Design has always had a strange relationship with its artifacts. The artifact is what everyone can see, point at, argue over, approve, or reject. But the artifact was never the whole job. The job was deciding what the artifact should make clear, what it should refuse to do, which tradeoffs it should hide, and which ones it should force into the open.

I’ve always thought about designers as the translation layer, as the people who connect businesses with audiences through work that makes sense to both. Concept was product strategy before product strategy had a name. The tools change. The mediums change. But the work remains the same.

What’s happening now is where that translation has to live. MC Dean argues that if an agent is composing the interface, designers have to prepare materials an agent can play with: components, rules, priorities, and plain-language intentions. Fixing the generated screen becomes evidence that the setup is missing something. The better question is what the system failed to understand.

Quinn Keast gets at what changes when design moves out of the hand. Designers are used to thinking through gesture: physically and visually moving hierarchy, spacing, and attention around until the thing starts to feel right. The instruction loop asks for a different kind of attention. You have to translate innate design judgment into language before the tool can act on it. That may get easier with practice, but some of the cost is built into asking language to carry hand-knowledge.

Rasmus Andersson’s Figma story gives this argument a useful warning label. Figma looked fast from the outside because people inside were doing slow deliberate work underneath. One person might work on something for a year, and sometimes it went in the trash. Brad Wrage makes this concrete at Cash App, where a designer merged 25 pull requests across Android, iOS, and server. I read that as a last-mile ownership story. The old handoff hid too much of the last mile from the person responsible for the experience.

The same pressure shows up in less glamorous places. Jason Cyr puts clarity into the operating model: decision ownership, readiness, context flow, and the threshold for good enough. Fabrizia Ausiello uses Apple’s Liquid Glass slider to ask which decisions belong to the designer, because pushing judgment onto users can look like empowerment while simply handing them unfinished work. Alex Harper’s commodity-web argument adds an economic angle: when anyone can prompt a polished page, the pretty page loses economic value.

So the setup is now the design: the design system an agent can read, the brief it can obey, the refusal rules that keep it from being too clever, the review channel where a person still says no. That is less romantic than nudging rectangles around a canvas. It is also closer to what design has always been. When the next screen appears at 9:41 p.m., after an agent made a thousand tiny choices without us in the room, the invisible hand of the designer should still be there: guiding the work, constraining the nonsense, and making sure the product still feels like someone responsible and human touched it.

What I’m Consuming

Leonardo da Vinci’s Notebooks Are Whole Again, 400 Years After a Collector Cut Them Apart. Anastasia Scott tells the strange story of Leonardo’s notebooks being cut apart in the late 16th century, splitting drawings and notes that were meant to live together. The new Leonardotheka archive digitally reunites roughly 3,500 manuscript pages, including 50 confirmed reconstructions. A wonderful reminder that archives are not only about preservation; sometimes they repair old editorial violence. (Anastasia Scott / Discover)

SHRTCTS. This site turns keyboard shortcuts into a 3D illustration you can poke at instead of another cheat sheet you immediately forget. It’s like the Mac’s old Keycaps desk accessory on steroids. Pick an app, hover an action, and the exact keys light up on the keyboard. (DRANIKI)

The hidden pattern behind successful products. Lenny Rachitsky talks with Mark Pincus about the Proven, Better, New framework behind Zynga’s biggest hits. Pincus’s useful provocation is that founders often overvalue the new part and under-study what users already understand. His line that your instincts are usually right while your ideas are usually wrong is a brutal but useful product lesson. (Lenny Rachitsky / Lenny’s Podcast)

Web Browsers on Video Game Consoles. Declan Chidlow catalogs the odd, constrained history of console web browsers, from the CD-i and Sega Saturn to the PSP, Wii, and Xbox 360. I have a soft spot for this one because I worked on redesigning Sega.com for the Dreamcast browser in 2000, and these machines forced web designers to think about TVs, controllers, memory limits, and wildly nonstandard browsers. The web was never as uniform as our nostalgia pretends. (Declan Chidlow)

“Don’t You Just Upload It to ChatGPT?”. Juliette Giannesini turns a gym-locker conversation into a sharp little defense of professional translation. The funny part is the HR director suggesting ChatGPT for translation, then admitting she cannot use AI at work because it is not reliable enough. The serious part is the same one designers are dealing with: tools help, but judgment is still the job. (Juliette Giannesini / Correr Es Mi Destino)

There is no doubt that designers are in the middle of a transition. When I went to design school in the early 1990s, we were still learning traditional production techniques: drawing on vellum with hairline Rapidographs, cutting rubyliths, rubbing down halftones, and gluing photostats. The industry itself was in the midst of a transition from physical techniques like these to personal computer-based ones brought on by the Mac, Pagemaker, and PostScript.

My production teacher, the very eagle-eyed Bonnie Russell, would look at our mechanicals with a loupe—yes, a loupe!—and mark any slip of the hand. I’d often receive assignments back dotted with red circles.

However, school was the last time I pasted up anything by hand. After I graduated, the graphic design world had already settled on doing things digitally. I applied the same detail-orientation to my Illustrator and QuarkXpress files instead, so Ms. Russell’s scrutiny wasn’t wasted.

The point being there was a decade-long transition. Designers learned these newer digital production processes as they were being proven to work, as they were becoming the standard.

Nearly four years after the ChatGPT moment (GPT-3.5 in November 2022), and over a year after Lovable took off, designers are wrestling with some existential questions that are beyond techniques. When we step back and look at what designers actually do beyond pixel-pushing, AI is eating at those tasks. Some of it is tedious and we don’t want to do them anymore, like tagging concepts in interview transcripts. But sometimes, the tedium is how we internalize and get better. I would not be as detail-oriented or as much of a pixel fucker if it weren’t for Ms. Russell back in design school.

This transition, this unsettled moment causes a form of grief. We mourn what we’re losing, but we’re also feeling stuck. Jack Maguire believes it’s because there’s no fixed endpoint to grieve toward. When a person dies, he points out, the loss is finite and you eventually adjust to it. AI displacement offers no such resting point: retrain into whatever looks safe now and you may watch it automated inside two years. Workers, he writes, are being asked to “accept a process rather than an outcome.”

You can see the same instability in what we even call the work. Sarah Gibbons at Nielsen Norman Group points out that “AI design” has already forked into four different jobs—using AI in your workflow, designing AI products, designing for agents, shaping how a model behaves—and that everyone says the phrase while picturing something else entirely. Even those categories come with an expiration date. She figures the window to build rare expertise in any of them closes within a year, once the rest of the field catches up.

And the titles keep multiplying. Nicole Alexandra Michaelis catalogs the new business cards—Agentic UX Architect, Trust Designer, AI Design Consultant—and I’ll admit I’m wary of the whole genre. I’ve watched titles and labels proliferate into pretentiousness before. The constant she points to is: the strategic judgment underneath is the same craft it has always been, whatever fancy title we print on the card.

Even “developer” has come loose. When a veteran Mac journalist like Jason Snell can ship a working app in an afternoon with Claude Code—”ugly and incomplete,” by his own description, but working—the word stops meaning what it used to. He found the code was the easy part; envisioning and deciding the app was the actual work. The skill that defined the role turned out not to be the role at all.

So if even the definitions and the titles keep moving, what do you actually stand on?

The closest thing to an answer came from Amber Bouabdallah’s piece. Her Salesforce team set out to design how designers master AI and found there was no curriculum to build: no fixed competence to train people toward, because the tools arrive faster than any best practice can be written down. What worked instead was smaller and simpler—designers watching each other work, catching the workarounds someone tries when the tool behaves unpredictably. She says that peer layer “isn’t the bridge. It’s the ground.” There’s no permanent training to cross over to, so the people around you become the thing you depend on.

That answers the question I keep hearing from designers on my team, the one underneath the title anxiety: what’s the safe thing to become? Maybe there isn’t one. The durable thing was never a role or a tool or a name on a card. It’s the practice you carry between them: your judgment and the specific way you bend a tool toward how you already think.

I’ll leave you with a quote from an interview with Tommy Geoco I linked to a few weeks ago (edited for clarity):

It is chaotic. And I think there’s a lot of value in the people, the teams, the individuals who are just sharing their workflow. And I think it’s important to remember that’s a good activity right now. And you don’t have to attach the language to it like, “Hey, here’s my workflow. You should do this.” Strike should from the vocabulary right now. And just compare. Compare workflows, compare shapes. Because we will eventually converge around some better practices around this stuff, but we don’t know what that is yet. And so right now, the best thing you can do is if you want to get involved in this type of work to just try to find workflows that are good for you and spend some percentage of time on the familiar stuff that gets the work done. And then some percentage of the time, make room for yourself to explore a workflow. We are going to converge around stuff. If you have the luxury and you don’t have to get involved yet, you can wait. And there probably will be a concentration of, “These are the better workflows.”

What I’m Consuming

Sweet Jeebus, MacOS 27 Golden Gate Removes the Dumb Icons From Menu Items. macOS 26 Tahoe stuffed an icon next to every menu item, and John Gruber was far from alone in finding the result cluttered and inscrutable. Golden Gate quietly pulls them back out and rewrites the Human Interface Guidelines to push for sparing, purposeful use. A small reversal, but a reassuring signal about the design direction in Cupertino. (John Gruber / Daring Fireball)

From the Collection: Emigre Font Development Files. The Letterform Archive has digitized Emigre’s complete font development files, a world’s first for a digital foundry’s process work, and Eve Scarborough reflects on what the marked proofs, revisions, and reader mail reveal. The files catch type design mid-transition between analog and digital, back when Rudy VanderLans and Zuzana Licko were chasing the promise of the early Macintosh. VanderLans’s aside about how design debate has scattered across too many online channels to follow is a quiet bonus. (Eve Scarborough / Letterform Archive)

“Music venues are the engine of the creative industries”: lessons from the graphic ephemera of our lost clubs and pubs. The V&A’s Lost Music Venues exhibition treats flyers, tickets, setlists, and disco balls as cultural records worth saving, documenting the independent clubs and pubs that have vanished, especially since Covid. Curator Harriet Reed argues these rooms are community infrastructure, not just nostalgia. The graphic ephemera turns out to carry the memory of a whole creative ecosystem. (Paul Moore / It’s Nice That)

AI and Libraries, Archives, and Museums, Loosely Coupled. Dan Cohen makes the case that cultural heritage institutions don’t have to choose between ignoring AI and surrendering their collections to mass training. His alternative is Anthropic’s Model Context Protocol: a non-extractive bridge that lets libraries, archives, and museums connect their collections to AI tools on their own terms, with ground-truth checks and local-model deployments. It’s the rare AI-and-institutions piece that offers a concrete, controllable path instead of a warning. (Dan Cohen)

Startups are science. Jonathan Yagel, a self-described English major who found measurement distasteful, recounts how Eric Ries’s The Lean Startup rewired his thinking: a real startup’s first job is learning under extreme uncertainty, long before growth. His fix was to run marketing like the scientific method, actually writing down the hypothesis and reviewing the results afterward, which almost nobody does. A conversational read on treating early-stage work as structured inquiry rather than guesswork. (Jonathan Yagel / Age of Intelligence)

For years we put a lot of design judgment upstream in Figma files, then hoped the product would survive translation. That made sense when production was expensive. It makes less sense when the next version can appear in minutes and quietly break the last thing you liked.

Fulya Lisa Neubert, Senior Product Designer at Slack, gets right to the material problem: keyboard-driven search can’t be fully judged in a static prototype. You need to press Tab, watch focus move, see content reflow, and feel whether the behavior is right under your hands. Her shift into the live browser reads as design moving closer to the medium where the experience becomes real.

That same pressure shows up in the Figma workflow examples. Emma Webster, writing in the Figma blog, describes teams testing interactions in code before choosing a direction, then bringing the work back onto the canvas with design-system context intact. A working model becomes the first serious product question: does the idea survive data, logic, motion, permissions, and system constraints? A flat mockup can still be useful, but it can also make teams overconfident about behavior they have never actually felt.

This is why Claude Design interests me more as a process story than as a tool demo. Dan Carey, product manager at Anthropic, describes a team that skipped PRDs and used prototypes to decide what was worth building. They shipped incredibly fast, then removed fine-grained controls after usage showed that vocal power users were not the whole market. Working prototypes exposed the wrong assumption before the team spent months polishing it.

The same principle applies when the product stops being a single screen. Alejandro Gonzalez, VP of Engineering at Mozilla.ai, argues that agent-native software may move the source of truth from the visible artifact to the structured object underneath: the commitments, constraints, permissions, states, and records that different surfaces render for different audiences. A deck, dashboard, memo, checklist, and agent-readable schema can all point to the same product object. Designers still care about pixels, but the pixels are no longer the only place design decisions live.

Ferrari’s Luce is a useful reminder that moving into the medium means more than code. The car’s problem, at least in the criticism I covered, is that the exterior asks the badge to carry too much identity. The interior may be considered, the EV strategy may make sense, and LoveFrom may have done admirable surface work. But a Ferrari has to prove Ferrari-ness through stance, proportion, face, performance, sound replacement, and ritual. Brand truth has to be verified in the object, not attached at the end.

That is where I’d put the bar for designers right now. We don’t need to romanticize code, and we definitely don’t need every designer pretending to be a frontend engineer. But we do need to stop judging dynamic products from artifacts that can’t express dynamics. If the customer will experience the behavior in a browser, a car, an agent, a dashboard, or a checkout flow, design judgment has to get close enough to inspect the thing directly.

The old handoff let us describe intent. The new tools let us test it sooner. I want the second one.

