How language became a primary design medium for AI products — with a first-hand prompt rewrite on screen
Writing for humans in an AI world ft. Chelsea Larsson (Anthropic) | Config 2026 · Chelsea Larsson
26 min total·Actually worth watching closely: ~9 min·3 must-watch clips
- 0:13 – 3:20Skim
Are writers first to go? Look at what actually changed
She opens on the fear that content design is the thing AI eradicates first, and answers with what actually happened on her own team: a UI copy change that used to run across Figma files and documents, a PM, engineering and localization — about two weeks — now takes 20 minutes tops, straight into the repo as a PR to stamp and ship.
What got compressed is the execution, not the profession.
Mostly setup; the screen is the old flow next to the new one, so a glance at the two timelines gets you the conclusion — no need to follow every sentence.▶ Jump to 0:13Speaker · Chelsea Larsson - 3:20 – 7:14Listen
Moving upstream from shaping strings
With the word-by-word work taken over, the content designer's position moves upstream: no longer the sentence a human reads, but the instructions read by a model doing things for humans. She explains how that changes who you write for and what you ship.
The audience changed from humans to an AI that helps humans, and the job becomes helping the model be a good writer.
Pure narration, nothing to look at — fine to listen to while doing something else.▶ Jump to 3:20Speaker · Chelsea Larsson - 7:14 – 10:48Watch
Wording changes model behavior directly
Two cases, one each way. The memory export request got hedged because it read like a demand letter, then improved step by step once it was reframed as legitimate migration and later restructured as a checklist. On the other side, adding "sarcastic," "witty" and "dry" to the vibe description turned Claude into a bit of a creep — "let me get some context before I interrogate you."
Tone and framing aren't rhetoric; they're functional variables that trip the model's caution training and change the output.
The value here is the prompt text on screen — which words changed and how much the result moved. It's a sample you can copy; listening alone misses exactly the wording that matters.▶ Jump to 7:14Speaker · Chelsea Larsson - 10:48 – 15:00Skim
The capability gap: why nobody uses the feature
The gap between what people think AI can do and what it can actually do widens every day. Connectors are the example: buried too deep for almost anyone to find, even at the moment one would have made sense. Users were asking for the capability in plain language, while Claude — not realizing that connecting a tool was the right answer — built the tools from scratch.
Low adoption is often not a UI problem but a gap too wide — create the conditions for the model to hand the capability over inside the conversation.
The visuals are the click depth and the adoption numbers; catch that percentage and the fix, and you can skip ahead through the narration in between.▶ Jump to 10:48Speaker · Chelsea Larsson - 15:00 – 17:52Listen
Naming is defining a category
Talking to knowledge workers showed they weren't stoked on agents — they didn't want to be a manager to a bunch of uninhabited task loops, they wanted a tool that made them feel more skilled. So the team dropped the internal "Agent Mode" name for one built around working together, and made it the mental model underneath every work feature that followed.
Once a name lands it gets picked up across the industry — which means you've defined users' relationship with AI for them.
This is the research logic behind a naming call, carried by what she says rather than by the slides. Listening is enough.▶ Jump to 15:00Speaker · Chelsea Larsson - 17:52 – 21:10Listen
Cars and infinite scroll as cautionary tales
Two design revolutions held up for comparison: streets widened for cars, and the endless scroll of social media. Neither set out to harm anyone, yet both encoded values that shaped the decades after. She thinks AI design is standing at the same intersection right now.
Once a pattern sets, it spreads into the default — and defaults are all but irreversible.
A historical-analogy stretch; the language is the argument and the slides are only illustration.▶ Jump to 17:52Speaker · Chelsea Larsson - 21:10 – 24:53Watch
Writing human thriving into the goal
First a flyer for having your AI dispatch a human to do a task, as a live sample of a value already spreading. Then a study of 80,000 people: excited about AI and scared of it, wanting to be expanded rather than replaced. The moves she gives: we know what addictive design looks like, so design against it, and build a separate eval for models that just tell you what you want to hear.
Human thriving has to be designed for as an explicit goal; it won't just emerge once the capabilities are right.
The flyer is the most visually striking moment in the talk and is itself the argument; the study slide is worth pausing on to read the spread.▶ Jump to 21:10Speaker · Chelsea Larsson - 24:53 – 26:09Listen
Did I leave room for the person?
It closes on a self-check you can apply to any design: give users a meaningful role — is there still a meaningful role here for a person? Plus: don't forget what you do best. The greetings she wrote by hand in her first couple of months at Anthropic are still the number one thing people reach out to her about.
Writing for humans doesn't go out of date; good old-fashioned product delight is scarcer than ever.
The ending lands on one self-check and one small story; the screen holds only a line of hand-written copy, so hearing the passage in full carries more weight.▶ Jump to 24:53Speaker · Chelsea Larsson