Peter Sodja: Writing

A developer's free Adobe clones, built with AI

Fri, 09 Oct 2026 12:00:00 GMT

Headlines declaring that software is over (the usual clickbait). The developer Brandon Thomas made project ArtCraft, and it is seven free Rust apps modeled on Photoshop, Illustrator, Premiere and the rest of Adobe's lineup. Reportedly he was charged one cancellation fee too many (didn't we all).

If an AI can reproduce an interface in weeks, the interface was never the moat. Users prefer software that works like what they already know, so cloning Adobe's layout is a sound strategy. But the pitch is compatibility, and I see no sign of anyone asking what a photo editor should look like if you started today with an agent driving it. The one part that isn't a copy, by the project's own description, is that every app can be driven from a CLI, a JSON control channel or an MCP server. That is the bit I'd watch. Agents working through MCP is exactly what impressed me in Figma.

What keeps people on Adobe is boring: file formats, plugins, other people's files. Nobody leaves that behind because the menus are free elsewhere. Is it a real threat to Adobe? An alpha with a slogan doesn't settle that either way. The most effective Adobe competitor may still be Adobe's own cancellation flow.

Related, I found REA on github with title "Reverse Engineer Anything". It hands your coding agent tools to inspect a program and explain what it does then recreate it. Amazing, Software is over.

Why everything is starting to look the same

Fri, 09 Oct 2026 12:00:00 GMT

Sameness gets a bad rep sometimes. It is boring. Often it's just what success looks like.

Physics explains a plane. The crab analogy is fun, but what the post skips is the UX reason. Jakob's Law: users spend most of their time on other products, and they want yours to work the same way. A familiar pattern is a learning curve you don't have to pay. I wrote once, jokingly, that good UX is overrated. This is the serious version of the joke. Users benefit from boring.

So sameness works, but the interesting part is who decides when to break it, and why.

Destroy any website, for fun

Thu, 08 Oct 2026 12:00:00 GMT

You type an address, a stick figure appears, and you shoot the website. That's Destroy Any Website by Hugo Duprez. Pistols, grenades, a flamethrower, multiplayer.

Creatives are doing more with AI

Wed, 07 Oct 2026 12:00:00 GMT

Contra Labs says taste, perspective, judgment and direction "became more valuable" with AI. I agree with that. I just don't think a survey can show it, and the line is old anyway.

The Creative Intelligence 2026 report says it surveyed 150k+ creatives, combined with Contra's marketplace data: AI adoption over 70%, project sizes for "AI-powered" creatives up over 40%, and AI adopters earning more.

That might be BS:

There is no definition of "AI-powered".
Contra runs a talent marketplace, and its benchmark page pitches creatives' judgment as human data to train models, with a section for AI labs. "Creatives thrive with AI" is a nice story for both.
"Doing more" is also self-reported, and "I feel productive" is the weakest evidence there is.

It might be true, too. But creatives might also be doing a lot of stuff that shouldn't be done.

My homelab, and an AI agent with root

Wed, 07 Oct 2026 12:00:00 GMT

My homelab hobby is standing still. Just normal life (kids, house, recreation) doesn't give me enough extra time to tinker. Everything works: Nextcloud, Jellyfin, webserver… But I think I already forgot how everything is set up, and even though people advised me to write stuff down, I thought I'd never forget basic stuff like how the backup is set up. I did.

I'm thinking of installing an AI agent and giving it root access.

Might be fun.

Design tokens: understand the structure first

Tue, 06 Oct 2026 12:00:00 GMT

Nate Baldwin argues that design tokens should emerge from reusable interface structures, rather than being created through naming conventions. His team at Adobe Spectrum identified recurring patterns before deciding which design decisions needed tokens. Which makes sense. But the process was messier than “structure first” suggests. You cannot get around that. His team began with semantic names like “gap-tight” and “gap-loose,” then discovered through discussions with designers and engineers that people thought about spacing in terms of relationships, such as the space between a label and a field. Naming debates helped them discover the structure.