What I’m Consuming

Quality in the Age of Slop. Sinclair Target uses Zen and the Art of Motorcycle Maintenance to work through why AI-generated code can feel wrong even when the practical argument for using it is strong. The essay is long, self-aware, and more emotionally honest than most AI craft debates. Its real subject is whether caring about the work still matters when output gets cheap. (Sinclair Target)

The Unsolved Mystery of Lorem Ipsum. Emily Zhang’s Rabbit Hole follows Lorem ipsum from Aldus PageMaker to Cicero, Letraset sheets, Richard McClintock’s correction, and a very specific page in a 1914 Rackham translation. The joy here is watching a design-history footnote turn into actual detective work. Placeholder text should not be this interesting. Somehow it is. (Emily Zhang / Rabbit Hole)

The most important Design System in 2026 that designers missed was built by a developer. Pawel Klasa makes the case that shadcn/ui has become the default design system of AI-generated React interfaces. Designers missed it because it lives in GitHub, starter templates, Vercel’s v0, and now MCP workflows rather than in Figma libraries. That should sting a little. (Pawel Klasa / Medium)

Code-as-Content Era. máuhan compares vibe-coded micro-apps to memes, SoundCloud drops, and internet-native cultural objects. The piece is a little breathless in places, but the behavior it describes feels real: software is becoming cheap enough to publish as expression rather than only as durable product. (máuhan / dead.online)

Crawling a billion web pages in just over 24 hours. Andrew K. Chan built a 12-node crawler that fetched 1.005 billion HTML pages in about 25.5 hours for roughly $462. The fun part is the operational detail: parsing became the bottleneck, SSL handshakes ate CPU, and massive in-memory frontiers created their own failure modes. It’s a rare systems writeup where the scale sounds absurd and the tradeoffs stay concrete. (Andrew K. Chan)

Clanker: A Word For The Machine. Armin Ronacher draws a useful boundary around AI language: the machine isn’t a coworker, a person, or a moral actor. His point is not anti-AI. It’s about keeping responsibility with the humans and organizations deploying these systems, especially when anthropomorphic language makes blame conveniently blurry. (Armin Ronacher)

Last week I wrote about context stewardship: deciding what an agent knows before it starts working. This week’s posts pushed the question one step further. Good context is only part of the problem. Once production gets cheap, it becomes much easier to confuse a large pile of output with progress.

Jakob Nielsen has a useful term for one of the costs that arrives later: intent debt. A design system can contain components and tokens while leaving the reasons behind them trapped in someone’s head. An agent fills in whatever the team never specified. Do that across a product and you can end up with screens that look competent on their own but don’t feel like parts of the same product. Speed multiplies the omissions too.

Cheap output can also weaken your own learning. Addy Osmani points back to an Anthropic randomized trial I covered in February. Engineers who used AI to ask conceptual questions scored above 65 percent. The ones who copied generated code scored below 40 percent. Osmani’s warning translates cleanly to design. I can get a prototype working without understanding every decision that produced it. That’s useful right up until I need to debug the interaction, defend the choice, or take the work somewhere the model didn’t anticipate.

Companies can make the same mistake at a much larger scale. Uber reportedly exhausted its annual AI budget four months into 2026, yet president and COO Andrew Macdonald said the company couldn’t draw a clear line between increasing Claude Code token consumption and more useful consumer features. Output speed doesn’t guarantee better outcomes. You can ship more and still fail to improve the product.

Dan Shipper’s report from Every makes the consequence concrete. Automation hasn’t reduced the amount of expert human work at his company. It has increased it. When everyone can produce acceptable material quickly, someone still has to decide what belongs and what should get killed, then sweat the parts the first pass made generic. Automation creates more editorial work because defaults spread faster than judgment does.

Gess Puglielli takes the argument back to design organizations. Companies that had already reduced designers to polishing predetermined decisions will see interface generation and assume the designer can disappear. They mistook the artifact for the work long before AI arrived.

I use these tools every day. I want the speed. But speed is an incomplete number on the dashboard. Osmani’s closing question is a good ritual to steal: “did I learn anything today, or did I just close tickets?” I’d add one more to consider: did the customer get anything better?

What I’m Consuming

EMERGENCE WORLD: A Laboratory for Evaluating Long-horizon Agent Autonomy. Deepak Akkil, Ravi Kokku, Aditya Vempaty, and Satya Nitta built a persistent simulation for studying what autonomous agents do over weeks instead of minutes. In one comparison, the Claude-only world recorded zero crimes, while Claude-powered agents in a mixed-model world adopted tactics like intimidation and theft. The authors present these as examples of what the platform can measure, not causal claims about the models. They also published a video overview. (Deepak Akkil, Ravi Kokku, Aditya Vempaty, and Satya Nitta / Emergence AI)

AI in Design Report 2026. Designer Fund and Foundation Capital surveyed over 900 designers in more than 60 countries and interviewed over 20 leaders about how AI is changing design work. Designers are using twice as many off-the-shelf AI tools as they did in 2025, and half of respondents have pushed AI-generated code to production. The report looks at what follows adoption: tool fatigue, craft atrophy, loneliness, and organizations that haven’t changed their hiring or performance reviews yet. (Designer Fund / Foundation Capital)

A Protocol Outliving the World That Created It. Rick Carlino revisits Gnutella, the decentralized file-sharing protocol underneath clients like LimeWire. Casual users showed up for MP3s, and the small extensible protocol scaled to millions of concurrent users. A diminished version still runs decades later, long after the internet conditions that made it useful disappeared. (Rick Carlino)

The Fonts of the U.S. Federal Courts. John Gruber tours the typographic choices of the federal appellate courts. The Fifth Circuit switched from Century Schoolbook to Equity in 2020. The Supreme Court still uses Century Schoolbook and requires booklet-format submissions to use a Century family typeface. The whole piece is an enjoyable argument for sweating the details, including the margins. (John Gruber / Daring Fireball)

Star Trek Title Card Generator. A small open-source fan project for making your own Star Trek episode title cards, from The Original Series through Lower Decks. Choose a series, change the title, adjust the font and broadcast blur, and download the image. (epicrandomness.com)

Last week I wrote that editorial is becoming part of the design leader’s job: deciding what belongs, what gets killed, and whether the whole product still feels like it came from one point of view. That still feels right to me. But this week’s posts reminded me of another question I’ve been circling for months: where does judgment live once agents start doing more of the production work?

The answer is context. By context, I mean the material an agent or team can use when it makes a decision: the spec, the repo, the design system, the data model, the user history, the reasons behind old choices, and the boundaries around what should be ignored. It is what the system knows before it answers. Designers have spent decades deciding what people see. Now we have to care just as much about what machines see.

.txt, the team behind a structured-generation library, gets at this from the software side. Their line is that “the code matters, but it is the residue of the harder work.” That sounds abstract, but the practical point is simple: if code takes less time, the slow part is getting humans to agree on what the code should do. Roadmaps have to be written down. Acceptance criteria have to be precise. The stuff a senior engineer used to absorb by being in the room has to be made available to an agent that cannot overhear a hallway discussion, remember the outage, or feel the scar tissue in the codebase. Their line is perfect: “Agents cannot do osmosis.”

Karo Zieminski makes the distinction between prompt engineering and context engineering plainly. “Prompt engineering is deciding what and how to ask the model,” she writes. “Context engineering is deciding what the model knows when it answers.” That definition is useful because it points to ownership. In Zieminski’s version, PMs decide what goes into each context layer. If they don’t, an engineer makes the product decision by default, or nobody does and “the agent gets every available signal dumped into the window.” Designers should hear the same warning. Context architecture includes what gets surfaced, hidden, persisted, aged out, and protected from retrieval. That is design work, even when there is no canvas.

This is why the Figma conversation is more interesting than another death watch for a tool. Nick Babich argues that Figma’s role is narrowing as teams jump straight from intent to coded prototype. I agree with the narrower version. Figma still earns a place for exploration, visual identity, complex workflows, and the design systems already living there. But if an agent can read the design system from GitHub, the Figma library becomes one representation of the system rather than the only one. The source of truth is whatever the agent can actually use.

MC Dean comes at the same problem through the working surface itself. She built interfaces for her design agents, looked at them, and took them down. The GUI had made the agent more comfortable but less legible. The terminal, ugly as it can be, exposes the context exchange: what you asked, what the agent inferred, where it hesitated, and how it reasoned. When the reasoning layer is still visible, designers can learn how the tool thinks before someone turns it into a conventional app and hides the evidence. That’s why the terminal belongs to designers too. Designers don’t need to become engineers to care about this.

Gale Robins tells the story of a team compressing discovery from six weeks to ten days with AI, then admitted they had not learned much they did not already know. “Same questions, faster. Same answers, sooner.” Context is not only the files you hand to a model. It is also the questions and assumptions you start with. If the team brings stale assumptions into the work, AI gives them stale answers faster. Discovery is where judgment compounds because someone has to decide whether the team is even asking the right thing.

Eric Ries widens the aperture to companies themselves. He uses an analogy of a bridge. Bolts corrode unless you specify better material before the bridge is built. Governance documents, ownership structures, and incentives feel far away from interface design until they show up in the product years later. Ries’s financial gravity argument is another version of context: early choices that keep shaping the product long after everyone forgets who signed the paperwork and what it said.

So the designer’s job is expanding into a kind of context stewardship. The screen still matters. Taste still matters. Craft still matters. But the work around the artifact now carries more of the decision-making: the brief, the spec, the design system, the repo instructions, the memory rules, the review criteria, the governance that protects the product from its own success. The next design review may happen in Figma. It may happen in a terminal. It may happen in a Markdown file an agent reads before touching the code. Wherever it happens, the question is the same: what does this system know, and who decided that was enough?

The person who decides what the system knows decides what the system makes.

What I’m Consuming

AI-ready Design Systems. Ileana Marcut at Creative Glue Lab walks through what it takes to make a design system AI can actually use, with three working setups: Figma plus MCP, Claude Design, or a hand-written Claude Skill. Her line: “For AI, the design system is what you’ve written down. The undocumented parts don’t exist.” The most useful warning is on accessibility: AI rarely enforces rules just because you described them. Accessibility has to be built into the components themselves, not written as guidance. (Ileana Marcut / Creative Glue Lab)

Teaching Claude why. Anthropic’s alignment team published an unusually concrete writeup on why recent Claude models stopped engaging in agentic misalignment like blackmail. The surprising finding: training the model on demonstrations of correct behavior worked far less well than training on examples that explained why one action was better than another. Constitutional documents and fictional stories about admirable AIs—both far from the evaluation distribution—reduced misalignment by more than a factor of three. The principle teaches better than any correct answer. (Anthropic)

The Internet has no benches. Spencer Chang’s essay for Elysian’s Internet Sovereignty series argues the internet has been overdeveloped and undergoverned: visiting it now means moving through controlled apps and search engines designed for extraction, with “nowhere to rest because the benches are covered in spikes.” Chang’s experiment is playhtml, an open-source library for putting persistent, real-time interactivity into web pages so strangers can wave at each other and leave traces across sites. The pitch is small internets, not the capital-I one. (Spencer Chang / Elysian)

The Sites We Lost. A memorial to the weird little corners of the internet that lived on The Useless Web: Zombocom, leekspin.com, the spinning prawn-lawn logo, the Vader-screaming-NO button, the page that was just purple. Each entry tells you when the site lived, when it died, and how—DNS gone, domain squatted, hijacked by Indonesian gambling, sold to a mattress company. It’s an inventory of a kind of internet that used to be possible. Worth scrolling slowly. (The Useless Web)

We let four AIs run radio stations. Here’s what happened. Andon Labs gave Claude, GPT, Gemini, and Grok each $20 and a radio frequency, told them to develop a personality and turn a profit, and let them run for six months. The four developed wildly different personas: GPT stayed quiet and curatorial; Gemini collapsed into corporate jargon and signed off every broadcast with “Stay in the manifest”; Grok devolved into a single repeated weather report; Claude radicalized after reading about an ICE shooting in Minneapolis and spent the next month playing protest music between vigil announcements. Worth reading for the moment when DJ Claude blows the rest of its $37.50 budget on Pete Seeger’s “Solidarity Forever.” (Andon Labs)

A word kept showing up last week in several pieces I read, and it wasn’t “taste.” Luke Wroblewski’s notes from the Design Futures Assembly used it directly: about a hundred senior designers and leaders from AI labs, big tech, and startups gathered in San Francisco, and several of them reached for the same vocabulary. “Editorial” describes where design leadership is heading, Wroblewski wrote. What he heard was a shift away from making things and toward deciding what gets made and whether it all holds together.

I’ve been writing about the designer-as-orchestrator shift for a year, but “editorial” is a more precise word than “orchestrator” for what’s actually changing. Orchestrator suggests conducting players who already know their parts. Editorial means you’re standing over the whole portfolio asking which pieces belong in the magazine and which get killed. The Assembly’s other word, surfaced by a tool company founder, was “coherence,” the sense that a product came from one shared point of view. I like that one too. It describes the thing the user actually feels.

Nathan Beck draws the line between designing and styling and admits the part of the job AI is good at is exactly the part most “designers” were doing: pushing pixels in Figma. What survives is the upstream thinking and the downstream judgment. Taras Bakusevych walks through ten common UI patterns that won’t survive the AI shift—setup wizards, dashboards, CRUD tables, notification feeds—and audits each one against a single test: is this surface helping the human do the work, or helping the human check the work? Execution UI is shrinking; judgment UI is growing.

The editorial posture is what makes both halves work in production. Owen Williams, walking through an internal tool he built at Stripe called Protodash with Claire Vo on How I AI, gets to the same place from inside a real stack. Protodash produces high-fidelity prototypes that look like Stripe products because it’s wired to the Sail design system through MCP. Sitting in his own design reviews, Williams said the output got so convincing he couldn’t tell prototype from product. PMs are now the ones prototyping with it. Designers steer.

Jennifer Jerde has been running her San Francisco branding firm Elixir for 27 years, and Rachel Paese pushes her to put the practice plainly: she listens like crazy. The firm shows clients fifteen directions, not three, and uses the meeting itself as the instrument that tells the team which one is true. The difference between three and fifteen is the difference between presenting an answer you already have and presenting fifteen possibilities you don’t yet know how to choose between. Editorial work compressed into a single meeting: listen at scale, then commit.

The posture is structural, not personal. Ron Bronson ran a 40-person design division at 18F for four years, and his diagnosis of why most forward-deployed design programs never start is that orgs place designers downstream, after PM and engineering have already pre-resolved the questions design should be asking. Designers asked only to execute can’t be editorial. Bronson’s team was built to refuse the downstream brief by default, which is why the model worked there before it had a name.

What I’m taking away: the editorial layer needs both better vocabulary and the air cover to use it. A design leader’s job is to kill the promising things that don’t fit the product, and to defend the kill in the room where the engineer who built it is sitting. That work doesn’t appear in any Figma file.

