interview

Joel Lewenstein

interview

Joel Lewenstein

interview

Joel Lewenstein

interview

Joel Lewenstein

Case study

When code is no longer the constraint

Joel Lewenstein

interview

Joel Lewenstein

Case study

When code is no longer the constraint

Joel Lewenstein

interview

Joel Lewenstein

Case study

When code is no longer the constraint

Joel Lewenstein

interview

Joel Lewenstein

Case study

When code is no longer the constraint

Joel Lewenstein

interview

Joel Lewenstein

Frontier ai lab - 1,500+ employees - San francisco / distributed

Frontier ai lab - 1,500+ employees - San francisco / distributed

Frontier ai lab - 1,500+ employees - San francisco / distributed

Frontier ai lab - 1,500+ employees - San francisco / distributed

summary

How should design operate when building code isn't the bottleneck?

At Anthropic, Claude writes most of the code. 
That changed what design is for.

Claude now writes the majority of code across Anthropic. Features that used to take six weeks ship in forty-eight hours. Engineering speed has roughly 10x’d; design speed, even with AI tooling, has moved closer to 3 to 5x. That asymmetry is the defining tension of the team’s year.

The response has been pragmatic: custom internal tools that compress exploration, automation agents that handle routine review work, and a baseline expectation that every designer ships code directly into production within their first week. What’s left over is the work that matters most — the last 20%, the deciding of what to make at all, the aliveness to culture that AI still can’t do well.

summary

How should design operate when building code isn't the bottleneck?

At Anthropic, Claude writes most of the code. 
That changed what design is for.

Claude now writes the majority of code across Anthropic. Features that used to take six weeks ship in forty-eight hours. Engineering speed has roughly 10x’d; design speed, even with AI tooling, has moved closer to 3 to 5x. That asymmetry is the defining tension of the team’s year.

The response has been pragmatic: custom internal tools that compress exploration, automation agents that handle routine review work, and a baseline expectation that every designer ships code directly into production within their first week. What’s left over is the work that matters most — the last 20%, the deciding of what to make at all, the aliveness to culture that AI still can’t do well.

summary

How should design operate when building code isn't the bottleneck?

At Anthropic, Claude writes most of the code. 
That changed what design is for.

Claude now writes the majority of code across Anthropic. Features that used to take six weeks ship in forty-eight hours. Engineering speed has roughly 10x’d; design speed, even with AI tooling, has moved closer to 3 to 5x. That asymmetry is the defining tension of the team’s year.

The response has been pragmatic: custom internal tools that compress exploration, automation agents that handle routine review work, and a baseline expectation that every designer ships code directly into production within their first week. What’s left over is the work that matters most — the last 20%, the deciding of what to make at all, the aliveness to culture that AI still can’t do well.

summary

How should design operate when building code isn't the bottleneck?

At Anthropic, Claude writes most of the code. 
That changed what design is for.

Claude now writes the majority of code across Anthropic. Features that used to take six weeks ship in forty-eight hours. Engineering speed has roughly 10x’d; design speed, even with AI tooling, has moved closer to 3 to 5x. That asymmetry is the defining tension of the team’s year.

The response has been pragmatic: custom internal tools that compress exploration, automation agents that handle routine review work, and a baseline expectation that every designer ships code directly into production within their first week. What’s left over is the work that matters most — the last 20%, the deciding of what to make at all, the aliveness to culture that AI still can’t do well.

Engineering is going 10x faster. Design is going 3 to 5x faster. That ratio is the story.

Joel Lewenstein

HEAD OF PRODUCT DESIGN, ANTHROPIC

Engineering is going 10x faster. Design is going 3 to 5x faster. That ratio is the story.

Joel Lewenstein

HEAD OF PRODUCT DESIGN, ANTHROPIC

Engineering is going 10x faster. Design is going 3 to 5x faster. That ratio is the story.

Joel Lewenstein

HEAD OF PRODUCT DESIGN, ANTHROPIC

Engineering is going 10x faster. Design is going 3 to 5x faster. That ratio is the story.

Joel Lewenstein

HEAD OF PRODUCT DESIGN, ANTHROPIC

KEY INSIGHTS

What we learned

01

The last 20%

AI gets UI work to "pretty good" quickly. The final push to "right" still takes disproportionate human effort, and that gap isn’t closing as fast as everything else.

02

Designers as toolmakers

With building costs near zero, designers build their own tools instead of requesting them. Anthropic runs an internal marketplace of bespoke design tools, many spun up in an afternoon.

03

Code us the clay

Every designer ships a pull request in their first week. The point isn’t learning to code — it’s expressing design intent without translation loss.

04

Automate to protect

Routine work gets handed to agents the team builds itself. Automation isn’t a substitute for craft; it’s what creates the standing to protect craft.