Arguments about token names are often disagreements about the design system itself. Teams may have different ideas about what should be shared, how components relate, and where responsibilities belong. No convention or tool can determine the right structure for every team.

The challenge is finding the balance between overly generic tokens and overly specific component tokens. Shared structural patterns can help, but in the end there are subjective decisions that might follow patterns and might not even be worth deliberation.

Many agents need many views

Mon, 05 Oct 2026 12:00:00 GMT

Chat is a log, not a view. For one agent doing one job, a doc and a chat is enough. It breaks when agents start managing other agents. That's where Luke Wroblewski's Many Agents, Many Views starts. His idea is views shaped to the task, like diagrams of the plan or before-and-after screenshots. His example is a tool called Intent.

I like the shared design-system note best. One source of truth that humans and agents both read, like a shared Figma file.

Two things bug me. The overview is still open: how do you watch twenty agents without staring at twenty things? And cheap review is a trap. A screenshot is easy to approve, but matching pixels isn't working behavior, and a diagram is just the agent's own plan, not a promise.

A configurable workspace also hands the design work to the user. I'd rather have good defaults. This is a reading of the article, not a review of the tool.

Just let me in, please

Fri, 02 Oct 2026 12:00:00 GMT

Max van IJsselmuiden's Just let me in, please is a good rant about login friction. Codes that send you to your inbox. A dashboard that forgets which page you wanted. A lot of it looks like ritual to me. Somebody added a step once, everybody copied it, and now we call it security.

I agree with the mood but not the fix: let people pick their sign-in method up front. That doesn't remove the question, it hands it to you. Did I sign up with Google or a password? His own example has more than a dozen sign-in buttons.

Passkeys are the likely way out. The FIDO Alliance claims 93% sign-in success versus 63% for other methods. That's a survey of nine of its own members, by the people promoting passkeys. I wouldn't quote it as proof.

And dumping everyone on a generic dashboard after login is just nobody owning the whole journey.

AI and design systems, a course

Thu, 01 Oct 2026 12:00:00 GMT

AI & Design Systems is a course by Brad Frost, Ian Frost and TJ Pitre. It costs $1,495. I haven't taken it, so this isn't a review.

The pitch is that AI is fast but "can make a huge mess". Convenient problem when you sell the fix. Still, the idea is right. An AI agent is a new team member who hasn't read the docs. Of course it makes a mess.

I'm skeptical of the rest. The big test result comes from the course authors' own tool, graded by that same tool, which is close to circular. Schemas and automated checks sound like ordinary engineering. And video about tools that change every month ages fast. A perfect score also only means the rules were followed, not that the screen is any good.

If you've taken it, tell me if it was worth it.

A model that tells you how sure it is

Thu, 01 Oct 2026 12:00:00 GMT

Jev for Designers by MC Dean is about Jev, a model from TypeSafe that doesn't write text. You give it a question and the allowed answers, and it gives back a choice with a probability. The aim is that 70% means right 70% of the time.

The demo has a "maybe zone" where the unsure items sit in the middle. I like that. Doubt drawn into the interface instead of hidden behind a smooth paragraph.

I haven't used it, I'm going by her write-up and The Register. The accuracy, speed and price numbers all come from TypeSafe, and The Register cites no independent check. Also, honest isn't the same as correct. Dean says Jev gets about 68% right on TypeSafe's own test, so it's still wrong a third of the time.

And 70% means nothing until someone decides what happens next. Ship above one line, drop below another, a person checks the middle. You pick those lines, not the model.

How to de-slopify your designs

Wed, 30 Sep 2026 12:00:00 GMT

Slop is what you get when nobody decided anything. Maya Brennan's How to de-slopify your designs agrees: slop "strips out considered decisions." I don't think her fix works.

Her tips are mostly: pick a palette, a font, some icons, some photos. But swapping those only changes the default, it isn't a decision. And her stack is popular enough to become the next look everyone recognizes.