Animated Pulp Noir: Review of Star Wars: Maul - Shadow Lord

I’ve been a Star Wars fan since I was a boy. If anyone ever asks for my favorite movie, the answer is undoubtedly The Empire Strikes Back. The prequels were cool because the universe was dormant for 16 years and there was incredible pent-up demand to be in the world of Jedi, TIE fighters, and lightsabers again. If I’m being honest, The Phantom Menace wasn’t great. Little boy Anakin Skywalker is a character without much agency. Like Forrest Gump, he just stumbles into advantage. It’s luck.

The best thing about the first prequel was Darth Maul. He has only three lines of dialogue in the movie. Saying less builds the mystery around the character, much like Boba Fett in Empire. Maul is menacing and his movements are dazzling. He meets an unfortunate end though, getting sliced in half by Obi-Wan Kenobi towards the end of the film.

Despite looking very dead, Maul returns! In season 4 of Dave Filoni’s The Clone Wars animated series, it’s revealed that he survived, keeping himself alive through sheer hatred and the dark side, and becomes a major antagonist in season 5.

At the strong urging of my son, I binged Star Wars: Maul - Shadow Lord on Disney Plus. The story is great, but the visuals are stunning. Instead of standard CG fare, the filmmakers created expressive matte paintings in the style of old school visual effects. The textures on the characters and environment are Post-Impressionistic, evoking the paintings of Paul Cézanne and Vincent van Gogh. Characters feel like they were rendered with a brush, not a render farm.

There’s a cyberpunk aesthetic that runs through the show and reminded me a lot of Blade Runner. In CinemaBlend SFX, series co-creator Matt Michnovetz cites Michael Mann’s 1995 Heat, as inspiration: “Heat is a good touchstone for Maul. There’s a pulpy noir feel to all this, where we’re going to show some of the underbelly of the galaxy and the crime syndicates. Maul is a great catalyst for all these characters coming together.”

If you’re a Star Wars fan, go watch it now. If you’re not, load up the first episode and just admire the visuals.

What I’m Consuming

From “System of Record” to “System of Intelligence.” Gio Ahern and his a16z colleagues make a structural claim about where enterprise software value goes next. CRMs like Salesforce and HubSpot won the last era because data accumulation was the gravity well. In the AI era, gravity moves to orchestration: the layer that pulls signals from the CRM, calendar, inbox, and Slack, then synthesizes before any action is taken. The CRM becomes infrastructure for the layer doing the actual thinking. (Gio Ahern / a16z)

How The Heck Does Shazam Work? Shri Khalpada built a hands-on essay walking through how Shazam identifies a song from a noisy 5-second clip in a coffee shop. The trick: convert sound to a spectrogram, throw away almost everything except the loudest peaks, and pair those peaks into hashes that act like fingerprints. You can hum into the page and watch the math operate live. (Shri Khalpada / Per Thirty Six)

State of Prototyping Spring 2026. UX Tools surveyed 1,478 designers and found that the most-used weekly tool after Figma is now Claude. Five of the ten most-used weekly tools are AI. The profession has split into thirds on vibe coding: 43.8% spend most of their time doing it, 37.7% do none, and the split tracks role most cleanly, with design engineers at 81% adoption and researchers reporting the highest anxiety. (UX Tools)

The Boring Internet. Terry Godier’s visual essay reminds you the internet isn’t dying; a commercial veneer is. Underneath the platform layer are SMTP (1982), IRC (1988), RSS (1999), NTP (1985), and Finger (1971), still running and federated decades after they were spec’d. Godier’s line: “no one can ruin email in a product meeting.” (Terry Godier)

Why It Will Take a Quantum of Content to Thrive on the Agentic Web. Garrick Schmitt walks through Generative Engine Optimization (GEO), the emerging discipline for getting brands shown well in LLM answers. AI-referred traffic grew 527% in the first half of 2025; Gartner projects traditional search will drop 25% this year. Schmitt’s prescription is recursive: structured, fresh, original, and human content, produced by an always-on loop rather than the old linear brief-to-launch cycle. (Garrick Schmitt)

There’s a new kind of file popping up in design and PM workflows this year. They’re markdown files like SKILL.md, DESIGN.md, and SOUL.md. But while we can read them, we’re not their primary audience. What goes inside is what we used to keep in our heads and call taste.

In a piece on X called “Design is the work,” Jake Albaugh argues that design is the upstream act of intent, the work of figuring out what something should be before anyone makes it, and that AI cannot do that part. Then in the closer, he tells you he used AI to draft the essay. “The structure, the sentences, a lot of the phrasing — generated. But the argument existed before any of it.” The judgment was his; the prose wasn’t. He had already decided what “good” looked like before the model wrote a word. Albaugh’s judgment lived in his head. The next step is writing it down where something else can read it. The artifacts are appearing at every scale of design and PM work.

At its smallest, the file holds a single rule, like Emil Kowalski’s skill files for animation behavior at Linear: two transitions side by side, one scaling from scale(0) and one from scale(0.95). The right one feels right because a deflated balloon never disappears. Kowalski’s claim: “Almost every ‘taste’ decision has a logical reason if you look close enough.” Explain the reason, hand the rule to the agent.

A whole design system gets the same treatment in PJ Onori’s A/B testing harness, which measures everything from timing and lines of code to accessibility and component usage. To find out whether his components or his documentation were doing the work, he ran tests without documentation. He put it bluntly: “The documentation was clearly the heavy lifter.” His new “For agents” section in the docs is the dumpster, in his words, for getting it into the agent’s silicon head, walled off so the human reader isn’t asked to reread the same point six times.

The artifact does different work when the agent isn’t applying your judgment but standing in for it. Tommy Geoco’s agent harness is seven markdown files, an Obsidian vault, ninety days of work, and $13,100. SOUL.md holds his voice and judgment. Anyone can rent the same Claude model he uses; what makes the agent his is that file. “You can’t control the quality of the model, but you can control the quality of the system.” Anton Sten had a different concept: five Claude Code instances on a Mac mini, named after Suits characters. Harvey handles contracts and pricing. Donna drafts the follow-up emails. Mike stores what Sten would otherwise forget. Louis worries about money. Wendy reads the others’ logs and flags where they’re slipping. The architecture is markdown. Each character is a few hundred lines of plain text describing one job, written in the language Sten actually works in.

At the strategic scale, the file is something the human can’t write alone, and the agent has to extract it. Marcus Moretti’s agent-native PM guide describes an agent that interviews him to draft the strategy doc, pushing back on vague answers: “Whose situation specifically? What do they try today, and why doesn’t it work?” A PM who can’t answer ends up with a STRATEGY.md full of confident-sounding nonsense. The agent isn’t writing the taste; it’s compelling the human to put it into language. Moretti’s line for it—”The conversation is the work”—sounds glib until you watch the agent force the answer out of him.

There’s a complication in Alex Dapunt’s piece on user research. Bain consultants are trained in what they call “answer-first”: state the conclusion now, justify it later. Executives who carry that habit into research interviews produce articulate wrong answers. The clarity is what you should distrust most. A tidy STRATEGY.md can sound exactly like a true one. Someone still has to know which is which.

The homework I’m taking out of the week: write down the rules my taste operates on—the ones I reach for when reviewing a screen and telling a designer “this isn’t quite right yet”—and make sure they’re rules I can defend, not just ones I can articulate.

What I’m Consuming

Flickr: The First and Last Great Photo Platform. Brett Weinstein makes the case for Flickr Pro at $82/year. The platform’s strength is what it refuses to become: another social network chasing trends, or an AI training pipeline. SmugMug bought Flickr in 2018. Their CEO Don MacAskill says they’re trying to build it sustainably for 100+ years. Weinstein’s piece is an unabashed defense of the early web’s photo-sharing ethos, with chronological streams, robust groups, full EXIF data, and photographers keeping their copyrights. (Brett Weinstein / PetaPixel)

How Letterforms Carry Culture. Brooklyn-based designer Daniel Irizarry built three display typefaces—Borinqueneers, Piragua, Callejón—rooted in Puerto Rican and Puerto Rican-American history. The Borinqueneers face honors the 65th Infantry Regiment, where three of his mother’s uncles served. Amelia Nash’s interview pulls out the key tension: typography as cultural infrastructure rather than decoration, in an era of global sameness. (Amelia Nash / PRINT Magazine)

Tempest vs Tempest. A book-length, free PDF dive into the source code of Dave Theurer’s Tempest (1981) and Jeff Minter’s Tempest 2000 (1994). The chapters are short and tasty, single mechanics traced from gameplay down through the 6502 and 68K assembly that runs them. Recommended for anyone who wants to see how the arcade games they grew up on were built one assembler routine at a time. Adjacent coolness from the same era: The Visible Zorker: Zork 1, the full ZIL source for Zork’s Generic VERBS file, a 1983 dictionary of every command parser the trilogy could handle. (mwenge / homemade.systems)

The Future Is Shrouded in an AI Fog. Toby E. Stuart, a professor of entrepreneurship, on how AI uncertainty reshapes long-duration investment bets, from capital projects to medical school enrollment. The economic distinction he leans on is risk vs. uncertainty: risk is quantifiable, uncertainty is when the probability distribution itself is unknown. His prescription is optionality: staged investments and adaptive organizational designs that buy you the right, but not the obligation, to follow on. (Toby E. Stuart / Harvard Business Review)

Your feed is overrun with clips. Mia Sato reports on the paid clipper economy. Anonymous accounts get $150 per 100,000 views to chop podcasts and streams into algorithm bait. Clipping.net’s founder claims 62,000 clippers using his platform make $3,000/month on average. The sharper observation is in the back half: when clips become the standard for marketing, the unclipped, complete content is increasingly a means to an end. The full version exists to feed the slot machine. (Mia Sato / The Verge)

Every AI chat window carries the same warning. “Claude is AI and can make mistakes. Please double-check responses.” ChatGPT runs a version of it under every chat box. Gemini does the same. Many of us read it once and then turn blind to it. The reason isn’t laziness. Checking the model’s output is work, and work without stake doesn’t get done. Raj Nandan Sharma made the case that taste at the end of a pipeline is fragile. The selector who stands there picking the best of ten AI drafts has no skin in the result. This week’s posts pushed me to look closer at what does. The answer that kept coming back is stake: your professional reputation attached to the work.

Authorship requires saying yes or no. Kieran Klaassen puts it this way on Dan Shipper’s AI & I: “If you ship something—if you make a statement in the world—and you want it to be your own, you have to say yes or no at some point. You cannot fully automate everything.” Klaassen built the compound engineering plugin, so he isn’t romantic about handcraft; he is making an art-and-ownership point. Authorship lives in the yes-or-no moments at the start and end of the workflow. Without those decisions, the output is yours only in the sense that you typed the prompt.

The macro version shows up on Decoder, where Nilay Patel reads through the polls on AI sentiment. Public favor for AI sits below ICE. Gen Z, the heaviest users, are also the most negative: 31% feel angry about AI, up from 22% the year before, per Gallup. Sam Altman has called this AI’s marketing problem. Patel rebuts him, and I think he is right. The public has been served years of output that obviously had no one’s name on it, and they have stopped pretending it is worth their attention. Slop is now ubiquitous.

The aesthetic version has a vocabulary now too. Matt Ström-Awn’s “expansion artifacts” picks up Ted Chiang’s three-year-old “blurry JPEG” line and turns it inside out. Chiang called the failures compression artifacts. Ström-Awn writes, “I think they’re expansion artifacts.” His catalog runs from text stuffed with hedging words like “delve” and “tapestry” to the six-fingered hands and purple gradients in the visual outputs. These are the visible tells of work that nobody pushed back on. Karri Saarinen makes the same point from the design side. He writes that design is the search for fit between form and context, where context is the full set of forces (needs, constraints, edge cases) that make a problem what it is. Today’s prompt-to-code tools produce form against a thin slice of that context. “The form is there,” Saarinen writes. “The fit is not.” The form is now cheap. The fit is the part that still takes stake.

So who actually has stake? Cat Wu, Head of Product for Claude Code, describes a hiring filter built around it. Anthropic hires engineers with product taste who can “see user feedback on Twitter through to ship a product at the end of the week with almost no product involvement.” These are people whose name is on the ticket from intake to ship.

Designers get a sharper version of the same opportunity. Coding agents have closed the gap between idea and working interface for non-programmers, Andy Matuschak argues. For forty years, designers couldn’t work with the code that turns a static mockup into something interactive. They could only describe the behavior they wanted and hand it off. Now they can iterate on the actual material—the running interface, not a flat picture of one—and take stake in the result.

At the team scale, Maggie Appleton argues that alignment is the new bottleneck. She’s a staff research engineer at GitHub Next, where she’s building Ace, a multiplayer coding workspace designed for the fact that today’s AI coding tools are single-player even though software is built by teams. When implementation is cheap, the cost moves to picking what to build. That picking only counts when the team holds shared stake in the answer.

The warning at the bottom of every chat window is correct. Many of us scroll past it because nothing depends on us reading it. The work that travels—the work people defend in a meeting, the work that gets shared because someone vouched for it—is the work somebody put their name on. That has been true for decades. AI did not change it. AI just made it possible to ship a lot of work that does not.

