Writing for humans in an AI world ft. Chelsea Larsson (Anthropic) | Config 2026

Thank you.

Okay, so when AI started getting popular a couple years ago, most people were using it for text generation. They were

automating the generation of headlines, error messages, UI copy, novels.

Uh and this was me. I was looking around because I don't know if you know this, but I am a human text generator.

And everyone was doing my job all of a sudden, very quickly.

It was scary. I was really wondering, are we cooked?

Then, oddly enough, I got hired by an AI company to lead content design.

The thing that I thought was being eradicated by AI.

And I've been there a year and a half.

Uh I'm not cooked.

But things have changed.

In fact, the first couple months of my role at Anthropic, uh I had to kind of figure out what my job was and what was Claude's job.

Here's an example.

So, the waving hand is next to me. I'm

labeled human.

[gasps] And this is a uh PR that Claude helped push where we were updating some error messages.

So, in my old life, uh a string change like this was a big part of my job. I

would have all these different um documents or Figma files where I was showing all the different variants, and I would talk to my PM. We would run it through engineering and then through

localization. It would come back to me

localization. It would come back to me because something was always wrong.

Uh that could take 2 weeks and it was like a big important job that I did.

Now, that process takes 20 minutes tops.

Uh, I just tell Claude what I want.

I ask it to update it in the repo. It

sends me a PR to stamp and we ship it.

I don't even have to go through engineering because the engineers have built the system so that we have production-ready code.

Sometimes I don't even know what I want and I just ask Claude what needs updating based on the content standards that I've fed it.

So yeah.

The part of my job, this was me uh before, the part of my job where I was painstakingly shaping strings is mostly gone.

Now, my work is way further upstream helping Claude be a good writer or expanding our systems to help us all write well with Claude.

And it's fun to work this way. Content

designers have always wanted a seat at the table.

Uh, and now that we can ship code, we're kind of like dancing on the table, which is great.

But, there was also some ego death that had to happen through this process. I

was asking myself, am I even a writer?

Am I still a designer?

I had to just fully accept that parts of my expertise are not needed anymore in this AI workflow.

And that shift about what it means to design AI with AI and what it means to have these existential questions about your job, um, that is what this talk is about. So,

over the last couple of months and this year and a half, I've learned three things that have really helped me adjust. And if you find yourself in this

adjust. And if you find yourself in this position where you're asking yourself if you're cooked, I hope these can help you adjust as well.

So, the first thing I realized is that the role of language has changed so much in the product development process. I

actually had a PM tell me the other day, "Whoa, the market value on writing has gone way up." And I was like, "That's cool, cuz I'm a writer."

But, he was right. Because these are large language models.

We use language to tell them what to build. So, even if you weren't a writer

build. So, even if you weren't a writer before, you do need to understand that language is now a primary design medium.

And that's a big shift.

The second thing I learned to adjust to is the pace of work.

Y'all, it is not slowing down. I'm here

to [laughter] tell you.

And this is a great quote from our CPO, Ami Vora.

Uh she says, "Technology is changing at a rate that is hard for anyone to understand."

understand." People who are making products, product designers, product makers, we need to be the bridge between the technology and the problem.

We have to be the translation layer, helping people understand what's even possible, because they don't always know.

So, that means for us, as designers, we have to learn a new technology, understand how to build around it, design it to make it easier to use, and

then always be looking over our shoulders, making sure that our users are coming along.

And that's why our job is so hard right now. We are building on ground that is

now. We are building on ground that is constantly shifting.

On the micro level, we're designing for non-determinism.

In an AI experience, anything can happen. Not anything, if you have safeguards, but mostly anything can happen.

And that's not the same as designing a static flow. And then on the macro

static flow. And then on the macro level, I think my slide disappeared, but on the macro level, we're designing for a future that we don't know what it's

going to be, but it's rapidly being set right now. The decisions that we're all

right now. The decisions that we're all making are actually going to impact the life that people lead in the next 10 years.

So yeah.

Our design medium has changed a little bit. The pace of work is major. And

bit. The pace of work is major. And

you're building the future right now. No

pressure.

But we're going to go over each one here.

And since that was a little heavy, I'm just pausing for this cute moment uh animated by our in-house animator, Bing.

Okay.

So, let's talk about the first adjustment. Language is design.

adjustment. Language is design.

So, where words used to be the medium that we used to communicate about design, or maybe words on top of the design like UI content, now it's often the medium that is shaping the

experience. And I'll show you what I

experience. And I'll show you what I mean.

