UX.md

Happy Friday! Tony Alicea proposes a UX.md standard. UX context for agents, living alongside design.md, containing things like research synthesis, a user model, a glossary and more.

I think that's a great idea.

A lightweight standard format can encourage experimentation and tool creation, even if it's wrong and will change and all that jazz.

What would a lightweight ux.md file in your situation look like today?

Number of the week

85%

AI agents can spend much more time verifying than creating, which seems like it would matter for UX too. 85% is the share of the 11-day Bun-to-Rust rewrite that went to verification rather than to writing the code. How building software is changing at Anthropic

Quotes

"The writing assignments I give my students are gym tasks, not work tasks." Bruce Schneier.

"Years ago, we didn't have SQL. There were people whose job was to generate software that would query large data sets. Their job title was COBOL programmer. Then SQL comes along and it gives you this convenient way so people could just specify. That didn't mean programmers went away. It just meant the job changed a little bit." D. Richard Hipp, via Simon Willison

"How do I make soup with the fire, not just look at the properties of the fire." Elena Yunusov.

"We spend so much time making things frictionless, but by making it frictionless we make it forgettable." Molly.

👀 Interesting this week

UX-Context Design: Using UX Knowledge to Inform AI-Generated Design, Nielsen Norman Group. Tony Alicea proposes the deliverable: a UX.md file holding research synthesis written as constraints, interaction standards, a glossary, user models and world models, sitting next to DESIGN.md (Google Labs open-sourced that spec in April).

"A persona has a stock photo and a first name with the intention of helping a human to empathize and remember. However, a model does not need persuading, it needs the underlying reasoning."

Information architecture is the foundation artificial intelligence is starving for,

"You don't have an AI problem. You have an information architecture problem wearing an AI costume."

Your people get AI. Get out of their way. Dan Maccarone puts two conversations side by side: a head of UX at a hundred-year-old manufacturer still waiting on IT to approve the tools her own PMs are already shipping with, and an engineering lead at a five-person team who defined the founder's job as a filter that removes the small stuff.

"They adapted so fast and so quietly that you didn't even see it, and the reason you still can't is that they've already learned it isn't safe to show you."

A fundamental flaw leaves LLMs strikingly vulnerable to attack.

"It's just one big sheet of tokens."

Rethinking friction in an AI world. Darren Yeo's team built a parents' communication app to be completely frictionless. Adoption spiked and so did support tickets, because parents lost attachments and missed details in the infinite scroll. They added friction back (truncated text you have to tap to expand, interaction pauses on attachments), time in app went up, and so did satisfaction.

Investigating three real-world incidents in our cybersecurity evaluations, Simon Willison on Anthropic's post.

A big week for AI denialism.

"Who cares? It's just marketing. Who cares? They're just doing what they were programmed to. Who cares? It's not like they're sentient."

For AI adopters, success and failure look identical, at first. Two historical archetypes: the bounded adopter (NYSE automated order delivery but kept human specialists on the floor and rejected a fully electronic book in 2000, then bought Archipelago for $9B in 2006; Borders let Amazon run its e-commerce for seven years and never built the muscle) and the project accumulator (1980s GM at $10B a year across robots, Saturn and NUMMI, where NUMMI outperformed every other GM factory and the learning still travelled too slowly to matter).

The accessibility paradox. Michael Buckley traces LAION-5B back to alt text: Common Crawl images paired with the descriptions people wrote to make them accessible, which then went on to train Stable Diffusion.

"The blind web, quite literally, helped teach AI how to see."

ICYMI 2026-07-25: Define 'Best', Jorge Arango.

"'best' is meaningless until you say best at what, on which cases, at what cost."

AI Worming through Word, Simon Willison.

The Parts Around the Task, Saeideh Bakhshi. Self-checkout, delivery, Waymo, talks at 2x, asking AI the question she'd once have asked a colleague. She's an introvert who used all of it deliberately, and she notices that almost none of her favourite memories came from time alone.

"They were not the task. They were just what happened around the task."

A Concrete Definition of "Product Sense" (and How to Build It).

Anthropic's first technical PM on token maxing, the jagged edge, and living in the future, Lenny's Newsletter.

The UX + AI MCP: 101 Methods for Designing and Building with AI, Ileana Marcut. 101 methods across research, IA, UX writing, critique, design systems and designing for AI, shipped as an MCP server.

The consciousness mirage in AI design, Slava Polonski. How much mind a product invites users to attribute to it, treated as a design dial rather than an accident.

The seams are where the system lives, Takuma Kakehi. Splitting a system into clean modules doesn't remove complexity, it moves it to the seams. Cities learned this through zoning.

Quoting Boris Cherny, Simon Willison. Buried on page 73 of the Opus 5 system card: it's the least prompt-injectable model Anthropic has shipped, across both PI evals and red teaming.

Your website is boring (but that might just set it free). Sam Belt.

"Nobody falls in love with a warehouse."

Battling AI fatigue as a designer and developer: a practical guide. Adir SL.

"For me writing is thinking and I don't want to outsource it to a machine."

The UX of Contrast: Lessons from Indika's Game Design. Alisa Smelkova on a game where lighting a candle in prayer returns a bright retro chime, the kind of sound that says "you did it" in Duolingo rather than "you were heard" in a church, and where the faith points you spend the whole game collecting unlock nothing.

The MOS 6502: the people's princess, Neel Dozome.

The aggressively mediocre fight, Chris R Becker.

The AI Hype Index: Unsexy AI, MIT Technology Review.

Claude Opus 5 review: this model is brilliant (but annoying), Claire Vo for Lenny's Newsletter. Her framing: we've hit an intelligence overhang, so the variables that matter now are personality and verbosity rather than raw capability.

Health and happiness,
Peter