than a week, sometimes a day.

“We've been moving pretty fast for several quarters now, so I think it it's not fully Mythos.”

Kat Wu · 11:00

“We want to remove every single barrier to shipping things.”

Kat Wu · 11:00
#anthropic#culture#shipping speed#process

Hot Take· 5

Hot Take05:00

AI PMs Are Getting It Wrong: The Job Is Now Speed

Kat Wu says most PMs still approach the role with slow, multi-quarter roadmap coordination that AI has made obsolete. Because model capability now improves so fast, feature timelines have collapsed from six months to as little as a day, so the winning PM skill is shortening the distance from idea to shipped product and defining what must work out of the box.

  • Pre-AI, tech shifts were slow, so PMs planned on 6-12 month horizons and heavily coordinated with partner teams because code was expensive to write.
  • AI has cut feature timelines from 6 months to 1 month, and sometimes to 1 week or 1 day.
  • Less emphasis on aligning multi-quarter roadmaps with partner teams; more on getting something out the door fast.
  • The best AI-native PMs shorten idea-to-user time and define the most important tasks that must work out of the box.

“the timelines for a lot of our product features have gone down from 6 months to 1 month and sometimes to 1 week or even…”

Kat Wu · 05:30
#product management#ai#shipping speed#anthropic
Hot Take15:30

The PM, Engineer, and Designer Roles Are Merging

Kat argues the traditional role boundaries are collapsing: PMs do engineering, engineers do PM work, and designers PM and ship code. The Claude Code team specifically hires engineers with strong product taste, because engineers who can go from user feedback to shipped product with almost no PM involvement are the most efficient way to ship.

  • All roles are merging; PMs, engineers, and designers increasingly do each other's work.
  • You can either hire engineers with product taste or hire more PMs to guide engineers.
  • The Claude Code team focuses on hiring engineers with great product taste to reduce shipping overhead.
  • Some engineers go from Twitter feedback to a shipped product by end of week with almost no product involvement.

“I think all of the roles are merging. PMs are doing some engineering work. Engineers are doing PM work. Designers are PMing and also landing…”

Kat Wu · 16:00

“On our team we're pretty focused on hiring engineers with great product taste.”

Kat Wu · 16:30
#hiring#product management#engineering#careers
Hot Take51:30

The Hardest PM Skill: Being the Right Amount of AGI-Pilled

Kat calls out the hardest emerging PM skill. It's easy to build for a hypothetical super-intelligent model, where you just need a text box because the model figures everything out. The genuinely hard and rare skill is building for today's model: eliciting its maximum capability, guiding users onto the golden path, and patching the model's current weaknesses.

  • It is very hard to be the right amount of AGI-pilled.
  • Building for a hypothetical super-AGI model is easy: just a text box, since it can add tools and ask clarifying questions.
  • The hard part is eliciting maximum capability from the current model.
  • That means guiding users onto the golden path, playing to model strengths, and patching weaknesses.

“I think it is very hard to be the right amount of AGI pilled.”

Kat Wu · 52:00

“I think the hard thing is figuring out for the current model how do you elicit the maximum capability?”

Kat Wu · 52:30
#product management#ai#model capability#product taste
Hot Take1:04:00

Build Products That Don't Work Yet

Kat describes a recurring pattern: build products that don't work with the current model so you know exactly what's missing, then swap in each new model to see if it closes the gap. She points to code review, which failed a few times until Opus 4.5/4.6 made it reliable enough that engineers rely on it passing before merging PRs.

  • It's important to build products that don't necessarily work yet, so you learn what's missing.
  • When a new model ships, you swap it into the prototype you already made to see if it closes the gap.
  • Code review was attempted several times and only became reliable with the newest models.
  • The engineering team now relies on the code review passing before merging PRs, using multiple review agents across the codebase.

“Um it's pretty important to build products that don't necessarily work yet so that you know, okay, what is missing um for this product to…”

Kat Wu · 1:04:30
#product strategy#model capability#code review#prototyping
Hot Take1:09:30

A 95% Automation Is Not an Automation

Kat's advice for automating your own work: push it all the way to 100%. She sees people get an automation to 90-95% accuracy and give up, but if it doesn't work 100% of the time it isn't really an automation. The last 5-10% takes real effort and teaching, but only then can you actually rely on it.

  • Anytime you do a manual task repeatedly, try to automate it with Claude Code or Cowork.
  • People often get an automation to 90-95% accuracy and then give up.
  • If an automation doesn't work 100% of the time, it's not really an automation.
  • The last 5-10% takes elbow grease and teaching the model your preferences, but delivers the real leverage.