KEY INSIGHTS

What we learned

01

The last 20%

AI gets UI work to "pretty good" quickly. The final push to "right" still takes disproportionate human effort, and that gap isn’t closing as fast as everything else.

02

Designers as toolmakers

With building costs near zero, designers build their own tools instead of requesting them. Anthropic runs an internal marketplace of bespoke design tools, many spun up in an afternoon.

03

Code us the clay

Every designer ships a pull request in their first week. The point isn’t learning to code — it’s expressing design intent without translation loss.

04

Automate to protect

Routine work gets handed to agents the team builds itself. Automation isn’t a substitute for craft; it’s what creates the standing to protect craft.

KEY INSIGHTS

What we learned

01

The last 20%

AI gets UI work to "pretty good" quickly. The final push to "right" still takes disproportionate human effort, and that gap isn’t closing as fast as everything else.

02

Designers as toolmakers

With building costs near zero, designers build their own tools instead of requesting them. Anthropic runs an internal marketplace of bespoke design tools, many spun up in an afternoon.

03

Code us the clay

Every designer ships a pull request in their first week. The point isn’t learning to code — it’s expressing design intent without translation loss.

04

Automate to protect

Routine work gets handed to agents the team builds itself. Automation isn’t a substitute for craft; it’s what creates the standing to protect craft.

KEY INSIGHTS

What we learned

01

The last 20%

AI gets UI work to "pretty good" quickly. The final push to "right" still takes disproportionate human effort, and that gap isn’t closing as fast as everything else.

02

Designers as toolmakers

With building costs near zero, designers build their own tools instead of requesting them. Anthropic runs an internal marketplace of bespoke design tools, many spun up in an afternoon.

03

Code us the clay

Every designer ships a pull request in their first week. The point isn’t learning to code — it’s expressing design intent without translation loss.

04

Automate to protect

Routine work gets handed to agents the team builds itself. Automation isn’t a substitute for craft; it’s what creates the standing to protect craft.

in action

Omelette, prototyping in native code

CHELSEA LARSSON · HEAD OF CONTENT DESIGN

Omelette is an internal design tool that turns Claude's HTML generation into a sandbox for prototyping, sharing, and iterating. The team taught Claude the full Anthropic design system — fonts, colors, content guidelines, existing product surfaces — so every output already looks native.

The real superpower comes when you're not prototyping in a playground — when you're pushing to production. That's the end product your audience sees.

Nate Parrott

Product designer, lab

Designers dictate a brief and generate dozens of side-by-side variants, then comment on specific elements for Claude to revise. One experimental feature — "weatherman mode" — lets the user green-screen themselves in front of a design and point at it. The idea only exists because prototyping got cheap enough to justify building it. Non-designers on the education team now ship visual work they previously couldn't get out of their heads.

in action

Omelette, prototyping in native code

CHELSEA LARSSON · HEAD OF CONTENT DESIGN

Omelette is an internal design tool that turns Claude's HTML generation into a sandbox for prototyping, sharing, and iterating. The team taught Claude the full Anthropic design system — fonts, colors, content guidelines, existing product surfaces — so every output already looks native.

The real superpower comes when you're not prototyping in a playground — when you're pushing to production. That's the end product your audience sees.

Nate Parrott

Product designer, lab

Designers dictate a brief and generate dozens of side-by-side variants, then comment on specific elements for Claude to revise. One experimental feature — "weatherman mode" — lets the user green-screen themselves in front of a design and point at it. The idea only exists because prototyping got cheap enough to justify building it. Non-designers on the education team now ship visual work they previously couldn't get out of their heads.

in action

Omelette, prototyping in native code

CHELSEA LARSSON · HEAD OF CONTENT DESIGN

Omelette is an internal design tool that turns Claude's HTML generation into a sandbox for prototyping, sharing, and iterating. The team taught Claude the full Anthropic design system — fonts, colors, content guidelines, existing product surfaces — so every output already looks native.

The real superpower comes when you're not prototyping in a playground — when you're pushing to production. That's the end product your audience sees.

Nate Parrott

Product designer, lab

Designers dictate a brief and generate dozens of side-by-side variants, then comment on specific elements for Claude to revise. One experimental feature — "weatherman mode" — lets the user green-screen themselves in front of a design and point at it. The idea only exists because prototyping got cheap enough to justify building it. Non-designers on the education team now ship visual work they previously couldn't get out of their heads.

in action

Omelette, prototyping in native code

CHELSEA LARSSON · HEAD OF CONTENT DESIGN

Omelette is an internal design tool that turns Claude's HTML generation into a sandbox for prototyping, sharing, and iterating. The team taught Claude the full Anthropic design system — fonts, colors, content guidelines, existing product surfaces — so every output already looks native.