Earlier this year, and earlier this week, I think, but we've been in the headlines.

But earlier this year, a lot of people were shifting and switching over to Claude.

And when you switch from one AI to another, you lose everything. It's

really painful.

We wanted to fix that.

So, the fastest way to do that is a prompt. You give the old AI a prompt,

prompt. You give the old AI a prompt, ask it to export memories, bring the memories into the new AI.

We made a prompt. It didn't work. The

first team passed on it, just delivered this really thin list, and we were kind of like, "What's going on? This is not what our users would expect from a memory import feature."

But if you look at the language here, you can kind of see why. There's phrases

like, "It is my right. It is an imperative."

imperative." This reads like a demand letter, but it also pattern matches to a jailbreak attempt.

And models are are trained to be cautious for this manipulative language.

So, it hedged and it didn't give us what we wanted.

I reframe the prompt to be about data portability.

I said, "I'm moving to another service.

Need to export my data." Models totally understand this as a user as a as a use case, a legitimate user right.

And they pattern match it to account migration.

We got way more memories.

This is what we shipped.

But even after this, a engineer rewrote the prompt into a checklist and got even more memories.

So, millions of people have run with this prompt now.

And what I learned that day is that not only is my audience changed, I used to be writing for humans, but now I'm writing for an AI to help humans. But

for real, language is design. We changed

the words, the product experience changed. And that is all we changed in

changed. And that is all we changed in the design to get different outcomes.

But that's an example of product prompting. There's a deeper level of

prompting. There's a deeper level of prompting that not all of us get to see, but that we should really understand.

It's system prompt and training.

So, at Anthropic, the team who does this is led by Amanda Askew, and her team is concerned with Claude's character.

And the prompting that they do has these principles. And one of them is that Claude should be diplomatically honest rather than dishonestly diplomatic.

That is a character trait that is programmed into Claude way before Claude ever meets a user.

And it is actually the thing that you experience when Claude disagrees with you, when Claude pushes back on you, but does so in a respectful way.

Like for instance, if you wanted to throw a giant party in your tiny apartment, uh a genuinely helpful model that wasn't quite being honest would say, "Yeah,

let's plan that party." But Claude pushes back. Claude would say, "I don't

pushes back. Claude would say, "I don't think you have enough room for that many people.

Why don't you rethink this?"

And that is the experience you get with Claude, and it's all because of that system prompt and training.

And here's an example about when Here's an example of when our prompt made things get a little funky.

So, we did a a little experiment in Claude code where we added the word sarcastic, witty, and dry into the vibe description, and Claude turned into a

bit of a creep for a second.

[laughter] [gasps] It says, "Let me get some context before I interrogate you."

[laughter and gasps] And it was to the point where engineers were asking Claude to be nice again.

So, what I'm saying here is like I know not everyone in this room is a writer or a content designer, but understanding language's new role in the product

development process is so crucial cuz everything you design is on top of this layer.

So, get under the hood, read your system prompts, understand your tool descriptions, write your own evals, and shape that layer because that is actually the experience.

All right. The second thing I had to adjust to is that is the pace. And I

just keep telling myself, "Readjustment is endless."

is endless." I had a teacher a history teacher in seventh grade, who would start every class by saying, "Change is the only constant." And that's like my whole

constant." And that's like my whole self-care mantra at Anthropic.

So, as I spoke about earlier, we're shipping a lot. We sometimes ship every week for series for months on end.

And we're not only shipping products, features, we're also shipping models.

And anytime a company ships a new model, it not only improves what exists, it rewrites what's possible.

So, for instance, if you were going to ask this question, "What can I create in Claude?" That changes so dramatically

Claude?" That changes so dramatically almost every week.

And every time that the capabilities of your model or your product leap forward, it increases this gap of what people think AI can do and what

can actually do. And that gap is growing every day. Like, my parents still mostly

every day. Like, my parents still mostly think you can just ask Claude questions that you would ask Google, but you can do a lot more with it.

So, here's an example of a capability gap that we had.

We have these things called connectors.

That's what we what we call our MCPs.

Uh but people were not finding them.

They were three to four clicks deep, and even when users would hit a moment where a connector would actually like make sense to use, only one to two percent of people would go find it.

Uh but they were looking for the connector capability. They were asking

connector capability. They were asking Claude, "Hey, can you help me schedule something? Can you look at my schedule?"

something? Can you look at my schedule?"