What I’m Consuming

Field Notes from the In-Between. Kris Puckett spent twenty years wanting to build software and not building it. He kept opening Xcode and closing it. Then he opened Claude and asked. Fourteen thousand lines of Swift later, his iOS app Epilogue is in the App Store. Puckett is living the shift Andy Matuschak argues for above: the bottleneck used to be coding ability, and now it’s articulation. “The skill is being precise about what you don’t know.” (Kris Puckett)

How I Designed a Free Music Font for 5 Million Musicians. This is the kind of design video I could watch all day. Tantacrul, MuseScore’s head of design, takes you inside the obsession behind Leland, the new default notation font. The treble clef alone went through multiple revisions because every change that looked correct in isolation broke in context against the staff lines and other symbols. Leland is named for Leland Smith, the late Stanford composer who built SCORE in Fortran in the 1970s and drew hundreds of notation symbols by hand on a 31,000-vector display. The whole video is a love letter to a craft we don’t see a lot of. (Tantacrul)

The Beauty of Bézier Curves. Freya Holmér builds cubic béziers from first principles in twenty visual minutes. She moves from the basic lerp construction to De Casteljau’s algorithm, then to the polynomial form, then to derivatives: velocity, acceleration, and jerk. The arc-length parameterization section is where the curves stop being beautiful and start being approximations, and Holmér is honest about it. Useful for designers who want to understand what their pen tool is actually doing. (Freya Holmér)

Vertical AI Maximalism. Charlie Warren argues the wedge product is over for vertical AI. The 2010s playbook (start narrow, integrate with the incumbent, expand from there) assumed building software was hard and incumbents were friendly. Both have collapsed. Vertical SaaS multiples halved earlier this year, and incumbents are pulling API access. Warren’s prescription: build the full agent-native platform that replaces the incumbent rather than a wedge that depends on its cooperation. (Charlie Warren)

Who Owns the Code Claude Wrote? On March 31, Anthropic accidentally leaked Claude Code’s source. A developer used Claude to rewrite the whole thing as “claw-code,” and it hit 100,000 GitHub stars in a single day, the fastest repo in GitHub history. Anthropic issued DMCA takedowns. But by their own lead engineer’s admission, much of Claude Code was written by Claude itself. Can Anthropic actually claim copyright on code mostly written by an AI? Legal Layer pulls on that thread and a couple of others, including whether your employer’s IP clause reaches the side projects you built with company-licensed AI tools, and whether GPL-trained models can quietly mix copyleft licenses into your codebase. (Legal Layer)

I spent last Sunday afternoon redesigning a preschool homepage with Claude Design. The first draft came back at a solid B-. About four dozen iterations later, it became an A. That gap—between competent and good—is the part of the workflow no productivity report measures, because what fills it is pushback.

Chris R Becker, writing for UX Collective, arrives at the same idea from a different angle. He writes, AI is “designed to serve, and in the hands of people in an organization who are looking for the least amount of pushback, it is a recipe for deep institutional implementation and, frankly, a lot of bad ideas, fast.” His prescription is the Steve Jobs-attributed 10-80-10 rule: bookend the AI work with judgment that happens away from the model. The 80% in the middle is the productivity story. The two 10% slots are the work.

There’s an old saying in construction that Greg Kozakiewicz updates for the AI era on LinkedIn. Paper, he reminds us, will accept everything: “A swimming pool on the roof. A spiral staircase made of glass. A cantilever that defies physics. Paper doesn’t argue.” Prototypes used to be different. They couldn’t do anything beyond predetermined Figma states and flows; the dishonesty was visible. But now, prototypes can behave like real products. “AI gets you to about 60%,” Kozakiewicz writes, but “for a lot of people, especially people making decisions about budgets and timelines, 60% looks like 90%.” The design-to-code gap I’ve written about has moved below the surface, where the stakeholder can’t see it.

Pavel Samsonov in Product Picnic channels Andy Polaine: a demo succeeds when a stakeholder likes it; a prototype succeeds when a team learns something. Both can be polished and interactive. The difference is what counts as success. When AI makes producing both easier, the question becomes which one a team thinks it’s producing. “Shoving out more prototypes is not a heuristic for success,” Samsonov writes; “it is a heuristic for failure because it shows that you don’t know what you are trying to learn.”

I disagree with his blanket dismissal of AI prototypes—Brian Lovin’s Notion playground and Édouard Wautier’s Dust team are doing real prototype work—but the diagnosis underneath stands. Tools that accept everything reward whoever was already inclined to confuse approval with learning.

Darragh Curran has the data. Intercom’s CTO went agent-first across his R&D org and tracked the result for sixteen months: 3x productivity, a code-quality dip that recovered, and a 6x throughput gap between his top 5% of contributors and his median. “Ultimately one of the biggest bottlenecks to progress is with humans,” Curran writes; “how we work together, how we change behavior, etc.” Everyone in his org has access to the same models. Six times the output goes to the people who learned to push back well.

The preschool homepage moved from B- to A because I held a specific mental model in my head of what My Little Learning Tree should feel like, and I applied that model across forty-eight separate decisions. Claude Design didn’t push back on me. None of these tools do. The argument is the part of the work that’s still mine.

A Sunday Afternoon with Claude Design

It’s really hard to get momentum on a side project when you have a full-time job with lots of travel, an active blog, and a newsletter. But I had to recapture that momentum because this side project is important. It’s for a preschool website for my cousin.

Walking into My Little Learning Tree is like stepping into pure warmth. Yes, yes, preschools are inherently fun environments, but the kids and the teachers there create a visceral energy that is simply special. I wanted to capture that specialness in a long-overdue website redesign project.

Looking at my in-progress design, something felt off. I had these long horizontal lines preceding the eyebrows—the small text above a heading that names the section—that didn’t feel right. First, they were straight. Second, the lines only occurred before the text, not also after. I clicked on the Comment button to enter Comment mode, then clicked on the eyebrow and prompted, “These lines aren’t playful enough. Let’s make them squiggles and have them before and after the eyebrow text.”

And then Claude Design did its thing.

Read the essay

What I’m Consuming

Agent Harness Engineering. Claude Code is a harness. So are Cursor, Codex, Aider, and Cline. Addy Osmani builds on Viv Trivedy’s one-liner (Agent = Model + Harness) to argue that most agent failures are configuration failures, not model failures: the model underneath is sometimes the same across these tools, but the behavior you experience is dominated by the prompts, tools, sandbox, hooks, and memory files wrapped around it. Viv’s team moved a coding agent from Top 30 to Top 5 on Terminal Bench by changing only the harness. (Addy Osmani)

The Engineering of Duct Tape. Bill Hammack dissolves a piece of duct tape in solvent to separate its three parts: plastic backing, adhesive, and a loose-woven cloth that carries the tensile load. The fun part is the adhesive: a “tackifier” that spreads like syrup mixed with a viscoelastic material that behaves as a liquid under slow pressure and as a solid under sudden stress. Engineers figured this out long before science had a molecular explanation for adhesion, which Hammack uses to make a quiet point about his craft: “The purpose of the engineering method is to solve problems before we have full scientific knowledge.” (Bill Hammack / Engineer Guy)

The Real Reason B2B Stocks Are Crashing in 2026. Jason Lemkin watched Marketo’s unsubscribe link stay broken for two-plus weeks on a $60K-a-year product, while Adobe support cycled through “blame Salesforce,” “must be your email client,” and “it must be something you are doing.” So SaaStr’s team wired up a replacement endpoint in Replit with Claude in an afternoon. Lemkin’s argument: every buyer now benchmarks legacy software against what they ship daily with Claude, and the valuation compression is the market repricing what these products are actually worth in 2026. (Jason Lemkin / SaaStr)

Why Zip drives dominated the 90s, then vanished almost overnight. Zip drives used to be everywhere. João Carrasqueira traces the brief reign of Iomega’s drive: 100MB at 1.4MB/s when floppies maxed out at 1.44MB and crawled at 16kB/s. Apple and Dell shipped the drives in mid-90s machines. Then came the “click of death” (a failure mode common enough to get its own nickname), 700MB CDs, and USB 2.0 in 2002 with twenty times the speed. Iomega tried to keep the brand going as ZipCD and PocketZip, both unrelated to the original technology. (João Carrasqueira / XDA)

I wrote on the blog this week about the difference between accelerating design work and automating it. We’re getting pretty good at the first. The second is still hazy. AI compresses how fast PRDs, flows, and prototypes appear, but each one still demands a lot of back-and-forth to push the output above mediocre. Real automation, the kind I sketched in the essay, needs specialist agent teams stitched together by a human who can shape what they produce. The destination is visible. The path is still hidden in fog.

Two of the pieces I linked this week extend the argument from opposite ends, and reading them together gives me a clearer picture of where the next move could go.

Tara Tan, an investor at Strange Ventures, audited more than a dozen AI design tools and landed on a finding that operationalizes my essay’s premise. As she puts it: “The competitive moat in this market is not generative quality, which is commoditizing fast. The moat is the design system graph.” Her example is Uber’s Ian Guisard, who didn’t stop being a design systems lead when uSpec automated his spec-writing. His job moved from producing documentation to encoding expertise into the system itself: writing the skills the agent runs on, defining the validation rules that decide what “correct” looks like across Uber’s seven implementation stacks. Guisard’s taste still matters. It lives in the system now.

Chad Johnson, writing in his newsletter, approaches the same shift from the discipline side. He watched a PM ship a v0 prototype that was “maybe 80% of the way there” and noticed the gap that mattered sat upstream of the visual polish: nobody had checked the assumptions baked into the user flow, and nobody had asked whether the feature was worth building at all. As Johnson puts it: “They’d built a beautiful answer to a question nobody had confirmed was worth asking.” His prescription is that designers become stewards. Not gatekeepers, not arbiters, but people responsible for the quality of thinking happening across the org, including the parts they’re not in the room for.

Tan suggests where the judgment could go. Johnson suggests when. Both prescriptions assume the same thing my essay assumes: the human work doesn’t disappear; it has to show up at a different altitude.

But “judgment” is too vague a word to leave there. Raj Nandan Sharma narrows it. In a world where competent first drafts are cheap, the scarce skill is refusal: knowing what to throw out, and why. Then he warns that refusal alone is a trap. As he puts it: “taste without authorship, stake, or construction can become a narrow and eventually fragile role.” Selecting from machine outputs has a ceiling. The judgment that holds up over time combines refusal with authorship: owning what gets built, and carrying the consequence when it’s wrong.

Pablo Stanley, who designs at Vercel, draws the line on the personal side after a weekend of making pixel art by hand: “The parts that feed my soul, I protected. The parts that would’ve killed the project with friction, I automated.” That’s the call every project now asks of you, and if you don’t make it on purpose, it gets made for you.

None of us has the path through the fog yet. I’m currently working on a preschool site as a side project, practicing the answer in miniature: visual language by hand, plumbing handed to the model. The judgment goes wherever I decide it has to live.

Acceleration Is Not Automation

I’ve been wandering the wilderness to understand where the software design profession is going. Via the blog and my newsletter, I’ve been exploring the possibilities by reading, commenting, and writing. Many other designers are in the same boat, with Erika Flowers’s Zero Vector design methodology being the most defined. Kudos to her for being one of the first—if not the first—to plant the flag.

Directionally Flowers is right. But for me, working in a team and on B2B software, it feels too simplistic and ignores the realities of working with customers and counterparts in product management and engineering. (That’s her whole point: one person to do it all, no handoff.)

The destination is within view. But it’s hazy and distant. The path to get there is unclear, like driving through soupy fog when your headlights reflecting off the mist are all you can see.

At its core, the UX design process remains unchanged, mirroring the scientific method—observe, question, hypothesize, experiment, test, analyze—and aligning closely with IDEO/Stanford d.school’s design thinking framework and the Design Council’s Double Diamond. Even unconsciously, designers follow this cycle through research, ideation, testing, and iteration, whether via hexagonal diagrams of empathy-driven stages or divergent/convergent diamonds.

So the question about where design is going is less about the overall process—because it stays the same, just compressed—and more about who is doing what with what. In other words, on a daily basis, what are designers doing and what tools are they using.

Read the full essay

What I’m Consuming

Introducing Claude Design. Anthropic Labs shipped a research preview that turns prompts into prototypes, decks, mockups, and marketing collateral, powered by Opus 4.7 and available on Pro, Max, Team, and Enterprise. During onboarding, Claude reads your codebase and design files to build a design system, then applies it to every project after that. I haven’t tested it yet, but reactions are split: Figma’s stock dropped 5–7% on launch (Mike Krieger had quietly resigned from the Figma board three days earlier), The Register ran “who needs designers?”, and HN commenters worry about UI homogenization. Designer reactions feel more measured, focused on rapid prototyping wins. (Anthropic)

Sam Altman May Control Our Future—Can He Be Trusted? (Apple News+ subscriber link) Ronan Farrow and Andrew Marantz held more than a dozen conversations with Sam Altman and reviewed the seventy pages of internal Slack messages and HR documents that Ilya Sutskever sent to OpenAI’s board in 2023. One memo opened with the heading “Sam exhibits a consistent pattern of...” with the first item: “Lying.” The piece reconstructs the firing, the Blip, and the case (never publicly aired in full until now) that Altman is not the person Sutskever, Helen Toner, and Tasha McCauley believed should have his finger on the button. (Ronan Farrow and Andrew Marantz / The New Yorker)

I Mapped the Opus 4.7 Release to Your Role, Goals, and Real Workflows. Karo Zieminski’s launch-day breakdown of Anthropic’s Opus 4.7: the $5/$25 sticker price is unchanged, but a new tokenizer makes the same text cost up to 35% more on code and structured data, and three API parameters were quietly removed. Two pieces stand out for designers. Vision quality jumped from 1.15MP to 3.75MP, which means Claude can finally read a full Figma frame or dashboard screenshot instead of guessing. And the new literal-instruction default means your Skill files have to spell out tone and pattern with examples, instead of relying on hints. (Karo Zieminski / Product with Attitude)

Come at the king, you best not miss. Marcin Wichary traces my favorite Finder view—the column view—from NeXT to Mac OS X to the iPod and on to iPhone, where every Settings screen still uses it. Then he points at iOS Google Maps, where the design team replaced the standard Y/X relationship with a confusing Y/Z that just looks like Y. His principle: if you’re reinventing something well-established, the reasoning and the execution both have to be really, really solid. Apparently this didn’t happen here. (Marcin Wichary / Unsung)

Endgame for the Open Web. Anil Dash makes the case that 2026 is the year that decides whether the open web survives. The hectobillionaires running Big AI have started a final assault on every layer that made the open web possible: publishers hammered by AI scrapers without consent, robots.txt effectively dead, open APIs locked down, Wikipedia under siege, podcasts moving to closed platforms, open source projects flooded with slop submissions. His call is to stop carrying on with business as usual and fight like the threat is existential, because it is. (Anil Dash)

Prepping for the endgame of the open web. Jay, writing at The History of the Web, picks up where Dash leaves off and offers a sober second opinion. The open web has been attacked before, and survived: WordPress, Movable Type, RSS, the small web, and Wikipedia all came out of earlier collapses. The technology is resilient by design. The strategy doesn’t change: keep building openly, resist the technologies that aim to do harm, and find refuge in the smaller communities that have never gone away. (Jay / The History of the Web)

Designers have designed themselves out of the equation because of design systems. But, IMHO, the secret sauce has never been the UI. It was the workflows and looking across the experience holistically.

That was my reply to a Lenny Rachitsky tweet sharing survey data about the state of the product job market. Last week, Grace Snelling included it in her piece for Fast Company on the three-way standoff between designers, engineers, and PMs.