The real superpower comes when you're not prototyping in a playground — when you're pushing to production. That's the end product your audience sees.

Nate Parrott

Product designer, lab

Designers dictate a brief and generate dozens of side-by-side variants, then comment on specific elements for Claude to revise. One experimental feature — "weatherman mode" — lets the user green-screen themselves in front of a design and point at it. The idea only exists because prototyping got cheap enough to justify building it. Non-designers on the education team now ship visual work they previously couldn't get out of their heads.

Claude Code, designing 
in production

NMeaghan Choi — PRODUCT DESIGNER

The design process ends with a merged PR, not a Figma frame. Meaghan built a set of custom Claude Code skills — /prototype, /preview, /fresh — that let her build working designs inside unfamiliar codebases in minutes, then hand the PR off to automated agents that shepherd it through review, CI, and merge.

AI gets you to 70% fast. The last 30% — consistency, edge cases, the emotional logic of how a flow feels — is where design skill shows. And it's where most AI output falls apart.

Meaghan Choi

Product designer

Claude verifies its own work by hovering over it and taking screenshots. The /loop command runs a background job that monitors each PR until it merges, so the designer can move on without tracking status.

Claude Code, designing 
in production

NMeaghan Choi — PRODUCT DESIGNER

The design process ends with a merged PR, not a Figma frame. Meaghan built a set of custom Claude Code skills — /prototype, /preview, /fresh — that let her build working designs inside unfamiliar codebases in minutes, then hand the PR off to automated agents that shepherd it through review, CI, and merge.

AI gets you to 70% fast. The last 30% — consistency, edge cases, the emotional logic of how a flow feels — is where design skill shows. And it's where most AI output falls apart.

Meaghan Choi

Product designer

Claude verifies its own work by hovering over it and taking screenshots. The /loop command runs a background job that monitors each PR until it merges, so the designer can move on without tracking status.

Claude Code, designing 
in production

NMeaghan Choi — PRODUCT DESIGNER

The design process ends with a merged PR, not a Figma frame. Meaghan built a set of custom Claude Code skills — /prototype, /preview, /fresh — that let her build working designs inside unfamiliar codebases in minutes, then hand the PR off to automated agents that shepherd it through review, CI, and merge.

AI gets you to 70% fast. The last 30% — consistency, edge cases, the emotional logic of how a flow feels — is where design skill shows. And it's where most AI output falls apart.

Meaghan Choi

Product designer

Claude verifies its own work by hovering over it and taking screenshots. The /loop command runs a background job that monitors each PR until it merges, so the designer can move on without tracking status.

Claude Code, designing 
in production

NMeaghan Choi — PRODUCT DESIGNER

The design process ends with a merged PR, not a Figma frame. Meaghan built a set of custom Claude Code skills — /prototype, /preview, /fresh — that let her build working designs inside unfamiliar codebases in minutes, then hand the PR off to automated agents that shepherd it through review, CI, and merge.

AI gets you to 70% fast. The last 30% — consistency, edge cases, the emotional logic of how a flow feels — is where design skill shows. And it's where most AI output falls apart.

Meaghan Choi

Product designer

Claude verifies its own work by hovering over it and taking screenshots. The /loop command runs a background job that monitors each PR until it merges, so the designer can move on without tracking status.

Clontent, guidelines as 
an active agent

CHELSEA LARSSON · HEAD OF CONTENT DESIGN

Clontent is a Slack-integrated content design agent built on Claude. Rather than asking designers and engineers to read a style guide, Clontent embeds the guidelines directly into product conversations and can ship string changes through a pull request.

Engineering is going 10x faster. Design is going 3 to 5x faster. That ratio is the story.

Chelsea Larsson

HEAD OF CONTENT DESIGN

The agent asks clarifying questions before answering to avoid shipping bad content on autopilot. Each week it reports what it has learned from recent conversations, and the content team decides what becomes canon. Evaluation-to-shipped-string cycles close in 15 minutes. More than 480 back-and-forths with the agent since launch earlier this year.

Clontent, guidelines as 
an active agent

CHELSEA LARSSON · HEAD OF CONTENT DESIGN

Clontent is a Slack-integrated content design agent built on Claude. Rather than asking designers and engineers to read a style guide, Clontent embeds the guidelines directly into product conversations and can ship string changes through a pull request.

Engineering is going 10x faster. Design is going 3 to 5x faster. That ratio is the story.

Chelsea Larsson

HEAD OF CONTENT DESIGN

The agent asks clarifying questions before answering to avoid shipping bad content on autopilot. Each week it reports what it has learned from recent conversations, and the content team decides what becomes canon. Evaluation-to-shipped-string cycles close in 15 minutes. More than 480 back-and-forths with the agent since launch earlier this year.

