LongCut logo

Full Tutorial: From Design to Code with Claude Code in 40 Minutes | Meaghan Choi

By Peter Yang

Summary

## Key takeaways - **Designer shipping to prod feels magical**: Shipping code to production as a designer feels both terrifying and magical, bridging the historical gap between design and development roles. [01:11] - **AI code tools streamline design polish**: AI code tools allow designers to handle the final 10% of polish themselves, addressing small details and pixel changes that previously required engineers. [01:48], [02:06] - **Claude Code aids exploration and codebase understanding**: Claude Code supports three main workflows for designers: zero-to-one exploration, understanding the existing codebase, and final polish, with the latter two used most frequently. [02:51], [03:33] - **Code is the ultimate source of truth**: Understanding how designs translate into code is crucial for creating realistic product experiences, making code the ultimate source of truth over specs or Figma. [06:16], [06:26] - **Custom .md files tailor AI for designers**: Custom cloud.md files allow users to direct AI like Claude to adopt specific personas or instructions, such as acting as a tool specifically for product designers. [19:13], [20:44] - **Fluid roles accelerate feature development**: At Anthropic, fluid roles and rapid prototyping, often without extensive specs for smaller features, allow for quick iteration and faster feature releases. [22:47], [24:31]

Topics Covered

  • Designers can now ship code to production.
  • Why code is the ultimate source of product truth.
  • Customize your AI tools for unique workflows.
  • AI tools foster fluid, collaborative team roles.
  • How to become an AI-native designer.

Full Transcript

I myself probably spend an equal amount

of time now in quad code as I do in

Figma. Instead of handing off a Figma

mock to an engineer, I'll hand them like

a draft PR that I have with like a

description of what I was trying to do.

So, right away I say like from my eye, I

can tell, oh, this doesn't look exactly

how the app I designed did, right?

There's all these little details. And

so, this is actually where I spend most

of my time in cloud code. Last 10% of

polish. At Anthropic, anyone can code

and it's such a collaborative

environment that I feel like our roles

don't matter. you have like another

slide uh where you have this like whole

flow diagram. Can you walk through this

one too?

>> Yeah, definitely. So,

>> okay, welcome everyone. My guest today

is Megan uh as a design lead for clock

code. Megan not only designs features

but also ships to prod on a regular

basis and I'm really excited to talk to

her about how AI can help designers work

faster and get her to demo her exact

workflow to go from design to prototype.

So, welcome Megan.

>> Thanks so much. I'm super excited to be

here and yeah, just really excited to

talk about how I use cloud code myself.

>> Yeah. All right. So, um, so I guess

let's start with this. Um, how does it

feel to be finally a designer who ships

to prod on a regular basis?

>> Honestly, it feels a little terrifying

and simultaneously very magical. I

think, you know, over the years, I've

always really just been interested in

development, but it felt like at the

time, especially when I started my

career, you really need to hone your

craft in either design or either

development. It was just hard to bridge

the gap into between both unless you're

a really special unicorn.

>> And then just in the past, I'd say like

5 years, it just became so much easier

to code. and all those like little

details that I always felt I couldn't

really ask my engineers cuz it'd be a P2

or it would take me a really long time

to set up and fix, I can just ship them

and have them. And so like that really

polished product that you always want is

just there and ready for you. Uh it's

pretty amazing.

>> Yeah, exactly. I always feel bad asking

my engineers be like, "Hey, can you

tweak like five different copies like

make this pixel change or something?"

>> And and yeah, now you can just do it

yourself right?

>> Mhm. Exactly. And sometimes, you know, I

think when you're actually porting

designs or something that you have in

this idea into prod and then you

actually see it, you're like, "Oh, that

actually didn't look how I thought I

would." And then you feel even worse

asking them to change it back or

changing again. And this just gives you

the power to do it all yourself.

>> That's awesome. That's awesome. So,

maybe uh you can walk through like the

entire design process and and where like

AI and cloud code can help.

>> Definitely. So I would say for me I

actually have a slide for this I can

share which is really useful. Let me

share my screen.

>> Yeah.

>> Right now there's a few different ways

my workflow has changed. I would say a

big one that I have is kind of I would

bucket it into three big workflows that

I now use cloud code for specifically.

Uh this first one is one that I think

folks who are maybe less technical reach

for the most often. uh and it's one that

you can do even in cloud AI today with

artifacts uh without needing necessarily

the code tool on top of it and this is

0ero to1 exploration so you can kind of

explore new ideas but very often this is

kind of in a silo and it's not

necessarily in your codebase I think the

superpower of these 0ero to one

explorations if you're doing it in cloud

code is that it can actually be in the

codebase and understand your code and

where it fits and so it's a lot more

real than doing just it on a side on a

prototype and then the other two

workflows that I do most common now are

to um I'll go to the third one here

actually understand the codebase often

when I'm developing a new feature or I'm

working on a new project I'll actually

spend my first pre-brainstorm era uh

asking claude about how it currently

works uh how it was implemented today

down to like even what the system prompt

is and how it's configured how it's

architected uh to really understand the

use cases sometimes I'll even ask like

how would you design this differently

Uh, and then there's a phase where I'll

go into Figma and then do the designs

and then I'll go back into code once

it's ready and I'll kind of do the final

polish pieces if an engineer has built a

lot of the infrastructure that needs to

be there or sometimes I'll try it myself

at a first stab and see how far I can

get with uh cloud code.