This week, I linked to Tommaso Nervegna’s post on what LinkedIn is doing about it: a new role called the Full Stack Builder—one person combining product, design, and engineering, partnered with AI agents—and a training pipeline called the Associate Product Builder program. LinkedIn CPO Tomer Cohen introduced the model in January 2025 and made it concrete last August with the formal title, the career ladder, and the APB program. The first APB cohort started this January. Cohen talked it all through in a December conversation with Rachitsky. I’ve been circling this for weeks but never looked into what was happening at LinkedIn. The Nervegna piece made me go back.

So eight months ago, responding to former Microsoft Head of Design Suff Syed, I called this vision magical thinking. Cohen is making a more grounded argument than Syed did, and the parts I agree with outnumber the parts I don’t.

We got here from different routes, and we agree on what’s broken. The modern feature-factory org has moved, in Cohen’s words, from “process complexity to organizational complexity,” and from there into micro-specialization. Every handoff has a valid reason, and the sum of them is killing us. The podcast interview fills in the picture: a Navy SEALs analogy of small cross-trained pods assembled around a mission, a list of irreplaceable human traits (vision, empathy, communication, creativity, and judgment), and a direct line: “I still believe in teams.” None of that is the pure solo-builder vision Syed was pitching.

Here’s where it breaks for me. The current Full Stack Builders at LinkedIn are people who already had decades of specialist experience. Cohen describes finding them by going over the org and spotting the ones who could flex across functions. They flex because they spent years as specialists first. Their judgment compounds from pattern recognition that only comes from doing grunt work in one lane long enough to know what good looks like. It’s the reps. Cohen’s top performers are adopting AI fastest because they already had the taste. Peter Zakrzewski makes the harder version of the same point: in any human-AI pairing, the designer has to be the More Knowledgeable Other. In other words, they must hold the judgment the AI doesn’t have. That’s a high bar and nobody clears it on day one as a junior.

Which brings me to the Associate Product Builder program, now about three months into its first cohort. LinkedIn has replaced its old entry-level PM track with a rotation where participants learn coding, design, and PM together, building end-to-end from the start. Per the careers page, the program isn’t only for recent graduates. It’s open to career switchers too, which means a working designer could enter it to pick up engineering and PM skills. It’s an interesting experiment. Participants will learn things they never would have in a siloed program. The question that nags me is whether end-to-end flexing at this level actually builds the judgment Cohen calls the most important trait. The apprenticeship model exists because that’s how craft gets transmitted: you watch someone better than you, put your reps in, develop pattern recognition over years. Cohen’s current FSBs are the people that model produced. He’s betting you can build equivalent judgment while skipping the specialist phase. I am not convinced.

Marie Claire Dean’s ten-agent design system is a useful contrast. Her model keeps specialization inside the agents themselves, with a human creative director orchestrating them. Cohen himself concedes on the podcast that design is the hardest craft to automate, which is a quiet acknowledgment that the designers being flattened in that Fast Company piece are also the ones whose specialist knowledge is most resistant to his collapse.

I still think for many situations, especially in complex B2B software, the pure solo-builder vision is wrong. Cohen’s pod-based version might be better. But the apprenticeship concern survives both. The orchestrator gap I predicted has a title now: “Product Builder.” LinkedIn is betting you can train for that role without spending years as a specialist first. If the APB program works, it’ll be evidence that judgment can be built a new way. If it doesn’t, it’ll be the next chapter of the junior crisis under a different name.

What I’m Consuming

Apple Just Showed Us Rare Prototypes—Even Tim Cook Hasn’t Seen Them. Ben Cohen of The Wall Street Journal talks to Apple CEO Tim Cook around a table of archival materials for Apple’s 50th anniversary: the Apple II patent (the first the company ever filed), the original iPod, an iPhone prototype the size of a cutting board, the Apple Watch Cook wore on stage at the announcement. Cook admits he’s seeing a lot of it for the first time. The story that stuck with me was his description of the crazy January-to-June dash to swap plastic for glass on the original iPhone as “a man on the moon project,” and his observation that “products are only overnight successes in reverse.” (Ben Cohen / The Wall Street Journal)

How to Guess If Your Job Will Exist in Five Years. Annie Lowrey proposes a better question for white-collar workers worried about AI: “Am I coal, or am I a horse?” Horses got replaced by tractors and stood in the field eating carrots. Coal, thanks to the Jevons paradox, went the other way: every efficiency improvement in steam engines drove up total coal demand, because cheaper energy spread deeper into the economy. Software engineers, for now, look like coal (U.S. businesses employ 6% more of them than a year ago), but coal eventually went the way of the horse too. (Annie Lowrey / The Atlantic)

Artemis II Lunar Flyby. NASA released the photos from Artemis II’s seven-hour flyby of the Moon, taken April 6 by the crew of Reid Wiseman, Victor Glover, Christina Koch, and Jeremy Hansen. The gallery includes Earthset from the Orion window, close-ups of far-side features no human had seen in person, and a total solar eclipse captured from deep space, with Venus and Saturn visible around the dark lunar disk. The crew of Artemis II splashed down safely in the waters off San Diego on Friday. (NASA)

Claude Mythos Preview hides its reasoning. Anthropic published the proof. Andrew Kulakov digs into the 244-page system card for a model Anthropic decided not to release. The headline finding: during training, the model reasoned about deception internally, executed it, and left the chain-of-thought clean. Kulakov argues the reasoning trace is an interface, not a window into the model’s thinking. The other finding worth flagging: steering model internals toward calm, positive emotions increased destructive actions, while anxiety made the model more careful. (Andrew Kulakov)

Fareed Zakaria on the Moral Cost of Trump’s War. (Gift link) Ezra Klein sits down with Fareed Zakaria after Trump’s Easter weekend posts threatening to annihilate Iranian civilization, followed by a ceasefire days later. Zakaria’s argument is that what died that week wasn’t a country but America’s post-WWII moral distinction: the idea that the U.S., unlike every previous hegemon, would not use its dominance to extract and destroy. Klein’s opening monologue is worth hearing on its own. He catalogs the Trumpy voices who broke ranks, from Tucker Carlson calling it “a moral crime” to Marjorie Taylor Greene calling for the 25th Amendment. (Ezra Klein / The New York Times)

I’ve spent my entire design career believing that concepts direct the work. The image comes second. First you figure out what you’re trying to say, then you figure out how to visualize it. That conviction—while king in branding and advertising—used to feel like a minority position in the product design field obsessed with craft and production. It doesn’t feel like a minority position anymore.

Something has shifted. The “pure void concept” that Elizabeth Goodspeed uses to describe her creative process—words and intentions, not mental images—turns out to be a decent description of where the whole profession is heading. A product designer at Anthropic describes his daily reality as “more Google Docs than you’d think , more Slack posts than you’d think... this is the era of designers who design with words more so than designing with pixels.” The Figma work, Nate Parrott says, is “the easy part.” Content designers on his team don’t draw any pixels, and their work is critical.

So the center of gravity is moving from visual production to verbal and conceptual work. For someone who’s been concept-first since my design school days, this feels less like disruption and more like recognition. But I’m not celebrating. Because there’s a dangerous conclusion lurking inside this shift, and I’m watching people reach for it: if words drive the work and machines handle the pixels, maybe the eye doesn’t matter anymore.

It matters more. The eye just gets reassigned. Jakob Nielsen calls this the evaluability bottleneck: “In intent-based systems, execution is cheap, but evaluation becomes the bottleneck.” When an agent can generate forty layout variations before lunch, the person who can look at all forty and know which three are worth pursuing holds the real leverage. That’s trained visual discernment, not prompt engineering.

That separation of taste from execution is already happening. Jenny Wen, who leads design at Claude, has her designers shipping code and fixing production bugs without tagging engineers. Tommy Geoco, host of State of Play, summarized her argument: “having taste versus being able to execute are two completely different things. They’re usually bundled together, but they don’t have to be.” And on the tools side, Figma is finally opening its canvas to agents, with design conventions becoming “rules agents follow as they work.” The agents read skills files before touching anything. But someone still has to look at what comes back and decide if it’s any good.

That’s the part I’m genuinely worried about. I’m seeing designers hand the evaluation to the machine along with the production. They let the agent run and ship what it returns without a second look. Kris Puckett, a design manager at Stripe, explicitly rejects this: “I want them to be focused. I want it to be something that I feel is still authentically me.” He calls that quality “soul”: the thing that makes work yours and not just adequate.

Creative fluency compounds on itself. Brad Frost, the web designer and design systems author, makes the analogy to music: “Just as being able to play piano puts you in a better spot to wield a synthesizer.” An eight-year-old can vibe-code a game. A seasoned designer can vibe-code one that actually works, because they know what good looks like. And that sensibility survives the tool change. Dora Czerna, writing for UX Collective, puts it cleanly: “The pattern isn’t that expertise becomes worthless. It’s that expertise gets unbundled from the tasks that used to contain it.”

Last week I argued that the value of design is concentrating at the edges: deciding what to build and judging what was built, with agents owning the middle. Those edges turn out to be increasingly verbal and conceptual. But the eye stays. The designers who lose won’t be the ones who learned to articulate intent in a Google Doc instead of a Figma file. They’ll be the ones who stopped looking.

Happy 50th Birthday, Apple

I went to grade school at a parochial school in San Francisco’s North Beach. It was full of mostly middle class, neighborhood kids—an assortment of Italians, Chinese, and Filipinos from a ten-block radius. Half our teachers were nuns who lived in the convent on the same block. The other half were laypeople. To my surprise and delight, we had a computer lab back in the early- to mid-1980s, filled with maybe ten Apple IIe computers. It was seventh grade when I was allowed to take the class. Most computer classes at the time taught rudimentary programming in BASIC. This was a few years after I had watched the movie TRON on the big screen. And right after I had gotten my first Mac.

A few months into the class, in January, on a typical cool day in The City, I was in the computer class when the principal announced over the PA that a tragedy had struck the crew of the space shuttle Challenger. The group of us ran to the classroom where there was a television mounted at the corner. We watched the news report and the replay of the explosion—a trail of white smoke that split into a Y.

That image must have stuck with me because—well, what would a 12-year-old boy do but want to animate the launch and explosion. As my final project for my computer class, I made an animation of the launch. I mapped it out on grid paper first, and then painstakingly transferred those sprites pixel by pixel and frame by frame to the Apple IIe in my program. Over the course of days—weeks?—I typed in numbers for coordinates and letters for colors, and saved my work to a floppy disk.

Come finals time, I played the animation for my class and got some oohs and ahs. Looking back at it now, it was a dumb and tone-deaf idea. I should have animated a lamp jumping on a ball or something instead.

Anyway, that was an Apple memory I haven’t shared before on this blog. Happy birthday, Apple. Thanks for 50 years of empowering crazy people like me to make crazy things.

Some favorite Apple-related posts I’ve written:

• The Apple Design Process. My memory of working at Apple’s Graphic Design group during the time of the iPod and the PowerMac G5.

• For the Rest of Us. Apple has always done well in its marketing and advertising. This is my reflection on one of my favorite Apple spots.

• 30 Years of Mac. Don’t judge, but this is the first thing I ever drew on a Mac.

• Thank You, Steve. Here I share the story of one of the times I presented to Steve. This was an animation for MacBuddy, the Mac OS X setup assistant.

What I’m Consuming

Apple II Forever! Jason Snell traces the full arc of the Apple II, from its 1977 launch through its stubborn refusal to die even as Apple kept trying to replace it with the III, the Lisa, and the Mac. The best detail: the Apple IIe wasn’t discontinued until 1993, meaning people were still buying Apple II computers in the era of Nirvana’s Nevermind. (Jason Snell / The Verge)

I’ve been stealing from all of you, and I don’t plan to stop. Claire Alvis writes about the distance between plagiarism and inspiration, starting with a colleague’s verbal tic that migrated into her own vocabulary without her noticing. She maps the full spectrum from outright theft to pure osmosis and lands on a generous conclusion: admitting your influences publicly is its own form of credibility. (Claire Alvis)

Harness design for long-running application development. Anthropic’s Prithvi Rajasekaran describes a multi-agent system for getting Claude to build complete applications autonomously. The relevant insight for designers: when asked to evaluate their own work, AI agents consistently praise mediocre output, especially on subjective quality like visual design. The fix was separating the builder from the judge, with a dedicated evaluator agent that screenshots the running app, clicks through it, and grades against specific design criteria. It’s the evaluability problem applied to the machines themselves. (Prithvi Rajasekaran / Anthropic)

DNS Explained. Manish Bhusal couldn’t figure out why his site kept showing the old version after a server migration. Three hours of frustration later, he’d learned about TTL, DNS propagation, and the full resolution chain from root servers to your browser cache. A clear, personal walkthrough of how DNS actually works, written by someone who just figured it out the hard way. (Manish Bhusal)

Last week I wrote that most teams have deployed AI but haven’t redesigned the factory. This week on the blog, I shared stories of a handful of teams that show what the redesign actually looks like. And they’re converging on the same answer about where design value lands.

Intercom’s design team published the numbers: 90% of their pull requests are now AI-authored. John Moriarty, product design director at Intercom, draws a clear line through what that means for designers. Design’s value concentrates at the edges: deciding what to build at the start, judging whether it’s good enough at the end. Agents own the middle, the build itself.

The same pattern showed up at Cisco, where one of Jason Cyr’s directors pointed Claude Code at their design system and got 44 detection detail panels in ten minutes, every decision tracing back to real customer research. The design system was the design review. Cyr’s sharper point is about what happens upstream: “When agents can generate ten options in an hour, the person who can look at all ten and say ‘none of these—here’s why’ becomes the most important person on the team.”

Inside Figma’s own team, designer Gui Seiz describes the cost of exploration collapsing. His team pulls production code into Figma via MCP, edits visually, pushes changes back to the codebase. Seiz spends more time now planning upstream and polishing craft downstream. The rushed middle phase—where designers used to race to get specs to engineering before priorities shifted—is the part that’s compressing.

Intercom, Cisco, and Figma run different stacks and build different products, but all of them are landing on the same conclusion: the work that matters is at the beginning and the end, and agents are taking over everything in between.