Clontent, guidelines as 
an active agent

CHELSEA LARSSON · HEAD OF CONTENT DESIGN

Clontent is a Slack-integrated content design agent built on Claude. Rather than asking designers and engineers to read a style guide, Clontent embeds the guidelines directly into product conversations and can ship string changes through a pull request.

Engineering is going 10x faster. Design is going 3 to 5x faster. That ratio is the story.

Chelsea Larsson

HEAD OF CONTENT DESIGN

The agent asks clarifying questions before answering to avoid shipping bad content on autopilot. Each week it reports what it has learned from recent conversations, and the content team decides what becomes canon. Evaluation-to-shipped-string cycles close in 15 minutes. More than 480 back-and-forths with the agent since launch earlier this year.

Clontent, guidelines as 
an active agent

CHELSEA LARSSON · HEAD OF CONTENT DESIGN

Clontent is a Slack-integrated content design agent built on Claude. Rather than asking designers and engineers to read a style guide, Clontent embeds the guidelines directly into product conversations and can ship string changes through a pull request.

Engineering is going 10x faster. Design is going 3 to 5x faster. That ratio is the story.

Chelsea Larsson

HEAD OF CONTENT DESIGN

The agent asks clarifying questions before answering to avoid shipping bad content on autopilot. Each week it reports what it has learned from recent conversations, and the content team decides what becomes canon. Evaluation-to-shipped-string cycles close in 15 minutes. More than 480 back-and-forths with the agent since launch earlier this year.

FEATURED ANTHROPIC TEAM MEMBERS

NAME

POSITION

Joel Lewenstein

Head of Product Design

Meaghan Choi

Product Designer

Nate Parrott

Product Designer, Labs

Kyle Turman

Product Designer

Matt Gallivan

User Researcher

Chelsea Larsson

Head of Content Design

FEATURED ANTHROPIC TEAM MEMBERS

NAME

POSITION

Joel Lewenstein

Head of Product Design

Meaghan Choi

Product Designer

Nate Parrott

Product Designer, Labs

Kyle Turman

Product Designer

Matt Gallivan

User Researcher

Chelsea Larsson

Head of Content Design

FEATURED ANTHROPIC TEAM MEMBERS

NAME

POSITION

Joel Lewenstein

Head of Product Design

Meaghan Choi

Product Designer

Nate Parrott

Product Designer, Labs

Kyle Turman

Product Designer

Matt Gallivan

User Researcher

Chelsea Larsson

Head of Content Design

FEATURED ANTHROPIC TEAM MEMBERS

NAME

POSITION

Joel Lewenstein

Head of Product Design

Meaghan Choi

Product Designer

Nate Parrott

Product Designer, Labs

Kyle Turman

Product Designer

Matt Gallivan

User Researcher

Chelsea Larsson

Head of Content Design

Join the team

Anthropic Design

Anthropic's design team is made up of product designers, content designers, user researchers, and design engineers shaping Claude, Claude Code, and Anthropic's product surfaces. We're hiring across every discipline.

Join the team

Anthropic Design

Anthropic's design team is made up of product designers, content designers, user researchers, and design engineers shaping Claude, Claude Code, and Anthropic's product surfaces. We're hiring across every discipline.

Join the team

Anthropic Design

Anthropic's design team is made up of product designers, content designers, user researchers, and design engineers shaping Claude, Claude Code, and Anthropic's product surfaces. We're hiring across every discipline.

Join the team

Anthropic Design

Anthropic's design team is made up of product designers, content designers, user researchers, and design engineers shaping Claude, Claude Code, and Anthropic's product surfaces. We're hiring across every discipline.


Get new case studies & report markdown

Download the markdown version of the report, ready to drop into any tool. Get notified as new case studies go live.

By subscribing, you agree to receive communications from Designer Fund and Foundation Capital in accordance with their privacy policies.

Methodology

This report draws from

0
Survey responses
0+
Interviews
0+
Public sources

©2026 Designer Fund, Foundation Capital. All rights reserved

Get new case studies & report markdown

Download the markdown version of the report, ready to drop into any tool. Get notified as new case studies go live.

By subscribing, you agree to receive communications from Designer Fund and Foundation Capital in accordance with their privacy policies.

Methodology

This report draws from

0
Survey responses
0+
Interviews
0+
Public sources

©2026 Designer Fund, Foundation Capital. All rights reserved

Get new case studies & report markdown

Download the markdown version of the report, ready to drop into any tool. Get notified as new case studies go live.

By subscribing, you agree to receive communications from Designer Fund and Foundation Capital in accordance with their privacy policies.

Methodology

This report draws from

0
Survey responses
0+
Interviews
0+
Public sources

©2026 Designer Fund, Foundation Capital. All rights reserved