>> Got it. Okay. And you probably do the

understanding codebase and building the

code more, right? Like working on a you

know mature feature, right?

>> Definitely. Definitely. I would say this

first bucket I actually I like rarely do

at work. It's something that I do maybe

myself in my personal time and personal

projects but when it comes to working as

a designer like with my engineering team

with a product team I'd say 90% of my

time is spent on the latter two use

cases here.

>> Awesome. And I think you have like you

have like another slide uh where you

have this like whole flow flow diagram

like explore.

>> Yeah.

>> Can you walk through? Yeah definitely.

So I think like this is where you get a

little bit more granular into what these

actually feel. I kind of broke this out

into what the typical kind of design

process looks like. You know, the double

diamond of like explore and discover,

then design, then prototype, then you

kind of polish and you finish. And so I

think the way I look at it is claude

code a lot of people will assume it's

really only for development. And so

maybe in like step four and five here,

but I think step 1, 2, and three are

actually where cloud code shines for

non-technical folks just because it lets

you do a lot of I would say like

technical adjacent tasks or like

engineering adjacent tasks because you

have access to the codebase. I actually

think one, two, and three here is

something that even our very technical

uh customers do in cloud code that they

don't even realize they're doing because

it's just so easy to get started. And

that's really, you know, exploring

different solutions that you could uh

understand, exploring the history of why

something was implemented the way it was

and then planning and like sequencing

how you can break out uh building

different features. Uh the reason I like

to do this more in cloud code actually

than in cloud AI although I'll often go

back and forth in between the two is

that um as a designer our job is really

to kind of take all the information all

the feedback from all different parties

including our users including engineers

including our product managers and

translate that to a working product

experience. But that working product

experience is in code. So the faster you

can understand how it's going to get

implemented in code, the more realistic

it's going to be of representing what

the actual product experience is.

>> Yeah. Like I I think the code is the

source of truth, right? And like I know

I I spent like 10 years building

products and

>> uh I always noticed that you know I I

write like a spec and a PD and then

whenever the Figma is actually you know

ready to go like people don't look at my

spec anymore because Figma is a lot more

higher than the spec and then eventually

they move into the actual code and and

like with this process you can you can

basically go into the code direct

directly right you can start making

prototypes exploring the code direct

directly.

>> Exactly. Exactly. In fact, sometimes in

the past, I've instead of handing off a

Figma mock to an engineer, I'll hand

them like a draft PR that I have with

like a description of what I was trying

to do and where I didn't necessarily I

wasn't able to get to completion myself

cuz you know, at the end of the day, a

lot of stuff is still quite difficult if

you're not uh very coding knowledgeable.

And so, it just is really a great

starting point to developing features a

little bit further down the process than

you typically would.

>> And uh you know, now you're like a

trusted member of the team, right? like

when you joined the team, how did you

get engineers or like get people to give

you like GitHub access and like start

doing this stuff? What kind of

conversations do you have?

>> Uh, you know, I think at Anthropic

anyone can code and it's such a

collaborative environment that I feel

like our roles don't matter. So here it

was actually very easy. In fact, I'd say

the team was super excited when they saw

that I wanted to do this. A lot of the

designers code even themselves. So even,

you know, before we had cloud code

directly integrated, you would see a lot

of designers make PRs and kind of do it

themselves. Okay.

>> Um, at companies where was a little bit

harder, maybe a little bit toned down,

um, part of my process is actually to

sit down with an engineer and have them

pair program with me for a day because

honestly different repositories are set

up so differently. I'm sure you know

this. I was a little bit less familiar

with this when I first started. Even

like how to get a preview app running

and get all the like dependencies

installed and my local dev environment

set up was really hard for me in a new

company. And so uh I think it helps you

build good partnership with your edge

partners. And then it also helps you

understand like the unspoken

semantics of how like code gets pushed

at your company and like expectations of

deployment times and like preview builds

and things like that.

>> And another quick question like how do

you do you guys do you guys have like

design reviews or like you know you have

to go present to Mike or like do you

just present the code or how do you guys

do that? Yeah.

>> Uh you know it's interesting. We do have

product reviews and we do have design

crit and sometimes we'll have design

deep dives or design reviews. It's a

mix. Sometimes you'll see Figma

prototypes presented. Sometimes you'll

see fully working apps presented.