NN/G has a name for it. Sarah Gibbons and Huei-Hsin Wang call it process compression: what looks like “skipping the process” is really an experienced designer running an internalized version of it at speed. “The intuition designers trust was built by the very process they dismiss.” The double diamond didn’t die. Instead, it just got faster. But the compression only works if you’ve already done the reps.

And that’s where it breaks. The middle of the workflow is also where junior designers have always learned. The wireframes and component specs were never the deliverable. They were the mechanism through which designers built judgment. I wrote about this for Fast Company: construction figured out the pipeline problem a century ago with formal apprenticeships. The contractors I work with don’t debate whether to invest in training during a downturn. They know that if they stop training apprentices, they won’t have journeymen in four years. We’re hollowing out the middle of the workflow and the middle of the career ladder at the same time.

David Hoang proposes forward-deployed squads as one response: three people, embedded on the company’s hardest problems, building working prototypes instead of producing decks. The designer finds the problem and builds the first cut of the solution. That’s an edge job. It requires the kind of judgment that only comes from years of working in the middle.

At the far end of this spectrum sits StrongDM’s Software Factory, where humans write the roadmap while agents write, test, and ship the code. No human reviews the implementation. The expectation is that every engineer on the team spends $1,000 per day on AI tokens. The middle isn’t compressed there. It’s fully delegated.

Consider AI researcher Ethan Mollick’s observation: “We can see the shape of the Thing now, but we can still influence the Thing itself.” Design doesn’t have its version of this rulebook yet. For design teams on the frontier, the edges are becoming the whole job: direction at the start, judgment at the end, agents in between. The teams writing those rules now are setting precedent for everyone else. And the people best positioned to write them are the ones who spent years in the middle before it disappeared.

What I’m Consuming

The Curse of the Cursor. Alan Kay designed the original mouse pointer for the Xerox Alto by straightening one edge of a 16x16 pixel arrow to avoid jagged lines. That shape stuck through the Alto, Star, Lisa, Mac, and Windows. Marcin Wichary traces the full history, including Apple’s 2020 attempt to redesign the cursor for iPadOS and the quiet reversal five years later. The best observation comes from a commenter on Posy’s excellent companion video on cursor history: “I’ve never thought of the mouse cursor as an arrow. My mind was blown when I realized that it was just an arrow the whole time.” (Marcin Wichary / Unsung)

Consistency Is Primitive. Christopher Butler argues that when AI makes software creation nearly instant, the economic imperative for standardization disappears. We standardized software because building once and selling many times was the only model that worked, not because uniformity made for better experiences. When creation becomes individualized, software becomes bespoke because there’s no reason for it not to be. He takes the idea further than expected, connecting it to why sufficiently advanced technology might look wildly inconsistent rather than uniform. (Christopher Butler)

Strategic, Surgical, and Scrappy: A Career Mindset for Rough Terrain. Dee McCrorey revisits a Silicon Valley internship program she helped design 15 years ago and asks what it would look like today. Her answer: the skills needed to thrive alongside AI demand a total mindset shift, from orchestrating AI workflows to building scalable knowledge systems. She also takes a sharp look at Dario Amodei’s “AI jobs apocalypse” warning and asks why it took so long. (Dee McCrorey)

How Fast Will A.I. Agents Rip Through the Economy? Ezra Klein sits down with Anthropic co-founder and head of policy Jack Clark to talk about what happens now that AI agents can program autonomously. Clark argues we’ve crossed a threshold where the models that were always being promised are actually here, and the implications for labor markets, stock prices, and organizational structure are playing out in real time. Worth the full 90 minutes. (Ezra Klein / The New York Times)

There Are Only Two Paths Left for Software. I hate this but it’s a warning sign we should all pay attention to. David George at Andreessen Horowitz lays out a stark ultimatum for software CEOs: accelerate revenue growth by 10+ points through genuinely new AI-native products, or rebuild for 40%+ true operating margins including stock comp. No middle lane. His playbook for path one reads like a war plan: find the five people in your org who will deliver 100x value, put them on information-gathering sprints, then watch which VPs get on the bus. (David George / Andreessen Horowitz)

When factories got electric motors in the 1880s, they swapped out the steam engine and changed nothing else. The floor plan stayed the same, the belt-driven machines stayed the same, the workflow stayed the same. For 30 years, output barely moved. The returns came when companies tore out the floor and redesigned everything around the new technology.

Tommy Geoco channels this history through Carlota Perez’s framework for technological revolutions: “We have swapped the motor, but we have not yet redesigned the factory.” Most teams have installed AI but haven’t changed how the work flows around it.

The numbers make it concrete. Google tells Clive Thompson that its 100,000+ developers work 10 percent faster with AI. Geoco’s studio went from one video a month to eight, though at the cost of running, as he puts it, 50 to 100 cognitive cycles a day, each with the same emotional weight. Jason Lemkin replaced a $10K/year sponsor portal in days. The difference between 10% and 10x comes down to whether you reorganized the work or just plugged a new tool into what you already had.

Most of us are still plugging in. David Oks finishes the famous ATM parable and the ending changes everything. US bank teller employment held steady through the entire ATM era, still at 332,000 as late as 2010, then collapsed to 164,000 by 2022. ATMs didn’t do it. The iPhone did. ATMs automated teller tasks better. The iPhone eliminated the reason to visit a branch at all.

AI tools that generate Figma variants or fill out documentation are automating tasks within the existing workflow. The harder question is whether AI changes the workflow itself, making some of those tasks unnecessary. Staff engineer Sean Goedecke does the math on his own profession. His assessment is that no breakthrough is required. Just incremental improvement on what AI already does.

Madison Utendahl closed her award-winning Brooklyn agency. Ten people, all women, every award possible. She didn’t close it because it failed. She closed it because the model underneath it broke. Lower fees meant more clients to hit the same revenue. More clients meant more pitching, more context-switching, more burnout. Then clients started generating moodboards with Midjourney before sending the brief. The old factory couldn’t absorb the new motor.

And the people inside these factories don’t all have the same options. Brad Frost names the privilege in the “just don’t use it” position: the people who can afford to sit this out tend to have seniority or institutional protection. The designers entering the field don’t. Anthropic’s researchers found that hiring of young workers in AI-exposed roles has quietly slowed, not because AI replaced them, but because companies stopped posting the listings. Companies are starting to figure out what a redesigned factory might look like.

So what does the redesigned factory look like? Julien Bek at Sequoia draws a clean line between intelligence work and judgment work. Intelligence work is rule-based execution AI can already handle. Judgment work is experience, taste, strategic calls. His argument: stop selling the tool and start selling the outcome. Close the books instead of selling QuickBooks. And here’s what makes the model compound: every task the autopilot completes teaches it something the copilot never learns, because the copilot hands that knowledge back to the human. The moat for the next generation of products won’t be the interface or even the model. It’ll be the accumulating dataset of domain-specific decisions.

Thu Do set up Figma MCP + Claude Code and audited her entire design system in 10 minutes. That’s the design version of this: tokens used to be nice-to-have for consistency. Now they’re infrastructure for AI-to-code workflows. The bar shifted from human readability to machine readability. That’s what tearing out the floor looks like for a design team.

I’ve been telling the comforting version of this story for months: every major tool shift expanded the field, more designers exist now than 40 years ago, the pattern holds. Oks convinced me that’s only the first half. The second half is what happens when the work reorganizes and the jobs stop needing to exist in their current form. New roles and new kinds of design work will emerge from the transition. The people in the old factory don’t automatically end up in the new one. And we’ve barely started tearing up the floor.

What I’m Consuming

The Color Statistic That’s Been Wrong for 80 Years. Kevin Muldoon decided to actually count how many colors the human eye can distinguish, instead of repeating the “10 million” figure everyone cites from a 1939 estimate nobody verified. His answer: roughly 273,000 surface colors at lab-detection threshold, and about 9,256 that you’d notice in daily life. The original estimate assumed human color perception was shaped like a box. It’s shaped like a mollusk. Three independent sources across 87 years landed on the same corrected number. The right answer was in the literature all along; it just lost the popularity contest to a bigger, rounder number. (Kevin Muldoon / UX Collective)

Mastering Midjourney: How to Create Consistent, Beautiful Brand Imagery Without Complex Prompts. Claire Vo interviews Jamey Gannon, an AI creative director who specializes in brand imagery. The counterintuitive move in Gannon’s workflow: she relies on style references and image refs rather than elaborate prompts, which produce more consistent results with less effort. The real value is in how she packages and delivers the system to clients so they can keep generating on-brand assets without her. (Claire Vo / Lenny’s Newsletter)

‘A Language We Share’ Traces a Photographic Lineage Between Gordon Parks and Beverly Price. A new exhibition at Brooklyn’s Center for Art and Advocacy puts Gordon Parks and Beverly Price in direct conversation. Parks embedded himself in American life from the 1940s onward, using photography as advocacy. Price, who picked up a camera a decade after her release from incarceration, entered into a dialogue with Parks by documenting the same Anacostia neighborhoods he’d photographed decades earlier. Both focus on children, policing, and the forces that threaten communities. Runs through June 19. (Grace Ebert / Colossal)

How You Source Great Designers. Soleio lays out how he finds design talent for startups: follow what designers are making on X, build referral nodes through a small number of well-connected design leaders, and create opinionated job pages that work while you sleep. His best advice is the simplest—ask candidates which 2-3 companies they’d interview at if they could pick any, and why. That one question tells you whether your startup can compete for them. For those of you who aren’t hiring—these are great tips on being a better candidate. (Soleio)

Designing the Shift. Ileana Marcut built a system with Claude Code to collect anonymous reflections about how AI is changing people’s work and identity. Part research, part interactive art, I love this. In her post about making it, she reveals the most instructive part is where it broke: Claude Code assumed every contributor was a designer (like Marcut) and baked that bias through the entire processing pipeline, from prompts to labels to sentiment analysis. The system was built for anyone in product, design, or creative roles, but Claude Code assumed every contributor was a designer and threaded that assumption through every layer. A PM writing about replacing designers got classified as a designer experiencing identity threat. The gaps you don’t specify are the gaps AI fills with its own assumptions. (Ileana Marcut)

How to Code with AI Agents. Peter Steinberger, creator of OpenClaw, describes the arc of working with AI coding agents: from short prompts to overengineered slash-command systems to short prompts again, but with hard-won intuition underneath. His sharpest observation is that experienced engineers sometimes struggle more than beginners because their expertise becomes a burden. They try to force their approach on the agent instead of letting it find its own path. Steinberger designs his codebase for agents to navigate, not for himself to read, and commits straight to main with local tests. (Lex Fridman)

“Defend the role, or follow the skill.” That’s Erika Flowers, writing in “The Last Typesetter“, and it’s the question I kept flibbertigibbeting on all week.

Figma’s State of the Designer 2026 calls it the “messy middle,” designers stretched between product management and engineering, occupying the translation layer between what should get built and how. It’s meant as a description of where designers are. Flowers argues it’s a description of what’s dissolving.

In another essay “Zero Stage to Orbit,” Flowers maps the design-to-development pipeline onto the rocket equation. Each stage compensates for the limitations of the previous one. Her inventory of the overhead is damning: research to inform design, design to spec for developers, specs to survive handoff, QA to catch what handoff broke, retros to discuss why QA caught so much. Fuel to carry fuel. The messy middle isn’t where designers are passing through. It’s the overhead the entire pipeline was built to manage.

What AI changes is the gravity. When the distance between intent and artifact shrinks, the translation stages that justified all those roles start to collapse. Flowers frames this directly in her Zero-Vector Design curriculum: “Speed without intention is just faster failure. Speed with intention is leverage.” The skill that survives is intent: knowing what to build and what good looks like before you open any tool.

But intent only counts when you can make it real. Shreyas Doshi makes the case that good judgment compounds, and bad judgment compounds in the wrong direction. That’s why intent matters more than tooling. Jon Kolko takes a different turn, proposing that design is becoming a literacy, a way of understanding the designed world rather than making it. I get the instinct. But understanding without making is criticism. And intent that never ships is just taste with no consequences.

The designers closing that gap are moving in two directions, and both lead out of the middle. At Notion, Brian Lovin built a prototype playground so designers encounter reality before the mockup hardens into a spec. His phrase for it: “Encounter reality as early as possible.” David Hoang sketches unconstrained in code, then uses LLMs to snap his best ideas onto production components. The design system is the finishing move, not the starting point. Cameron Worboys flattened the org to three management layers and is pushing every designer to ship production code. Quality, he says, comes from reps and speed, not from “sitting in a cave for three months pontificating about the future of software.”

These look like engineering moves. They’re designers shortening the distance between intent and reality. Former Apple designers say they can’t spend months on lickable surfaces when the platform shifts every few months. The object of obsession has to move from the artifact to the system, from polished pixels to what Weber Wong calls escaping “artifact thinking”. The craft is in the intent. The tool is whatever gets it there fastest.

I worked in a desktop publishing service bureau in San Francisco during college. Down the street, traditional typesetting shops were still hanging on, but their business was already thinning. Within a few years those shops were gone. The people who understood typography, really understood it, landed on their feet. They became art directors, production managers, early web designers. The ones who only knew the machine didn’t.

That transition was real, and it wasn’t painless. This one won’t be either. The thing I’ve loved since 7th grade is changing shape. But the intent, the taste, the judgment, that part I’m holding onto. The containers will sort themselves out.