“If an automation doesn't work 100% of the time, it's not really an automation. And that last 5 to 10% does take more time.”

Kat Wu · 1:09:30
#automation#productivity#ai tools#workflow

Explainer· 2

Explainer32:30

When to Use Claude Code vs Desktop vs Web/Mobile vs Cowork

Kat gives her mental model for the product suite. She uses Claude Code in the terminal for one-off coding tasks with the latest features; Desktop for front-end work with a live preview pane and for non-technical users who dislike the terminal; web and mobile to kick off tasks on the go without a laptop. The simplest split: code output goes to Claude Code, non-code output goes to Cowork.

  • Claude Code in the terminal is the most powerful surface and gets features first; best for one-off coding tasks.
  • Desktop shines for front-end work with a live preview pane and suits non-technical users uncomfortable in a terminal.
  • Desktop is also a one-stop control plane to see CLI, desktop, web, and mobile sessions at a glance.
  • Web and mobile let you kick off tasks on the go without a laptop; if the output is code use Claude Code, if not use Cowork.

“So, I tend to use uh Claude Code in the terminal when I'm just kicking off like a one-off coding task, and I want all…”

Kat Wu · 32:30

“So, the way that I split the products in my mind is if I'm building something where the output is code, I'll use Cloud Code…”

Kat Wu · 35:00
#claude code#cowork#tools#workflow
Explainer29:00

Why a Unifying Mission Is Anthropic's Secret Weapon

Kat names a unifying mission as one of the two most important ingredients in Anthropic's rise. Because everyone is hired to care about bringing safe AGI to all of humanity and the mission sits above any product line, the org can make fast decisions that cut across teams. Mission means teams willingly sacrifice their own goals and KRs for Anthropic's, and everyone stands behind the chosen priority.

  • A unifying mission and focus are the two most important ingredients in Anthropic's success.
  • People are hired to care most about bringing safe AGI to all of humanity.
  • Putting the mission above any product line enables fast, org-wide decisions.
  • Mission means teams sacrifice their own goals and KRs in service of Anthropic's, then stand behind the decision.

“We hire people who care most about bringing safe AGI to all of humanity.”

Kat Wu · 29:30

“Mission means that teams are willing to make sacrifices that hurt their own goals and their own KRs in service of Anthropic's goals and Anthropic's…”

Kat Wu · 31:00
#anthropic#mission#culture#decision making

Story· 1

Story36:30

How Cowork Built a Polished 20-Page Deck Overnight

Kat recounts using Cowork to build a conference talk deck. After connecting Google Drive and Slack and giving it her marketer's draft and the narrative she wanted, it worked for an hour, searched Twitter and internal channels for demos, and synthesized a 20-page deck she woke up to. Because it had access to Anthropic's design system, it looked like a designer made it; she just gave one round of feedback to trim the words.

  • She fed Cowork her PMM's draft outline, connected data sources, and told it the narrative she wanted.
  • It worked for about an hour, searching Twitter, the evergreen launch room, and the Claude Code announce channel for demos.
  • It produced a 20-page deck overnight that was pretty good with only a few tweaks needed.
  • Because it had access to the whole design system, the output looked like an Anthropic designer made it.

“And it synthesized all this together to this 20-page deck that I woke up to this morning.”

Kat Wu · 37:30

“And because Cohere has access to our whole design system, it actually looks like an Anthropic designer put it together.”

Kat Wu · 38:00
#cowork#workflow#productivity#ai agents

Q&A· 2

Q&A12:00

What Actually Happened With the Claude Code Source Leak

On the Claude Code source code that leaked publicly, Kat explains it was human error, not a malicious act. A person working with Claude on a package-release PR made a mistake that passed two layers of human review; the person is still at Anthropic, and the team has hardened its processes so it can't happen again.

  • The leak was the result of human error, not intent.
  • It came from a human working with Claude on a PR to update how packages are released.
  • The change went through two layers of human review and still slipped through.
  • The person is still at Anthropic; the team treated it as a process failure and shipped new safeguards.

“Um we realized that this was the result of human error.”