Sometimes you'll see a local preview

presented. It kind of depends on kind of

what the feature is at the time. And I

do say this a lot to kind of other

designers I work with now that uh this

skill set of code is available to a lot

of nontechnical folks. um you should

just consider this as a new way you can

present your work or new way you can

work and similarly to the early

questions for design or product like

should this be low fidelity should this

be high fidelity I think you should

consider code the highest possible

fidelity so and the effort it takes and

kind of what feedback you get will

change depending on which one you show

>> got it okay got it I mean ultimately

it's about like kind of getting a feel

for what the user will actually see and

feel in the product right so like you

just make prototypes for everything you

make prototypes for the core user

experience. I guess

>> mhm I think there are still some

instances where I personally find it

easier to be in Figma rather than in

prototyping. This could be a limitation

definitely of my own technical ability.

But for bigger changes that are harder

to do architecturally in code and

sometimes I'll ask cloud like hey like

how hard is this going to be to change?

Uh if it takes me like a minute to do it

in Figma and it might take 30 minutes to

do it in code it does make sense more to

do it in Figma. So, there's still that

trade-off I think we have today of where

it makes most sense given what you're

trying to accomplish.

>> Yeah, Figma is surprisingly I mean my

designers make fun of me, but like like

just trying to get auto layout to work

is is hard. Like I can't figure it out.

>> Mhm. Definitely.

>> Figma is not that easy either. Yeah.

>> Yeah. Yeah. It's definitely a power user

tool, I would say.

>> This episode is brought to you by

Linear. Let's be honest, most project

management tools suck. That's why I love

working with Linear because they share

my obsession for craft and quality.

Linear is incredibly fast, beautifully

designed, and comes with built-in

support for AI agents like cursor. As a

PM, I love using linear to capture

customer feedback, draft PRDs, and

manage development and communicate

across the organization. Product teams

at OpenAI, Versel, BRS, and Cash App all

use Linear to build complex products at

lightning speed. See for yourself by

visiting linear.app/behindthecraft.

That's linear.app/behindthecraft.

Okay, now back to our episode.

All right, so let's kind of like maybe

walk through some of this process with

like a real demo, right? Like like a do

you have like a little demo app you can

show us?

>> Yeah, definitely. So, let me walk

through. Unfortunately, I can't share

anything that I do internally, but I can

show a little demo app we have called

Claude Cafe. And you can kind of see the

mocks right here of what it looks like.

And um I already have it open and

running on cloud code right now. So I'm

in the repository Claude Cafe. I have

kind of Claude open and running. So the

first thing I tend to do is ask Cloud to

help me preview this app. And so

it's pretty straightforward to just talk

to Claude like I would my engineer and I

say help me uh run this locally. I do

think this is easier for some engineers

like they might say, "Oh, I'll just do

this in bash commands." But for me, I

don't always know the packages that you

need to install or the commands you need

to run the app. And so I find this a lot

easier. Um, one feature I really like

that we uh recently released that makes

this very possible uh that I've been

using for months and has now only been

available I think for the past month

externally is background bashes. And so

you can kind of see here it says

controlB to run in background. Um

running a dev server uh as you probably

know but maybe not all the nontechnical

folks don't know it's just an ongoing

process. And so if you hit controlB you

can kind of see it runs in the

background now and it'll update you as

it's going but you can kind of keep

chatting with Claude even as it's still

running. Oh, nice. So, is that like

basically having two agents going? Uh,

>> kind of

>> a little bit, I would say. I mean, like

right here, you can see that it's been

successful. We do have the ability to

run parallel agents. Uh, this just lets

you run background bash commands. So,

anything that's like an ongoing command.

>> Got it. Okay. You want to should we open

a local host and see how

>> Definitely. So, we can see what it looks

like now. Oh, sorry. It opened in

another browser, but I can pull it in.

This is what the app looks like. So you

can see it's running on preview and the

app.

>> It looks great. Yeah.

>> So right away I say like from my eye I

can tell, oh this doesn't look exactly

how the app I designed did, right? Like

the images, if you can see it, aren't

necessarily properly aligned. The header

is supposed to be left aligned. Like

there's all these little details that uh

I can see how it was implemented that

way. And there's just like things that I

would want to change to get it to that

I'd say like last 10% of polish.

>> Mhm. And so this is actually where I

spend uh most of my time in cloud code.

And so we will have the app running and

I'll there's a few different ways you

can do this. So in a previous video I

referenced using the Figma MCP server. I

would say sometimes I do that if it's

very polished, but Claude actually has

vision capabilities. So I'll typically

start instead of using the MCP server by

um exporting this as a PNG

>> to my desktop.

Uh, so let me send it to my desktop for

now. It's there.

And then I will drag and drop that into

claude

and I'll say, can you update the header

and image alignment

so it looks more like

this block?

>> Okay. Um,

>> it's like a little bit lower fidelity,

but I think the Figma MCP server for me,

sometimes I actually don't have it all

always logged in or set up properly.