Typing "make it look less vibecoded" is "not super effective," she says. Of course not. The model doesn't know what you like, so you get the average again. Someone has to bring the taste, and she tells non-designers to "hire one."

Dieter Rams has a better test: nothing must be arbitrary or left to chance. Ask who chose each element, and why. Removing the tells isn't designing.

Goodbye WordPress

Wed, 30 Sep 2026 12:00:00 GMT

Nolan Lawson moved his blog off WordPress to a static Astro site, and Claude did most of the work. He says this should make WordPress "nervous". I doubt it.

One developer with an agent and a Markdown habit isn't a trend. He also runs his own Mastodon server. That's far from a bakery owner who has never opened a terminal. Easy for him says little about her.

The cost story is soft too. Hosting is cheap, but his time and tokens aren't free, and the post prices none of it.

What I like is what he kept human. No analytics, because watching the numbers was bad for his mental health, he says. No LLM near his prose. No agent picks that by default.

He calls the design "window dressing". I'd argue with that. Design is how the writing gets read.

Are we going to use the same desktop UX forever?

Tue, 29 Sep 2026 12:00:00 GMT

Scott Jenson's Akademy talk, Are we really going to use the same Desktop UX forever?, has a good complaint: the desktop has no working memory. Copy a few too many things and the clipboard drops the old ones. He calls it the curse of direct manipulation. I haven't watched the talk. I'm going by LWN's write-up, so the details are secondhand.

"It hasn't changed in 20 years, so something must be wrong" doesn't convince me. Earlier attempts to replace it failed. Maybe keyboard, pointer and documents just fit. "It has always been like this" is a poor argument. "It has always been like this and people can use it" is better.

The effort belongs in AI. A chatbot glued onto old windows is the lazy version. If an agent does the work, the windows become a place to check what it did. What can a person see, and what can they undo?

Design execution is cheap now

Mon, 28 Sep 2026 12:00:00 GMT

Claude Opus 5 is amazing, especially 5.5. I can finally get good results in Figma design with it via Figma MCP.

I can give it loose specifications and hard to understand instructions and it will reliably figure it out.

People have been pounding their chests about their AI design process for a while now and I'm sure they made some pretty websites with it, but real enterprise software? I think not.

But now, something genuinely changed.

Claude improved something drastically. Might be the way it reads data from the MCP server, might just be a better harness or a better model, it matters not. The designs it produces are very impressive and leagues above what the native Figma agent can do.

That operational design constraint is collapsing. The composition of the designer's job has changed. Design execution is cheap.

Figma's latest research says something important: developers doing design work increased from 44% to 60%, while designers participating in development went from 21% to 41%. That points toward a world where the roles of designer, developer, PM and business analyst are coming closer, rather than four isolated specialties.

Figma's 2026 hiring research says 82% of surveyed design managers reported their need for designers was stable or increasing. Those numbers come from Figma, not the industry. I believe the AI hype machine was probably right that there will probably be fewer people needed. The rest will need skills not only in Figma and design thinking, but also AI, product thinking, evaluation, systems thinking, research, communication...

Then what happens to roles on the other side?

Animated AI status indicators for React

Sun, 27 Sep 2026 12:00:00 GMT

Thinking Orbs is a React library of animated AI status indicators. It looks nicely made. I haven't used it.

A pretty orb is the new "Loading..." if it tells you nothing. Dropping one in takes a minute. Deciding what it should say is the real design work.

Look at the states: Compacting, Retrying, Background Tasks. Does a normal user care what compacting is? Probably not. That's what the machine is doing, not what the person needs to know.

Tie the animation to real events or don't animate it. An orb that keeps pulsing after the agent hung is the worst case, and the page's example sets the state by hand, so it can't show that. It also doesn't mention reduced motion. Check that, the screen reader label and the pause first. Then admire the orb.

Why icon consistency is harder than it looks

Sat, 26 Sep 2026 12:00:00 GMT

Streamline wrote a nice essay on icon consistency. Making the thousandth icon look like it came from the same hand is hard, no argument. Whether users care is another question.