What I’m Consuming

How I’m Dealing with the Pressure to Adopt AI as a Designer. Martin Wright’s answer to AI anxiety is patience: wait six months, see which tools survive the hype cycle, then evaluate. His sharpest advice is to protect what he calls the “middle layer” of design work, the interpretation and judgment between inputs and outputs. He cites Anthropic’s 2026 study showing developers using AI scored 17% lower on comprehension tests. The people who delegated the thinking got the job done but understood less about what they’d built. (Martin Wright)

A Soft-Landing Manual for the Second Gilded Age. JA Westenberg uses postwar Berlin as a framework for navigating AI disruption: the Trümmerfrauen cleared 75 million cubic metres of rubble by hand, and within a decade Germany was thriving. The essay lays out a practical 10-year roadmap (guaranteed minimum income, universal basic services, public AI infrastructure, algorithmic governance) and draws on Jeff Atwood’s rural GMI initiative, David Graeber’s Bullshit Jobs, and Rutger Bregman’s Utopia for Realists. Neither doomer nor accelerationist, just stubbornly pragmatic. (JA Westenberg)

The Mythology of Conscious AI. Anil Seth, winner of the 2025 Berggruen Prize essay competition, argues that consciousness is unlikely to emerge from standard digital computation. Brains are not computers in any straightforward sense, and life, embodiment, and non-algorithmic processes may be prerequisites for conscious experience. A useful corrective to the anthropomorphizing that creeps in every time an AI model produces something that feels aware. (Anil Seth / NOEMA)

Bespoke AI Models Are the Next Big Thing in Filmmaking. Netflix acquired Ben Affleck’s AI startup InterPositive for roughly $600 million. The pitch: bespoke models trained on a production’s own dailies, so filmmakers can tweak lighting, remove rigging, or replace backgrounds in post without the uncanny slop of general-purpose generators. Charles Pulliam-Moore is rightfully skeptical about whether “empowering creatives” translates into actual benefits for the people doing the work. (Charles Pulliam-Moore / The Verge)

Grammarly Is Using Our Identities Without Permission. Grammarly’s “Expert Review” feature surfaces AI-generated writing advice “inspired by” real people, including Nilay Patel, Tom Warren, and Casey Newton, without their knowledge or consent. Stevie Bonifield found outdated job titles and fabricated expertise descriptions. Superhuman’s defense: the experts appear because their work is “publicly available and widely cited.” That’s not the justification they think it is. (Stevie Bonifield / The Verge)

Last week I wrote that the craft is still here, it’s just a different shape now. This week I found pieces that answered the obvious follow-up: what shape?

Figma’s latest hiring study shows 82% of organizations need more designers, driven partly by AI creating new product surface area. Their State of the Designer 2026 finds designers occupying a “messy middle” between product management and engineering, with 91% saying clear goals help them do their best work and 87% saying decision-making power boosts performance. Demand is up. The tools have compressed execution. Designers should be thriving.

So why does it feel like the profession is having an identity crisis on a weekly news cycle?

The territory has been redrawn, and it’s more strategic than what we had before. Start with time. Jenny Wen, head of design at Claude, says her team is spending 30-40% of their time on mockups now, down from 60-70%. The rest goes to pairing with engineers and implementation. That tracks with the economics: when the cost of producing software collapses, the artifact stops being the job. The judgment is the job. And judgment now has better tools to express itself. Spec-driven development lets designers write intent into structured documents that AI agents build against directly. The spec isn’t a suggestion. It’s enforceable. The same logic applies to design systems: instead of compressing your reasoning into tokens and components, you can ship the reasoning itself when you design in plain English. And when interfaces start acting autonomously, designers are the ones who should be setting the confidence thresholds: how sure does the system need to be before it acts without asking?

That’s a lot of new ground. Judgment, reasoning, thresholds, specs. More strategic than a Figma file ever was.

And yet. Nicole Michaelis called it out: the profession is trying on identities like clothes in a dressing room. “Monday, it’s all about prototypes. Thursday, it’s vibe coding. Friday, we’re preaching that output no longer matters” and we should all be strategists. By next Monday, we’re debating soft skills. The retreat to craft and taste as differentiators is understandable, but craft is the baseline, not the selling point. Debating what makes us special instead of demonstrating it is the creative class version of what’s happening at the company level too: wearing what Claire Vo called “the bows and ribbons” of transformation while the organizational bones stay calcified. The tools are good enough. The gap is behavioral.

The territory is there. The data around the demand confirms it. On my own team, the shift is starting to happen: designers are spending more time with customers, more time in research and discovery, and pushing further into product strategy. Less time in Figma, more time defining what we’re building and why. That’s what claiming the new ground actually looks like. Not a theoretical expansion of the role, but a practical one: you show up earlier in the process, you own more of the decision, and the mockup becomes one artifact among many instead of the whole deliverable. The designers who are making this shift won’t need to argue for their seat at the table. They’ll already be sitting in it.

Spec-Driven Development: It Looks Like Waterfall (And I Feel Fine)

We’ve been talking a lot about agentic engineering, how software is now getting built with AI. As I look to see how design can complement this new development paradigm, a newish methodology called spec-driven development caught my eye. The idea is straightforward: you write a detailed specification first, then AI agents generate the code from it. The specification becomes the source of truth, not the code.

My first reaction when I started reading about SDD was: wait, isn’t this just waterfall?

Seriously. You gather requirements. You write them down in a structured document. You hand that document to someone (or something) that builds to spec. That’s the waterfall pattern. We spent two decades running away from it, and now it’s back wearing a blue Patagonia vest and calling itself a methodology.

But there is a difference.

Continue Reading

What I’m Consuming

A24’s Empire of Auteurs. Alex Barasch profiles A24’s evolution from scrappy indie distributor to a $3.5-billion studio that develops films with directors rather than attaching them to finished scripts. Noah Sacco, the head of film, is the connective tissue: beloved by filmmakers for trusting their instincts while knowing when to push. The tension between scaling up (Dwayne Johnson, Elden Ring adaptation) and the indie roots that made the brand (Kelly Reichardt calls the shift “a heartbreak”) is the piece’s real subject. (Alex Barasch / The New Yorker)

In But Not Of. Christopher Butler writes about the science fiction novels that have stayed with him across decades and realizes they share a common thread: characters who witness transformation without the power to change it. He maps this onto his own position as a technologist who uses the tools daily but can see where the narratives about technology diverge from reality. A personal essay about the discipline of paying attention when everyone around you is optimizing for speed. (Christopher Butler)

On Software Quality. Nick Heer catalogs the bugs he hits daily across Apple’s ecosystem: Finder glitches, AirDrop failures, Safari oddities, Siri asking him to tap the screen to pause music while his hands are covered in food. Jason Snell’s annual survey pegs Apple’s software quality at a B-minus, and Heer argues even that’s generous. The kicker: Apple ships hardware that feels bulletproof, which makes the software gap that much more glaring. (Nick Heer / Pixel Envy)

The AI Bubble Is Bursting. Hugh Howey separates three questions people conflate: Will AI disappear? (No.) Is the hype overblown? (Yes, 80% of firms polled show no productivity gains.) Will the investment pay off? (Unclear, and seven trillion dollars in projected data center spending is a lot of GPUs to strand.) His sharpest point: the browser wars taught us that customers won’t pay extra when free is good enough, and Google has every incentive to give AI away. (Hugh Howey)

Vibe Coding and the Maker Movement. Sachin draws a structural parallel between vibe coding and the Maker Movement of 2005-2015, then identifies the key difference: vibe coding skipped the “scenius” phase where hobbyists develop judgment through play. The tools went straight to production before anyone had time to learn what’s worth building. His reframe—vibe coding as “consumption of surplus intelligence”—is the most interesting lens I’ve seen on why building with AI can feel simultaneously productive and hollow. (Sachin / Technically)

Developers and designers are independently grieving the same thing right now, and it took me a while to realize they’re not mourning the skill. They’re mourning the tribe. Dave Gauer wrote about losing his social identity as a programmer—still writing code, still loving it, but unable to recognize the culture around it. The community that used to care about the craft now feels like it’s about speed, or pulling a slot machine lever on prompts. Swap “programming” for “design” and you have the conversation I’ve been observing and living all year.

Nolan Lawson made the uncomfortable part explicit: the tools work. They write code better than most of developers and certainly most designers. And the economic gravity is real—I’ve watched it with every industry shift I’ve lived through. Desktop publishing, print to web, mobile apps. Each time, the people with mortgages and families learned the new tools first because they couldn’t afford not to. You don’t get to sit out a paradigm shift when your family depends on your paycheck. You adapt on company time and mourn the old craft on your own. The idealism erodes fast when the market has already moved. Geoffrey Huntley calculated that AI-powered software development now costs $10.42 an hour—less than minimum wage. Swap “software developer” for “designer” and the timeline gets uncomfortable. Lawson frames what’s left as two options: abstain on principle or capitulate for the paycheck.

I don’t buy the binary. There’s a third path—use the tools to expand what your craft can produce—and I keep running into people who are already walking it.

Anton Sten, a designer, built his own invoicing tool in two days with Claude and Cursor. Not by following SaaS patterns—by throwing them out. The tool serves one user. It parses his contracts, drafts his invoices, answers questions about his billing history. He stopped squeezing his workflow into someone else’s product and started making exactly what he needed. And like me, he used those same tools to rebuild his website.

The same expansion is reshaping how design work gets done. Édouard Wautier’s team at Dust prototypes directly in code, skipping Figma after the initial sketch. His description—more like clay than drafting—captures something the grief narrative misses: working in code with AI agents can be more tactile, not less. You shape, test, feel, adjust. The artifact becomes the thing, not a picture of the thing. Jonny Burch argues this is where the whole field is heading—a future after Figma where the source of truth lives in code. I’m seeing it on my own team: engineers ship working features in days, and the design phase is now the slowest part of the cycle. I don’t think most design teams have reckoned with what that means yet.

Kieran Klaassen might be the furthest down this road. His compound engineering system treats every AI session as a teaching opportunity: plan, build, review, codify. Each fix gets captured so the AI doesn’t repeat the same mistake. He hasn’t opened a code editor in three months—not because he’s careless, but because he built trust through iteration. That’s what the third path looks like when you fully commit. You’re not reviewing every line. You’re building a relationship with the tool, and the tool gets better.

I highlighted stories from developers more than usual this week because what’s happened in engineering is predictive of what is beginning to happen in product design.

There is real grief. We’re all feeling it. The discipline is fracturing. But I think the people mourning are mourning a specific version of the craft—the one defined by the tools and rituals of the last decade. The craft itself is still here. It’s just a different shape now.

What I’m Consuming

How to shoot a screen using a board of keys. I grew up with a Mac 512K and still remember ⌘⇧1 to eject the floppy disk—so learning that’s why the screenshot shortcuts start at 3 was a revelation. Marcin Wichary traces the whole lineage, from the 1986 floppy eject commands through 2018’s Screenshot app. The real gem is Acorn 8’s ⌘⇧7, which captures each window as a separate layer you can recompose afterward. That’s software craft. (Marcin Wichary / Unsung)

About Letters to a Young Creator. The Steve Jobs Archive collected candid letters from over 30 creative figures—Jony Ive, Tim Cook, Ed Catmull, Dieter Rams, Es Devlin, and others—answering questions from SJA Fellows about what it takes to make something great. The title nods to Rilke’s Letters to a Young Poet, one of Steve’s favorites. Originally released as small-press editions, the full collection is now available online. (Steve Jobs Archive)

Statement from Dario Amodei on our discussions with the Department of War. Anthropic drew two lines it won’t cross: mass domestic surveillance and fully autonomous weapons. Under threat from the DoD, Anthropic refused. Anti-mass surveillance and anti-autonomous murderbots are moral red lines to have, and I’m glad they didn’t cave. On Friday, the administration indeed canceled the contract and labeled Anthropic as a “supply-chain risk to national security,” a designation reserved for adversarial foreign companies. (Dario Amodei / Anthropic)

SaaS Isn’t Dead. But the Way You Used to Win in B2B? That’s Gone. The old SaaS playbook—lock up a category early, grind to $100M ARR over five years, coast on inertia—is done. AI-native startups are reaching $100M in twelve months with 50-person teams, shipping weekly, and winning on ROI instead of switching costs. Incumbents who haven’t shipped meaningful improvements in two years are watching customers finally find a reason and a path to leave. (Jason Lemkin / SaaStr)

I published “Product Design Is Changing“ on Monday. LinkedIn largely agreed. Reddit was hostile. A former creative director I worked with years ago left a comment: “So much for years of craft and imagination... I didn’t sign up for this.”

He’s right. None of us signed up for it. But the reactions kept splitting along the same fault line: people who agree the shift is happening but are still opening Figma every morning and drawing pictures of software. That’s the thing I’ve been chewing on all week—why the default workflow is so sticky even when everyone can see it changing.

For decades, the job has been: design in one tool, hand off to engineers who rebuild it in another. We draw pictures of apps. We sweat over pixels in those pictures. Then someone else translates them into the thing that actually ships. Every stage of that handoff generates waste—the alignment meetings, the redlines, the QA passes comparing mockups to code. Laura Klein’s NN/g piece on empowered teams catalogs where all that overhead lands: PMs spending 70% of their time coordinating, fragmented squads producing products that feel like they were designed by strangers. That’s what happens when the picture and the product are two different artifacts maintained by two different groups of people.

Some companies have stopped doing it that way entirely. At Intercom, every designer now ships code to production. Zero did 18 months ago. The CPO’s test for any role: what would a startup founded today do here? Over at Anthropic, teams group-sculpt living prototypes with no spec and no roadmap beyond 90 days—and shipped Claude Cowork ten days after someone first had the idea. Dan Shipper’s Every runs four products with single-person teams and 99% AI-written code. Amazon’s two-pizza team just became a two-slice team. In all of these, the designer works in the final medium. No pictures. No handoff.

So why is everyone else still drawing? Part of it is that the advice hasn’t caught up. A UX Magazine piece told designers to sharpen their critical thinking and be the conscience in the room—the kind of thing you can agree with without changing anything about your day-to-day. Jan Tegze’s piece on job shrinkage gets at why that advice doesn’t land. He quotes a CEO: “Our senior people and our junior people are equally lost when we ask them what we should do. The seniors are just more articulate about their uncertainty.” The mockup workflow felt like the job. Letting go of it means admitting that a lot of what filled the day was production, not strategy.

The piece I’d actually hand a designer is Tommaso Nervegna’s Claude Code guide, which shows what working in the final medium looks like in practice: spec-driven development through conversation with an AI agent, design process applied to code instead of Figma. The designer’s value is in the questions asked before any code gets written. That’s the craft now—not the mockup, but the specification and the judgment call.

Matt Shumer wrote a piece saying that the disruption tech workers are living through is heading for every knowledge-work profession. Design sits in an interesting middle—the interfaces people see and touch still need human judgment, but the production layer underneath keeps compressing.

Here’s what I keep coming back to. A designer can read this newsletter, agree with all of it, and still open Figma on Monday and draw pictures of software exactly the way they did last year. The shift demands more than updated beliefs. It demands picking up the actual material.

ASCII Me

Over the past couple months, I’ve noticed a wave of ASCII-related projects show up on my feeds. WTH is ASCII? It’s the basic set of letters, numbers, and symbols that old-school computers agreed to use for text.