And, uh, switching between dev mode for

me is, um, just my personal preference

of doing it through images. Uh, the

other way I tend to do this, which is a

little bit more advancedish, I would

say, depending on where you fall on the

spectrum of technical ability, is I'll

actually go in weapon spec mode. I'll

see kind of where the alignment or class

is and then I'll direct Claude

specifically to look at this page

feature uh and then make the updates

that way. So there's a lot of different

ways you can kind of get into this

workflow.

>> So like you actually paste the code here

to take a look look.

>> Um I will sometimes

I like to actually look at the classes

or the way that it's rendered and I'll

tell Cloud to search for that code line

and update it to the one that I think it

needs to be.

>> Okay. because it's like cloud has access

to all this. Yeah.

>> Yes. Yes.

>> Yeah. Makes sense. Yeah.

>> Um I'd also say in this case, you saw

that this kind of went directly to

coding. This is actually not my typical

workflow. Normally, what I'll do is I

actually go into planning mode first cuz

I think this is the best way to do it.

be like actually can you provide an

overview of the plan and your approach

also uh

break down this task into easy to

understand steps and changes. Um I

sometimes add that last one especially

if I'm trying to make bigger changes

because Claude is kind of designed to be

able to speak to you like you're a

software engineer. I'm not a software

engineer, so I often prompt for

additional explanations.

>> Yeah, this is what I do, too. I always

get it to make a plan or make a two

to-do list and let me review the plan.

>> It's too overeager some sometimes, you

guys.

>> Definitely. Definitely. Especially if

you're in auto accept mode, it'll

definitely start going on its own way.

This also helps you, I think, if you're

learning to code, get a better

understanding of the architecture of how

a software engineer would think about

making these changes, which is really

helpful for me. So, we'll see the

changes coming in. Um, cloud of course

is very complimentary. You'll have a

sleek new header. It's really nice. And

then if this I'll read through the plan

here so you can kind of see it'll

list the file where it's making the

changes. It's updating the pad heading.

It's moving so it's left aligned. That's

what I wanted it to be. I really like

when it actually shows the CSS for me

actually because it helps me know how to

do it in the future. Y

>> and you know this generally looks good

to me. Let's say and then so I'll go yes

uh and auto accept edits modes. This is

also your best friend if you're not

necessarily understanding every line of

code that's changing. Uh, auto accept

edit modes just lets CL claude make the

changes and then you can come back and

see the final product. Uh, one really

fun thing that I like to do is watch the

changes live. I know it's like sometimes

breaks in between and they'll rerender,

but I like to see things move around.

>> Yeah. Yeah. Yeah. It's great. Yeah.

Yeah. This that's why it's very

important to get local host set up first

so you can see what what it's doing.

>> Absolutely. Absolutely. I think design

is such a visual job that it's like much

more helpful to get it all together.

>> Is this thing real by the way? Do you

guys actually have a claw cafe?

>> No. No. I mean, I wish we do have cafes

at the office and you can order drinks.

Uh I wish they were themed this way, but

uh unfortunately this is just an app

that one of the designers on our team

made for a demo.

>> Got it. Got it.

>> I've been trying to do like a lot of

like prototyping too. And oftent times

AI tends to make this like very same

look like like this shaden tailwind CSS

look with the purple backgrounds and

stuff. Like how how do you avoid this

stuff? It's like super super annoying.

Yeah.

>> Is it just using images and like giving

more precise instructions?

>> Yeah, I think so. I I tend to avoid

using claude or any AI models in general

for like net new designs without any

basis to start off with because I think

you do get that exact same styling. You

could ask honestly any model and it'll

give you a very similar styling. Um, the

tips that I would recommend are one,

I'll often reference parts of the site

that exist already or inspiration that I

have that I like the way it looks and

I'll drop it in as an image and say,

"Hey, make it look like this." So, an

example was is like a few weeks ago, I

was adding a new setting to our cloudi

page. And instead of just asking it to

generate a setting, I was like, "Hey,

look at the existing setting page,

specifically this setting, and

reimplement it this way. That's the way

I like it to look." Or if I'm doing it

sometimes in cloudi chat, I'll actually

drop screenshots of sites that I really

like or like almost like a Pinterest

board style in thing. I'll say like make

it look something similar to this.

>> Okay. Kind of like a moo moo board so it

kind of understands what it's doing.

Yeah. Yeah. Got it. Okay.

>> Um and uh let's talk about I I noticed

that you have a bunch of stuff in the

output around like lowrisk, high risk

and maybe you can pull up your cloud

file. I think you have a really great

file. Absolutely. Let me find it right

now. This is what my cloud MV lick file

looks like.

>> Yeah, this is a great file like like so

I've been using cloud MB just typing

init and then it just like analyze the

codebase. But I think this is way way

better. This is like this is like

Megan's cloud MD if you can walk through

why why you added all this stuff.