Kat Wu · 12:00

“It's it's a process failure and the most important thing is to just like learn from it and to add more safeguards so that doesn't…”

Kat Wu · 12:30
#anthropic#security#claude code#incident
Q&A12:30

Why Anthropic Prioritized First-Party Over Third-Party Clients

Kat addresses the backlash over restricting subscription use in third-party clients. Demand for Claude has been high, and while the team worked to scale infrastructure and make the harness more token-efficient, it was forced to prioritize first-party products and the API, since third-party tools have different usage patterns. Everyone still gets some credits alongside their subscription.

  • Demand for Claude has been very high, straining infrastructure.
  • The team worked to scale infra and make the harness more token-efficient.
  • The harness wasn't designed for third-party products, which have different usage patterns.
  • Everyone gets some credits with their subscription, but first-party products and the API were prioritized.

“It wasn't designed for third-party products which have different uh usage patterns than our first-party ones.”

Kat Wu · 13:00

“But yeah, we we did have to make the hard decision that we needed to prioritize our first-party products and our API.”

Kat Wu · 13:30
#anthropic#pricing#product strategy#infrastructure

Takeaway· 4

Takeaway07:30

Ship Almost Everything in Research Preview

To move fast, the Claude Code team ships nearly all features labeled as research previews. Clearly branding a feature as an early, possibly-unsupported experiment lowers the team's commitment bar, letting them get something into users' hands in a week or two to gather feedback.

  • Almost all Claude Code features ship in research preview.
  • Features are clearly branded as early and possibly not supported forever.
  • This reduces the commitment required to ship and enables a 1-2 week turnaround.
  • The point is to get feedback and iterate rather than commit to a finished feature.

“So, uh for Cloud Code, what we do is we actually ship almost all of our features in research preview.”

Kat Wu · 07:30

“And what this does is it reduces it reduces our commitment for shipping something. We can just get something out in a week or two.”

Kat Wu · 08:00
#shipping#product management#claude code
Takeaway17:30

As Code Gets Cheap, Product Taste Becomes the Rare Skill

Asked which background is most valuable, Kat keeps returning to product taste. When code becomes cheap to write, the scarce and valuable skill is deciding what to write, what UX is right, and which of tens of thousands of feature requests are actually worth building. An engineering background helps for now because it gives a sense of how hard something is to build, aiding prioritization.

  • As code becomes cheaper to write, deciding what to write becomes more valuable.
  • Taste means choosing which of tens of thousands of GitHub issues to build and how to build them well.
  • Product taste can come from any background but is the most important skill.
  • Engineering background helps prioritization for the next few months because you sense how hard something is to build.

“I still think it comes back to product taste. Like as code becomes much cheaper to write, the thing that becomes more valuable is deciding…”

Kat Wu · 18:00
#product taste#careers#prioritization#engineering
Takeaway25:30

The Cost of Shipping Daily: Product Consistency

Kat is candid about the tradeoff of Anthropic's pace: product consistency. When code was expensive, teams carefully planned one product per use case; now they ship overlapping features on purpose to let users tell them which form factor is better. The cost is user confusion and a sense of an ever-faster treadmill, which the team is trying to ease with better in-product education.

  • The main sacrifice for speed is product consistency.
  • Historically each use case got one carefully planned product; now features intentionally overlap.
  • Overlap often exists to test two internally-loved form factors and let external users pick the winner.
  • New users struggle to know the best path, and people feel pressure to check Twitter daily to keep up.

“We're sacrificing product consistency.”

Kat Wu · 25:30

“I think with these agentic tools, not just Cloud Code and Cower, but like across the whole ecosystem, people feel this need to like check…”

Kat Wu · 27:00
#product strategy#tradeoffs#user experience#anthropic
Takeaway53:30

How to Build Model Intuition: Introspection, Trusted Testers, Evals

Kat shares three ways to develop a feel for a model. First, spend a lot of time using it and ask the model to introspect on why it did something unexpected, which reveals what misled it so you can fix the harness. Second, find the handful of people whose feedback about a model is genuinely qualified. Third, build a small number of great evals to quantify goals and progress.

  • Spend a ton of time talking to and using the model.
  • When the model does something unexpected, ask it to introspect on why, then fix the harness to close the gap.
  • Find the ~5 trusted people who are best at articulating what makes a model or model-harness combination good.
  • You don't need hundreds of evals; just 10 great evals help quantify the goal, progress, and gaps.

“I think it's spending a ton of time talking and using the model. One of the things I really like to do is to ask…”

Kat Wu · 53:30

“You don't need to build hundreds of evals for them to be useful. Just building 10 great evals is important for helping the team quantify…”

Kat Wu · 55:00
#evals#model evaluation#ai#harness