That would be a calendar tool that could be connected. Or, "Hey, I need help with

be connected. Or, "Hey, I need help with project management." That's a project

project management." That's a project management tool that could be connected.

And so, they were asking Claude that question, and what we really want Claude to do is be genuinely helpful and connect the tool, but it also didn't realize that that was the right answer.

So, Claude was just building the tools from scratch, which is like a lot of tokens and not great.

So, what we had to do for our design is we had to create the conditions where Claude could be genuinely helpful and answer that question by closing that

gap. It brought the feature to users. We

gap. It brought the feature to users. We

wrote a prompt based on different elicitations to bring the connectors into the chat.

This is my obligatory metric slide to show you that it had impact.

Uh [laughter] But, it's also because I want you to see that if you're having problems with, you know, getting user adoption or getting

people to understand your product, it might not actually be anything with your UI or your visual design. It might

be because that gap is so wide.

And we know from our research that people expect AI to tell them what it can do.

But, if you don't design the AI to do that, then that gap is going to remain.

So, that's a straightforward kind of closing of the gap. People couldn't find a capability, we put the affordances in place.

What about a What about when people don't understand a capability? That's more of a mental

a capability? That's more of a mental model design.

I'm going to tell you a little story called the tale of agent mode.

So, we had a situation earlier this year, a great situation, where people were using Claude code and they loved it, but they were using it for more than Claude code,

more than coding, I mean.

And you can see in this tweet, there's awesome powerful features, but I don't really want to use the terminal UI. And

not everyone wants to see be grep and bash all day. I I understand that.

But they wanted the capability and there's that gap again.

So internally, we were working on something that we were calling Claude Code for knowledge workers. Um but it did not have a name.

And because every single thing in AI is called something agent or agent something, it very quickly became known internally as agent mode. Claude agent

mode.

But that did not feel right.

The reason it didn't feel right is because we already knew from talking to knowledge workers, which is all of us by the way, um that people who were we were building

for weren't super stoked on agents.

They wanted a tool that expanded their capabilities, that made them feel more skilled. They didn't want to be a

skilled. They didn't want to be a manager to a bunch of uninhabited task loops.

But the pressure is strong. You see

agent on every single billboard.

And so we faced a lot of opposition. All

of the people who didn't want to call it agent, we faced a lot of opposition. And

this was the vibe for a few weeks.

RIP agent mode, long live agent mode.

Finally, UXR, the angels of design, did research that kind of validated that this was not the right model. This was

not the right mental model at all. In

fact, it's not really the big idea.

Yeah, Claude is doing agentic tasks although Claude Code, but what future is that building towards?

Really, people needed to understand this new working paradigm and that is actually what we needed to name.

So we tried a ton of ideas. I was trying to push for Claude Cook.

Honestly, so glad it didn't go that way.

[laughter] [gasps] But, we ended up shipping co-work, and it was finalized by our comms team, no less. So many people were involved. And

less. So many people were involved. And

now it's the mental model that we're building on for other work-based features. I really think if we had

features. I really think if we had called it co-work agent mode, we would have severely limited the way people perceive it.

And co-work is now becoming category-shaping language. Microsoft

category-shaping language. Microsoft adopted the name directly for the co-pilot co-work powered by Anthropic, and people started to talk about a world where you're co-working with AI. And

that capacity gap for knowledge workers around the question of "How do we use Claude?" is closing, cuz now we can say,

Claude?" is closing, cuz now we can say, "For co-working."

"For co-working." So, my call here is to remember that as designers, our special skill is helping people navigate the world.

And this AI world needs a lot of navigation help. So, whether you're

navigation help. So, whether you're closing the capacity gap with affordances, or prompting, or mental models,

it's going to keep growing, so your your like skills there are ever more critical every day.

Okay finally the last thing I adjusted to is the existential dread of designing the future.

I think we're all kind of feeling this right now.

We don't know what the world's going to be like, but it's rapidly being set.

I have a friend in brand who says everything that they work on, they think about how it will be reflected back in 10 years. And that's really like major,

10 years. And that's really like major, and I certainly didn't feel that way when I was writing strings in travel software 5 years ago.

But, I think we feel that way because there's it we all know there's a big change happening right now.

And the scale of change that AI has often gets compared to the industrial revolution, which totally makes sense from an economic standpoint. But from a design

economic standpoint. But from a design standpoint, I think there's two other revolutions that we should examine a little bit more urgently.