>> Yeah, definitely. So cloud.md is

actually extraordinarily powerful as an

ecosystem. I didn't fully understand it

when I first started using it, but we

made a lot of updates so that you could

have multiple hierarchical cloud MDs.

>> And so there's a sh and cloud.mmd is

kind of like shared knowledge that

claude can have access to about your

repository

>> for our entire codebase. We actually

have lots of cloudmd files shared across

the team and these are checked into the

repository. They're kind of shared with

things like the structure. That's

typically what you get when you get

slashinit that you mentioned. It like

understands your codebase. Uh over time

um we also added a personal cloud.mmd

file. So this is to my cloud running

instance. It doesn't have to do with the

codebase. It actually runs for my global

experience of cloud uh code running in

any repository and it's attached to me

as a user. This is where uh initially I

had actually asked the nam. I was like

hey could we make a version of cloud

code that is directed specifically to

product designers because it'd be really

amazing as a tool to use. and they said,

"Oh, like have you tried adding it to

your cloud MD? Like Claude is pretty

good at following instructions. We think

that you might be able to get pretty far

that way." And uh so I worked with

Claude to actually draft this cloud MD.

I iterated on it a bit and this is where

I've landed up as a way to direct Claude

to be a cloud for designers.

>> Okay. So it's like, you know, you know,

I'm a designer. You got to explain more

stuff to me. And then I I I like the

emojis like, you know, there's a

high-risisk modification.

>> Yes. Yes. Yes. Yes. I've definitely

recently I was guilty of causing an

incident. My very first one u from

pushing code to prod. Fortunately, it

was caught very quickly and resolved and

everyone's really great here. It

happens. Someone told me it happens to

everyone. If you're pushing code to

prod, you'll like cause an incident at

one point, but I did add that in after.

>> But uh but all the stuff you push to

prod is like in PRs, right? Like some

does some engineer have to re review it

or

>> Yeah, everyone's still reviewing it. But

I think things get through a little bit

sometimes things slip through the crack.

It happens to everyone. Got it. Yeah,

this is great. Uh and and basically, so

okay, so this is in your project's root

directory or is this in your global?

>> This is my global root directory. Yeah.

>> Okay. And basically, okay, let's kind of

explain how cloudd works. So basically,

it puts this into context each time you

have started a new conversation or

>> Exactly. Exactly. And so it will always

have awareness and it'll adhere to these

quite like we prompted cloud so it'll

adhere to your cloudmd files. So my

outputs frequently look very different

than what the engine would look like.

They're a lot more ver verbose and a lot

more detailed than anyone else.

>> And I have a feeling that you played a

pretty big role in shipping the new

output styles feature. Did you play a

big role in that or

>> uh I did not actually. I would say that

was a combination of a bunch of folks

across the team who were more on our go

to market side and marketing side and

comm side thinking about uh extending it

even more into like a non-technical

audience. So, I was really excited when

it came out because you can still use

it. I think alongside your cloud MD, I

felt the cloud. MD was enough for me and

I think that's one of the beauty

>> beautiful parts about cloud code. It's

so permutable.

>> It's like very customizable. You can

make it your yours. Yeah,

>> exactly. Exactly. Um but but can you

like uh can you actually go back to that

slide with the flow flowchart because I

kind of want to talk about like a real

example of a feature that you guys

shipped

>> like maybe you can share example of like

ideation stage all the way to launch and

like how you work with CAD and like the

the engineers.

>> Yeah, let's think what's a good example

of that. Why don't we take agents? I

think that's a really good one. Uh so

cloud code recent release sub aents or

agents and what these are are

programmable agents that let you oop

sorry

>> run cloud codes in parallel. So um this

was actually I think a really great

feature because it shows how like AI

tooling in general allows for a lot of

fluidities of roles. So agents was

originally conceived as an idea by an

engineer on our team Sid and he was

really rapidly prototyping the idea and

um I think Cat might have mentioned this

in her session but here at Enthropic

everyone uses cloud code very often so

we're like our best dog foodters like we

use the product all the time we give

feedback and so he built the feature he

posted internally a demo of it working

and then he launched it to our internal

working version of quad code from there

I reached out to him I like, hey, this

is a really cool feature. I started

using it. I was like, I have a bunch of

design um kind of updates I would

recommend. I think like we should

highlight this even more. I think

there's a way we could show parallel

agents. Uh and so we did some rapid

iteration cycles with myself, another

designer on the team, and Sid to kind of

iterate on the feature and get it to a

level of polish that we felt happy with.

And then once it felt good, we prototype

it a little bit more with the team. A

few other folks gave feedback who were

using it and then it was kind of ready

to go. So I feel like a lot of the

feedback cycle is actually using it

ourselves.

>> Wow. Okay. So basically like uh Sid um

maybe spent a week or a couple days just

like prototyping himself and he just

shipped it to internal users first,

right? Exactly. Exactly.