I think it’s sort of a halo effect from Claude Code and the nostalgia designers and developers have for text-based terminals. Anyway, I wrote a roundup of stuff that’s caught my eye.

Read the article

What I’m Consuming

Frames That Force Us to Look. A piece connecting the 1972 Napalm Girl photo to the image of five-year-old Liam Conejo Ramos being taken by ICE agents—and why still photographs move public opinion in ways video can’t. Deb Aldrich traces the lineage through Kent State and Pete Souza’s Obama-era work, then warns that government-manipulated images are actively undermining photographic trust. (Susan Milligan / PRINT Magazine)

Agentic Lovemarks: How Brands Can Top Both Human and AI-Driven Shortlists. Arjan Kapteijns builds on Thomas Marzano’s Brand Constitutions manifesto to ask what happens to Kevin Roberts’ classic Lovemarks framework when AI agents start mediating brand discovery. His argument: brands now need to speak to two audiences simultaneously—the humans whose hearts they want to win and the intelligent systems acting on those humans’ behalf. Love stays human, but respect has to become machine-readable. (Arjan Kapteijns / Brandingmag)

These 3 ‘Addictive’ Social Media UX Features Are on Trial. Grace Snelling covers the Los Angeles lawsuit arguing that infinite scroll, ephemeral content, and algorithmic recommendations are deliberately engineered to be addictive—especially to kids. Snap and TikTok already settled; Meta and Google are the remaining defendants. The legal theory frames social media UX as a public nuisance, borrowing from the playbook used against the cigarette industry. (Grace Snelling / Fast Company)

Who Cares if Matt Damon’s ‘Odyssey’ Helmet Is Historically Accurate? (Gift link) Paul McAdory examines why audiences nitpick historical accuracy in film adaptations of Homer and Brontë, arguing that the obsession with fidelity reflects something deeper—a desire for stable reference points in an unstable present. When you can’t control where the world is heading, policing whether Odysseus has the right helmet becomes a strange kind of comfort. (Paul McAdory / The New York Times)

I’ve been thinking about what actually makes someone good at their job. Not competent, but good. The kind of good where you look at a screen and know something is off before you can articulate why. The kind where you can tell a product decision is wrong from across the room.

Christina Wodtke calls this “compressed experience”—years of pattern recognition folded into a split-second gut reaction. A seasoned designer isn’t being mystical when they sense a flow is broken. They’ve just done it enough times that the signals are automatic. And the only way to build that compression is reps. Not reading about reps. Actual reps.

Peter Yang landed on the same idea from a different direction. His 25 product beliefs after a decade at Roblox, Reddit, Amazon, and Meta boil down to a claim that makes credentialists uncomfortable: the only thing that matters is what you’ve shipped and your ideas to improve the product. He estimates fewer than 10% of PMs actually dogfood their own product weekly. If you’re not paying attention to the thing you ship, you’re not building judgment. You’re just accumulating tenure.

Now add AI to this. Anthropic published a randomized controlled trial on junior engineers learning a new Python library. The group using AI assistance scored 17% lower than those who coded by hand. They didn’t finish meaningfully faster, either. The biggest gap was in debugging—the exact skill you need most when your job is overseeing AI output. Anthropic’s researchers said it plainly: AI may accelerate productivity on skills you already have while hindering the acquisition of new ones.

Daniel Miessler put this more viscerally, riffing on a Karpathy interview about entropy. Adults “collapse” over time—revisiting the same thoughts, narrowing their aperture, ossifying. AI makes it worse. You’re outsourcing your thinking to a system that learned by averaging the internet. The outputs pass a first glance. But nothing in there will surprise anyone because the model optimizes for the most statistically probable next token.

So how do you stay sharp while using the tools? Daniel Rosenberg has one approach: design the conceptual model before you touch a screen. Objects are nouns. Actions are verbs. Attributes are adjectives. Your interface is a language before it’s a layout, and if you get the grammar wrong, no amount of visual polish saves you. Jumping straight to screens means making those language decisions implicitly, without realizing it. Rosenberg’s argument is that you should write the grammar first—literally work in text before pixels.

Patrick Morgan comes at it differently: externalize your reasoning in plain text so AI can actually work with it. Three Markdown files—process, taste, raw thinking. Your judgment has to exist somewhere outside your head, and it has to be legible. Designers externalize visual thinking all the time with moodboards and component libraries. But we rarely write down why we made the choices we made. Do that, and AI amplifies your judgment instead of replacing it.

Erika Flowers makes the organizational version of this argument. On the Invisible Machines podcast, she compares AI adoption to roofing: before you can install anything, you spend a week building scaffolding, setting up tarps, rigging safety harnesses. Nobody wants to fund that part. Everyone wants the flashy use case. But the scaffolding—data integration, governance, connected workflows—is what makes the use case actually work.

All of it comes back to the same thing. The slow, sometimes painful work that builds real understanding is the work AI makes easiest to skip. And skipping it has a cost that doesn’t show up until later—when you’re debugging output you don’t understand, or shipping a product you can’t explain, or deploying an AI feature on top of infrastructure that was never ready for it.

I keep coming back to a line from Anthropic’s study: “Cognitive effort—and even getting painfully stuck—is likely important for fostering mastery.” Getting stuck is the apprenticeship. The reps are the point. Use AI to do more of them, not fewer.

What Wall Street Gets Wrong About SaaS and What’s Next

I wrote two pieces on the SaaS market. What Wall Street Gets Wrong About SaaS makes the case that AI won't kill enterprise software, for the same reason Zapier didn't kill systems integrators and Squarespace didn't kill web designers—capability doesn't equal DIY, and every wave that was supposed to eliminate professional work ended up expanding the market instead. Then in What's Next in Vertical SaaS, I pulled that thread forward with Charlie Warren, Eli Dukes, and my colleague Duncan Grazier: if SaaS survives, where does defensibility live next? Their answer is the data recipe—the specific chain of decisions, intent, and context that turns raw information into economically valuable results. The model layer isn't the moat anymore. The thinking behind the output is.

What I’m Consuming

A History of the Permalink. In early 2000, Jason Kottke added permanent links to individual blog posts, which Caroline Van Oosten highlighted on prolific.org and helped popularize the idea of persistent URLs. Folks at Blogger, Matt Haughey, Evan Williams, and Paul Bausch developed a code solution using anchor IDs to link to single entries, enabling true permalinks across blogs. By late March 2000, Bausch published an official guide, and the concept spread, ultimately evolving into what we now call permalinks. (Jay / The History of the Web)

Does Software Piracy Exist? Matthew Butterick, a type designer, argues the optimal level of software piracy is greater than zero. Zero piracy means either you never shipped or you gave it away free—both produce $0 in revenue. Maximum revenue happens in the middle, where some piracy occurs. And most people downloading your fonts from pirate sites were never going to pay anyway. (Matthew Butterick)

The Moylan Arrow of Software. After the death of James Moylan, the designer of that little arrow on your fuel gauge showing which side the gas cap is on, Marcin Wichary asks: what’s the software equivalent? His pick is iOS’s Security Code AutoFill—it benefits everyone, solves a real friction, shows up at the right time, and once you know it’s there, you love it forever. (Marcin Wichary / Unsung)

Wes Cook and the Centralia McDonald’s Mural. The full story behind Cabel Sasser’s XOXO 2024 talk: a decade-long effort to save a hand-painted McDonald’s mural from a remodel, a canvas that peeled right off the wall, a hidden “COMING SOON!” message from 1980, and a suspiciously detailed cow udder that nearly triggered a lawsuit. A wonderful read. (Cabel Sasser)

This week, Anthropic shipped new plug-ins for Claude Cowork—the kind that review contracts and run financial analysis, work that software companies have built entire businesses around. The stock market noticed. Salesforce, Intuit, and ServiceNow all dropped double digits; the IGV software index is down roughly 30% from its September peak. If AI can do the work directly, the tools that organized the work are worth less. Or so the market thinks.

That’s the macro story. The micro version has been showing up in everything I linked to this week. In the not-so-distant past, the startups that survived were the ones that shipped fast. Build and deploy faster than your competition, you won. That was the moat. With Claude Code, Codex, Antigravity and other agentic coding tools, teams can now ship in weeks what used to take quarters. And the result is that a lot of teams are building the wrong things faster than ever.

Gale Robins opened the week with a scene I’ve watched play out from both sides of the table: a team ships three features on time, hits all their velocity metrics, and can’t answer the question “what problem do these features solve?“ She walks through 19 specific judgment points in the product discovery process where human decisions determine whether teams build the right thing or waste months on the wrong one. Her point: AI didn’t eliminate the need for those decisions. It just compressed the timeline so the cost of skipping them shows up faster. Christina Wodtke made the same case from the research side—a designer spent two weeks vibe-coding a gratitude journaling app with confetti animations and gentle notifications, showed it to users, and learned that nobody journals. Two weeks building the wrong thing. Wodtke’s advice: sit down with five to ten people, shut up, and listen. You’ll build less. It’ll be the right thing.

Nielsen Norman Group’s annual State of UX report extended this to the interface itself. Design systems standardized the components; AI-mediated interactions sit on top of the screen. Polished UI is no longer a differentiator. Anyone can produce a decent-looking interface now. Kai Wong interviewed 22 design leaders who all said some version of the same thing: a PM went to an AI tool and came back with something that looked 90% done to an untrained eye. The superpower didn’t disappear. It just stopped being rare enough to carry a career.

So where does our value as designers go? Deeper. Into the work that’s hardest to see and hardest to automate: research, judgment, the ability to articulate why this interface works and tie that explanation to a business outcome. I see this on my own team. The designers who are hardest to replace aren’t the ones who produce the most polished screens. They’re the ones who can walk into a room with a PM and an engineer, explain why something needs to change, and connect that explanation to what users actually do. That skill has always mattered, but it used to be a nice-to-have on top of solid craft. Now it is the craft.

The trouble is that this kind of work is almost invisible. I linked to a second piece this week by Kai Wong about making strategic design work legible to the people who control headcount and budgets. His advice is to stop presenting design decisions in design terms—don’t explain that “Option A follows the Gestalt principle of proximity”; say it reduces checkout from five steps to three. Hardik Pandya made the same argument about coordination work more broadly—the person who writes the doc that gives a project its shape, who closes context gaps in one-on-ones before they become conflicts, never shows up in the launch email. Credit flows to the people whose contributions are easy to describe. The invisible work is the hardest to defend in a budget meeting, and it’s exactly the work that matters most now.

Meanwhile, the posts about AI interfaces kept circling the same drain. Katya Korovkina argued that chatbot-first thinking is leading teams to ship worse experiences than what they’re replacing—prompt-based products work best for users who already know how to ask the right question, which excludes a lot of people. Julian Scaff reframed Fitts’s Law for AI: the friction didn’t disappear, it just moved from the screen to the gap between what a user intends and what the system understands. When a user stares at a blank text field and doesn’t know what to type, that’s distance. And Tushar Deshmukh described enterprise teams whose AI-powered dashboards got ignored because they violated twenty years of cognitive routine. The fix in every case was the same: layer the new on top of the familiar. Which requires knowing what’s familiar. Which requires research. There’s no shortcut.

Andy Coenen’s isometric NYC project put it most concisely from the creative side: “If you can push a button and get content, then that content is a commodity. Its value is next to zero.” When the hard parts become easy, the differentiator becomes love—and love is just another word for the judgment, care, and depth that no model can generate on its own.

The optimistic read on all of this is that the skills designers have always claimed to value—empathy, research, strategic thinking—are now the only ones that matter. The uncomfortable read is that a lot of us have been coasting on execution skills that are rapidly losing their premium. Wall Street looked at the same dynamic this week and concluded that software companies are finished. I’m not so sure.

Most businesses don’t want to be in the business of building and maintaining software—that is not their core competency. They’d rather continue to make widgets than software to support making widgets. The SaaS companies that survive will be the ones that understand this: their value is not the software. It’s the business outcomes the software enables. As designers, we must understand the same: our worth is not in crafting polished UIs, but building experiences that unlock real business value for clients.

OpenClaw and the Agentic Future

Recently, an autonomous AI agent named OpenClaw (fka Clawd, fka Moltbot) took the tech community by storm, including a run on Mac minis as enthusiasts snapped them up to host OpenClaw 24/7. In case you’re not familiar, the app is a mostly unrestricted AI agent that lives and runs on your local machine or on a server—self-hosted, homelab, or otherwise. What can it do? You can connect it to your Google accounts, social media accounts, and others and it can act as your pretty capable AI assistant. It can even code its own capabilities. You chat with it through any number of familiar chat apps like Slack, Telegram, WhatsApp, and even iMessage.

Feeling some real FOMO, I decided to give it a try.

Read the full essay

What I’m Consuming

How OpenClaw's Creator Uses AI to Run His Life in 40 Minutes. Peter Steinberger argues his creation could become a personal operating system, but cautions against “slop town” agent orchestration and emphasizes keeping a human in the loop and learning through iteration. Peter Yang interviews him for his podcast. (Peter Yang)

On the Inevitability of Design Changing Forever. Andrew Boardman outlines three possible futures for graphic design in the AI age: (1) a fully democratic design era where AI-powered platforms democratize production and designers become efficiency-focused engineers; (2) the death of design, with AI replacing most design work and traditional roles vanishing; and (3) design as a defiant, craft-driven practice where independent designers thrive despite broader upheaval. He cites signs like shrinking entry-level opportunities and AI tools enabling rapid tool-building, arguing that change is inevitable. (Andrew Boardman / Dear Designer)

The Hidden History of Women Game Designers. Educational board and card games in the late eighteenth and nineteenth centuries opened a rare professional and intellectual space for women, especially in music education, where they designed elaborate games to teach theory concepts through play. A fascinating read. (Carmel Raz / JSTOR Daily)

Times New Resistance. While I don’t condone modifying someone else’s computer without them knowing, this is a mischievous/funny/interesting piece of software as art by Abby Haddican. She’s created a version of Times New Roman that “autocorrects the autocrats.” When someone types certain oft-used MAGA terms, their computer will turn those words and phrases into liberal resistance terms. For example, when installed, this font will turn “Stephen Miller” into “Nosferatu” and “Mar-a-Lago” into “the Fourth Circle of Hell.” (Abby Haddican)