LLM interaction, revised
September 2025
Back in 2021, I was designing AI interactions at AI21 Labs, two years before ChatGPT made everyone think AI meant having a conversation with your computer.
We were building Wordtune, a Chrome extension that helped people write better. Not by chatting about it, but by quietly sitting in the corner until you needed to rephrase a sentence. Click a button, get options, pick one, move on. Like having an editor who only spoke when spoken to.
The One-Button Philosophy
We believed AI should be like a good colleague—helpful when you need them, invisible when you don't. So we built a Chrome extension with one button. Select text, button appears, click for options, select the one that works for you.
The technical constraints were brutal. Extensions have to work everywhere—Gmail, Notion, that weird corporate wiki from 2003. We spent months on making something noticeable enough that people remember it exists, but subtle enough that they don't want to murder it.
It worked beautifully. We had over 2M users, we were generating revenue in an up going pace. Then the AI got smarter.
When Technology Outgrows Its Interface
By 2022, large language models could generate entire paragraphs, not just rephrase sentences. Users expected it. Our one-button solution seemed weak next to the other, open ended, solutions.
We had to adapt and add more abilities to help our users write in a better way, generating text to complement and enrich what has already been written, supporting our "AI as a colleague" philosophy, and avoiding the AI-generated sameness that started to spread.
We knew what capability we need to add, but combining it with our one-button UX was not an easy task.
We explored three approaches in parallel:
Add another button. Now users had two buttons floating near their text. Intrusive, but we needed the discoverability. We knew it's a short-term solution, but wanted to test it anyway.
Dropdown menu. Same highlight mechanism, the the user had to choose what kind of enrichment they wanted - add more data? a joke? trust AI to generate by itself? We'd turned a reflex into a philosophy seminar.
Arrow nudge. Detect when someone stopped typing, show an arrow suggesting we could continue. Problem? Detecting "stopped typing" is hard, and might interrupt the user's focus, sometimes we need to think about the next sentence by ourselves, "writing" over just "typing".
Ship, test, design, repeat
adding a second button felt like admitting defeat. A dropdown menu turned every interaction into a decision tree. A predictive arrow when users stopped typing was technically complex and often felt presumptuous.
The solution we landed on: one button that revealed options on hover. Rewrite or generate, and refining the predictive arrow interaction to perfection.
The craft was in making this feel inevitable rather than clever. We spent weeks on details nobody would consciously notice—the hover delay (180ms, not 200), the easing curve, how to present the sources our LLM based his generations on.
This is what people miss about AI interface design: the AI does the magic, but the interface makes it feel seamless, easy. Every pixel, every millisecond shapes how people perceive the intelligence as an extension of their mind, not just for chrome.
LLM interaction, revised
September 2025
Back in 2021, I was designing AI interactions at AI21 Labs, two years before ChatGPT made everyone think AI meant having a conversation with your computer.
We were building Wordtune, a Chrome extension that helped people write better. Not by chatting about it, but by quietly sitting in the corner until you needed to rephrase a sentence. Click a button, get options, pick one, move on. Like having an editor who only spoke when spoken to.
The One-Button Philosophy
We believed AI should be like a good colleague—helpful when you need them, invisible when you don't. So we built a Chrome extension with one button. Select text, button appears, click for options, select the one that works for you.
The technical constraints were brutal. Extensions have to work everywhere—Gmail, Notion, that weird corporate wiki from 2003. We spent months on making something noticeable enough that people remember it exists, but subtle enough that they don't want to murder it.
It worked beautifully. We had over 2M users, we were generating revenue in an up going pace. Then the AI got smarter.
When Technology Outgrows Its Interface
By 2022, large language models could generate entire paragraphs, not just rephrase sentences. Users expected it. Our one-button solution seemed weak next to the other, open ended, solutions.
We had to adapt and add more abilities to help our users write in a better way, generating text to complement and enrich what has already been written, supporting our "AI as a colleague" philosophy, and avoiding the AI-generated sameness that started to spread.
We knew what capability we need to add, but combining it with our one-button UX was not an easy task.
We explored three approaches in parallel:
Add another button. Now users had two buttons floating near their text. Intrusive, but we needed the discoverability. We knew it's a short-term solution, but wanted to test it anyway.
Dropdown menu. Same highlight mechanism, the the user had to choose what kind of enrichment they wanted - add more data? a joke? trust AI to generate by itself? We'd turned a reflex into a philosophy seminar.
Arrow nudge. Detect when someone stopped typing, show an arrow suggesting we could continue. Problem? Detecting "stopped typing" is hard, and might interrupt the user's focus, sometimes we need to think about the next sentence by ourselves, "writing" over just "typing".
Ship, test, design, repeat
adding a second button felt like admitting defeat. A dropdown menu turned every interaction into a decision tree. A predictive arrow when users stopped typing was technically complex and often felt presumptuous.
The solution we landed on: one button that revealed options on hover. Rewrite or generate, and refining the predictive arrow interaction to perfection.
The craft was in making this feel inevitable rather than clever. We spent weeks on details nobody would consciously notice—the hover delay (180ms, not 200), the easing curve, how to present the sources our LLM based his generations on.
This is what people miss about AI interface design: the AI does the magic, but the interface makes it feel seamless, easy. Every pixel, every millisecond shapes how people perceive the intelligence as an extension of their mind, not just for chrome.