>> There's no spec or anything or like

design or anything. You just shipped it.

>> Not for smaller features like that. No,

I think it can happen pretty

organically.

>> Okay. And then you guys went back and

forth and then uh you made it better.

And then when when how do you know when

it's ready to go? like like the internal

people were like, "Hey, let's let's get

this out."

>> I think like when we initially launch

new features, we'll often get a lot of

feedback from the team. You know, our

teammates are our best sources of

feedback and our most honest. And so,

you'll kind of see initially when we

launch a new prototype, we'll get a lot

of feedback on it as people discover it.

And then over time, it'll slowly taper

off because we've addressed all the bugs

and features and it just becomes a

natural freaking part of the product. We

do also look at analytics and adoption

of features to see if they're being

wellreceived with uh ants as we call

them or folks at Anthropic. And if we

see that steadily growing and we see the

negative feedback steadily going down,

there's just a moment where we feel

like, yeah, this is good to go. I also

think we are on the side, which is

always a little bit nerve-wracking for

me, but it's been really beautiful to

see on releasing early and letting our

community also get feedback because we

can continually update and improve

things as they go.

>> Oh, because you have like an early

preview program or something like

>> uh not even an early preview. We'll just

release it early once it feels like it's

reach a point of readiness where we'll

get better feedback from our community

because we've gotten all the feedback we

can from internal folks

>> and then when you get ready to release

it the go to the market just like you

know uh maybe getting coms and per

marketing involved and putting like a

post out there and they just exactly

exactly um often it'll be the engineer

who built the feature themselves uh

making a little screen recording of them

of their feature or they'll work with

our devril team or coms team but it's

all very very

That's great. You know, like because

that's not how uh typical company ships

this stuff, right? Like you know,

>> typical company is like, you know, you

know, this waterfall thing, right? Where

like you write a spec, you go debate it

about it internally and then you go make

a design and then you like you hand it

over to engineers like here here's a

perfect design spec and go go build it.

>> Maybe you guys still do that for like

really big features like you know

enterprise or some or something. But

like yeah, I mean it's way more fun the

way that you guys do it.

>> I think so. I think so. One thing that I

often reflect upon is like how fluid the

roles have become on the team and like

how quickly you are to

>> able to just like make the product

better and make the product that you

want it.

>> Uh like the slashexplanatory slash like

learn mode that you mentioned that was

released a few weeks ago that was

actually a custom command because we

have custom slashcomands uh that you can

kind of work on and customize yourself

that someone on the team built and

they're like hey I actually think this

will be useful for a lot of other folks.

And so they started to work on that in

that direction and that evolved into our

product feature. I think one thing to

note that uh I always really like that

we still do is there's still experts on

the team and what they do. So if a new

feature is going out the engineers will

often reach out to myself or the other

designer on our team Nate and say like

hey can you give a quick pass on this

like it feels good but we just want to

make sure that aligns with our system

and we'll stamp it. Often we'll just try

it and use it. And similarly, if the

designers are building a feature, we

have code review for that to make sure

that it's all kind of good and working.

>> Does Cat have like a spreadsheet that

tracks all the features in progress or

>> Oh, yes. Oh, yes. She has a master

spreadsheet of everything that's going

on. uh she's uh we're always constantly

trying to keep it up to date and there

are a lot of kind of bigger features

that do follow a little bit more of a

traditional product development process

but I would say while it follows that

traditional process the role of like who

drafts that and like who's contributing

to it is extremely fluid across our

team.

>> I mean that's the best right where

people don't have like people still have

their expertise but like people can just

like cross across functions and just do

what's best for a product.

>> Absolutely. Absolutely. May maybe you

guys should just use like a slash camera

or something like cat can run it every

morning, check the codebase and see what

kind of new features there.

>> I mean I think that sounds like it'd be

incredible for her. Uh I'm not sure how

well that would work because we often

have like random ideas spinning out

everywhere.

>> Got it. Okay. And I think you have one

more slide on like some pro tips of

using claw hot code.

>> Yes, definitely. Let me share that as

well. So pro tips, these are my pro tips

for cloud code. I recommend these for

folks who are getting started. One thing

I will say is that these pro tips hinge

initially on the idea that you're

comfortable in terminal uh or CLI. So if

you're not, my biggest recommendation

there is actually to just chat with an

engineer on your team. Like no engineer

is ever not going to be happy to explain

to you how to use terminal or what

terminal commands are. Even people

who've used like CLIs their entire life

might not know every single command out

there. Uh, another way to learn how to

use CLIs is to just ask cloud AI in

chat. Like I'm frequently asking cloudi

how to do a thing before I open cloud

code in terminal just to make sure I'm

in the right place or have the right git

branch, things like that. Sometimes ask

cloud code in cloud code to help me.