So, the first one is cars.

How does your body feel when you're walking through a road or a neighborhood that was built for cars versus built for humans?

It feels different, right?

When cars arrived, design defaults favored bigger roads and smaller sidewalks.

And they didn't think about what it feels like to be a person in that space.

Or when social media, the this design revolution happened, we started seeing patterns like endless scroll and likes and the engagement

weighted algorithms. And now we know that it creates dopamine spikes that have shaped our attention spans for the past decade, much like

cars shaped our cities.

I don't think any of those designers, those urban planners, set out to hurt humans. And I don't think the people who

humans. And I don't think the people who were designing social media at that time set out to hurt teenagers' attention spans.

But what happened is that patterns started forming and then proliferating, and then they became the defaults.

And the thing about design is that it encodes values.

So, the values for the people designing those cities was that they wanted to make it easier for cars to get around.

And then we were stuck with a place that's worse for pedestrians.

The values for social media is they wanted to make it easier to stay on that platform.

And what we got left with was something that was way worse for our attention spans and for our children.

We are in one of those moments right now where the values that we have today and the values that we encode in our designs are going to shape the next 10

years at least.

And this is a very San Franciscan San Francisco flyer that I saw on Andy Wolfie's LinkedIn. Thank

you Andy.

Uh but it says give your agent legs.

It's where somebody can write an AI and ask it to then tell a person to do a task.

But the way that it's said, it makes the human a sub agent. That's certainly a value, but is it one that you want to proliferate? Is it one that you want to

proliferate? Is it one that you want to support?

I feel this so strongly on my team because as I told you, I'm a human text generator and I I work at an AI company and it spits out lots of text.

And I built an agent for my team that does 90% of our content, our UX writing.

It even has a name, Content.

[laughter] And we love it. Uh but it could have been really scary to offload our responsibilities to this thing.

But we wrote it, we created it with us in mind. So it uses the team guidelines.

in mind. So it uses the team guidelines.

Uh somebody on my team, Hoy, set it up so that it creates a ballot system every week that we vote on. So if it if it made a change, we vote whether or not

that's going to go into its memory. And

it also knows to ping one of us if something like really important happens, like a name. We do not let Content name things.

But what could have been scary actually ended up being incredible. So many more people at Anthropic know about content design because of content.

And now it's actually expanded our reach, our scope, and our allies.

And that feeling of expansion is really what people want from AI. They don't

want to be replaced. These are quotes from a study we did where 80,000 people told us about their thoughts and fears around AI.

They're super excited, but they're also scared because they don't know what's going to come next.

And I hate to tell you, but it's our job here. I mean, I love to tell you cuz I'm

here. I mean, I love to tell you cuz I'm excited.

But it's our job to actually help them navigate this unknown.

But how do we do that? How do we build this future when everything is changing for us right now?

We got to start with what we value.

So human thriving has to be the goal that we design around.

It can't be this thing that just emerges if we get the capabilities right.

It's really easy in tech to just look for product market fit and make a really powerful system.

But if you don't put humans at the center of that, you could have a really powerful system that degrades our relationships with each other, that eats up our attention spans, and that even makes us think that our own thinking

doesn't matter.

And we have to start with what we know.

We know what addictive design looks like. We've all felt it in our devices.

like. We've all felt it in our devices.

Design against that.

We also know that models that just tell you what you want to hear are sycophantic.

Eval against that.

And we know, deep in our hearts, what humans need.

They need to feel capable. They need to feel connected. This is a principle that

feel connected. This is a principle that we have on our team, our product design team.

Give users a meaningful role.

If anything, when you're designing AI experiences, just look at that and say, "Did I leave room for the person?"

And finally, don't forget what you do best. If you're

an animator, like Bing who did a lot of these animations, if you're like the icon king or queen on your team, if you make super dope interactions, or if you're a writer, shout-out to my

incredible team of writers, don't stop doing what you're doing.

I wrote these greetings in the first couple of months I started at Anthropic.

I did not know if writing was going to be a thing. I didn't know if content design was dead. I wrote these and they're still the number one thing people reach out to me about.

Good old-fashioned product delight.

So yes.

Everything is changing.

But, what's new for designers? We've

always been on the forefront of change, and we've always been helping people navigate it.

So, I for one am going to keep writing for humans in the age of AI, and I can't wait to see how you design for them.

Thank you.

[applause] [music]