It's kind of a little inception, but

once you get there, these are the cla

code specific key bindings and or like

best practices I would recommend. Um,

the reason I call it being comfortable

with terminal is that a lot of these are

designed for developers. So, a developer

is very comfortable and very familiar

with these key bindings. So, as you use

terminal more, you'll be more familiar

with them.

>> Um, but it might be a little bit new if

you're not. So, the first one here,

shift tab to go between auto accept mode

and plan mode. This is like my most used

mode. I think I always start in planning

mode whenever I'm doing something just

so I can make a plan with cloud and

understand the changes that it's making.

And then uh auto accept mode lets cloud

kind of autonomously accept the code

changes and I'll wait till the final uh

code changes are done before review.

>> Um MCP lets you connect to MCP servers.

These give access to more tools that

Claude can have. So the Figma MCP is one

that a lot of designer use. Um a lot of

folks also use a PlayWrite MCP to

preview their app. That can also be

really helpful. You can talk to your

engineering team about which MCPS they

have set up. Um I think it's a thing

that's actually very valuable to share.

Um, controlV to paste images. I think I

showed how I use that today. It's really

just to show cloud what I'm working on.

>> Um, escape to cancel is a big one. Just

uh if Claude is doing something and it

makes you nervous or you're not sure

where it's going, you just stop Cloud's

progress right away by hitting escape

and it'll stop it. Double escape lets

you go back in history, which I also

really like. So, if you've sent a prompt

and you didn't like where it went, you

can hit double escape twice and like go

back to that prompt and then see uh and

then update the instructions or you can

see past instructions that you've sent

to cloud.

>> Got it.

>> D-res

lets you uh reaccess an old session. So,

if you like opened a cloud code session,

closed it, and then you want to get back

into that task, you can use resume. And

then slashmemory to update your cloud.

Mmd file.

>> When do you use the last one

slashmemory? Like we have a new tip that

you want to share or like

>> Yes. Yes, I use it a lot when I end up

in a space where uh I wish uh I do a

thing repetitively with Claude and I

want Claude to just do it automatically.

And so for a little bit I did add

preview my app before I start or preview

my app for every change but that became

a little bit too hectic. So I did remove

that and I'm constantly playing around

with it to optimize Claude for my

workflows. Uh so uh in my most recent

memory update, I actually asked Claude

to look at the existing component

library before it finalizes any

implementation decisions to make sure

that we're following design systems best

practices.

>> Okay, got it. Okay. Yeah. Like actually

a lot of uh some some of these other

tools are like trying to add design

system support. So, so does clock code

just like support that natively just

like access the files and understand the

files or

>> I think this is where like you kind like

design systems is the role that really

beautifully covers design and

engineering. Um, and if the design

system team actually starts to design a

system and a component library in mind

with a coding tool or an AI collaborator

reading through it, it actually is a lot

easier than for cloud code to be able to

read through our documentation and

implement features.

>> That makes sense. Makes sense. Okay.

Okay, let me ask you some other random

questions that I forgot to ask Cat. Um,

so, so like obviously a lot of the stuff

depends on the quality of the model and

like what the model team is doing,

right? So, how do you guys work with

that team like the AI researchers? So, I

would say there are a few kind of

dedicated work streams around cloud code

for like bigger features that we know we

want to work on or like bigger

improvements that we see feedback from

the community and those have kind of

dedicated pods within the research

product team and there's partnerships to

continue iterating on them. That's one

way. It's like the very standard way of

just like we got feedback from the

community or we noticed a thing and we

want to kind of train the model to get

better at this. Um the second way is

something uh that we are kind of looking

into around how we can help cloud code

help researchers better. So that's the

second way that we partner really

closely like the software development

flow is so wide and widespread. But one

thing we think we can do is train cloud

to be really good not just at like

production code but also at researching.

So that's another pod that we have. And

then a third one is um we're very

bottoms up at anthropic as a process

perspective. So sometimes from the

product team, sometimes with the

research team, someone will have idea an

idea of a thing that they want cloud to

do. It'll spin up and they'll just start

a mini pod of working together on a

brand new thing.

>> That's great. Yeah, that's fun.

>> Um got it. Okay. And and you probably

have access to all like their early

models and stuff so like yes. Yes, we uh

definitely uh try out all the models

ourselves internally at first and uh we

get a lot of very strong feedback

immediately if the model's not

performing as well uh because people

will switch.

>> Got it. Okay. Got it. Um all right, M.

Well, this is super awesome. Yeah. Um

so, how do you think the design and

maybe like the PM role will evolve like

you know or is already evolving? I mean

I I I guess everyone is just kind of

like a builder with their own specialty

or

>> Yeah, I think that's a great question. I

feel like

across tech, espec specifically if

you're in a coding adjacent role, we're

actually becoming more and more fluid in

our responsibilities if that makes

sense. you know uh historically I used

to talk a lot about the design PM

fluidity like depending on if your

product manager is a lot more design

oriented or your designer is a lot more

product oriented uh as you get more and

more senior or even in like smaller

teams you'll see the responsibility like

really kind of become fluid you'll see

that with EM as well like it'll become a

little bit fluid around who does product

strategy or things like that uh and I

think that's because the skill set of

being more product minded grew over time

but the product manager was still the

expert and the expertise and the

decisionmaker in that uh when it comes

to these coding tools, I kind of see it

very similarly where now we all have

kind of a skill set and ability to code.

So that responsibility is a little bit

more fluid. We can kind of lean into

each other's rules and help each other

out and collaborate a lot more, but the

ultimate decision maker should still be

the engineer and the reviewer should

still be the engineer. So

>> yeah,

>> it's more like fluidity and then an

expansion of our skill set. Now we just

have another tool in our toolkit that we

can reach for which is code.

>> Yeah. And I also think there's also a

very good trend uh around more senior

designers like yourself choosing to stay

in kind of like a builder role because

previously you kind of have to become

like a manager or something like you got

to work on like some management stuff,

right? But like you know I feel like

more and more now people who care about

the product and the craft are more

valued. I don't know maybe I'm biased

but

>> have you seen that? I don't know. I

can't say if I've seen that personally.

I would definitely I would definitely

believe it though. Like I myself

probably spend an equal amount of time

now in cloud code as I do in Figma. And

um you know I talk about this sometimes

with folks at Anthropic. It's just so

fun to build here because you can build

so quickly that it's hard to imagine not

wanting to be a builder. A lot of our

managers at Anthropic are building

themselves too though. Like I think the

manager role also has more access as

well because you can get into these

tools so much quicker.

>> Okay. So like let's say like I'm a

designer and like you know I want to

become like a AI native designer like

Megan and like or like maybe get a job

at anthropic like how how do I

>> get ready?

>> Yeah. Uh okay. So I would say if you're

like totally new to coding and it's

something that you want to get into my

recommendation is honestly to just start

like just start using it because it's

hard to know the unknown known. So you

can uh first start with like claw.ai and

artifacts that will actually give you a

good understanding of like what it looks

like to have a model that produces code.

um you can start and then from there I

would recommend you go into cloud code

directly. Uh it's a really great

partnership if you have an edge partner

to actually collaborate with us on

because your engineering partners will

be excited that you're you're wanting to

code because it'll help them get faster.

You'll learn a lot more. I think it's

like a nice moment to build that

partnership and that bridge. Um

>> a few things that I did is I did take an

intro to React and an intro to Tailwind

class. Most of the changes I make are

front end and it helped me get a little

bit deeper in the understanding of you

know how these libraries work and how

they're rendered uh directly. Not a

requirement but I do think it helps to

understand the code that's being

implemented. Um and then to get into

like the actual vibe coding and the

actual kind of AI side of things. I

would say the community is very open and

collaborative right now. Like we're in

that nice spot of new tech where people

are just excited and sharing what they

do. So, uh, X is a really good place. I

would say GitHub is a really good place

to learn. Uh, we have a few courses

actually that we offer at anthropic that

teach you how to vive code. Those are

really great because there is a little

bit of a skill to learning how to vive

code. That's even different from just

coding like a separate thing.

>> Yeah. Like you know just just like you

know probably number one tip I have for

people is just to like get to plan first

like

absolutely.

>> And also like another tip I have and I

don't do this a lot. It's just like

after it makes something, you just kind

of ask it to explain how you did it, you

know, cuz then you can actually learn

like how what what the code is actually

doing.

>> Absolutely. Absolutely. I think it's

really really helpful to actually get

into the code itself. Uh it also lets

you do more complicated things in the

future and you can detect when cloud is

wrong because sometimes cloud will go in

like a direction that you need to course

correct it a little bit.

>> Yeah. Yeah. When it starts saying you're

absolutely right that that's something

that's going wrong. Yeah. Exactly. Cool.

All right. Well, make it out. Where can

people find you or, you know, learn more

from you?

>> Uh, yeah. So, you can follow me on uh X

I, uh, my handle's a little bit

complicated because my name is spelled

complicated, but it's me E A G HAN H_C.

Uh, I will often share tips and tricks

there on how I use Cloud Code for

nontechnical folks. And I'll also share

some of the launches that are upcoming

and how I use them in a way that's maybe

not necessarily geared uh directly

towards technical developers. I I also

have gotten feedback that it's helpful

for devs to learn there because then

they can see how their nontechnical

peers are using cloud code as well.

>> Yeah, I I love your post like they're

very practical and like yeah I I I think

you know I I posted that this thing like

cloud code is kind of like it's really

just a really good agent. So it can like

right now it's most used for coding but

it can be used for design for product

stuff. Absolutely.

>> Absolutely. Absolutely.

>> Yeah. All right. Well, keep working on

it. Keep working on it. I'm I'm a power

user. Yeah. All right, Megan.

Loading...

Loading video analysis...