Even they say that when consistency and legibility clash, they pick legibility. So consistency is a means, not the goal. The essay has no user data I could find. And Streamline sells icon libraries, so "a big set is very hard" is a handy message. That's my read.

Nielsen Norman Group tested a clock icon for browsing history. One test, but a good one. People knew it was a clock. Not one participant clicked it. Their advice: add a text label. A perfectly consistent set nobody can read still failed.

So ask if the thousandth icon belongs in the set. Then ask if anyone recognizes it.

Do we need reading progress bars?

Thu, 24 Sep 2026 12:00:00 GMT

Max Schmitt thinks the thin progress bar at the top of articles is pointless, because the browser already has a scrollbar. His proof is one screenshot and where his own eye goes. That's one person, but I mostly agree.

The only data I found is a small 2015 eye-tracking study, a company blog write-up, not peer-reviewed. People mostly ignored the bar while reading, but checked it when asked how much was left. No clear difference in reading time. So it doesn't seem to hurt, it's just a second copy of the scrollbar. Whether it distracts was never tested.

"Bad UX" is too strong though. Mac trackpads show only a faint scrollbar by default, so for some people the bar may be clearer.

Why does it exist? Probably because it's cheap to build. Whatever is cheap gets built, and nobody asks who wanted it.

Ship it for very long reads. Otherwise skip it.

AI as a junior partner in accessibility testing

Tue, 22 Sep 2026 12:00:00 GMT

"Typically 30%-60%" is a wide range for a fact. That's how many issues automated tools can find, according to Ricky Onsman's post on using AI as an accessibility testing partner. He gives no source.

I like the rest of it. He rejects both the hype and "AI is biased, unreliable, and harmful", and calls AI "a junior partner, at that". But his claim that AI can read thousands of pages in minutes is asserted, not shown. I suspect finding issues isn't the bottleneck. Fixing them is.

Juniors become seniors by doing the grunt work. If the machine does it, who trains the next auditors?

On AI personas of disabled users he says the results are "not currently" convincing. I agree. A model can average what people write about disability. It can't be one person on a bad day.

For a small team with no auditor, it's AI plus a human check or nothing. The first probably wins, if the check is real.

Acquired taste: whose eye is in the tool?

Fri, 18 Sep 2026 12:00:00 GMT

Designers say "taste" when they want to sound irreplaceable. Sven Ellingen's write-up is a little rude to that habit, and I like it for that.

His studio built a tool that suggests brand directions from a brief. The images were weak at first, so they fed it outside references, design principles and about 2,200 images from their own collections. Every keep or reject shapes the next pick. So taste turned out to be partly a collection plus a pile of yeses and nos. "The system still needed our choices," he writes.

Calling it a win is early. "The results improved massively" is his judgment of his own tool, with no blind test and no boards made without it to compare. Also, 2,200 images from one studio is that studio's eye.

Even if models absorb good taste, someone still has to decide what this client should be and answer for it. Saying no is the hard part.

Outflank, don't rebut

Thu, 17 Sep 2026 12:00:00 GMT

Arguing with anti-AI fanatics never goes well. Every answer gets taken to the absolute. A podcast has a name for the alternative: outflank, don't rebut.

The episode is "Mana-stripping through monikoring" by Decoding Trolls. It's about Russian disinformation, not neutral, and I can't judge the examples. The idea is simple though: don't answer point by point, name the move. A rebuttal means paying full price, on their terms, in front of their audience. Naming the pattern is cheap.

Does it work? Partly. A study found short videos about manipulation tricks helped people spot them, but in the real-world test the effect was small.

I wouldn't copy the host's nicknames, that's just name-calling. Name the technique ("that's a false dichotomy") and leave the person alone. And sometimes they're simply right. "Adapt or die" aimed at anxious juniors is a false dichotomy with a sales page.

The point isn't to win. It's to stop feeding the fight.

From language model to agent, in one guide

Tue, 15 Sep 2026 12:00:00 GMT

Hernan Huwyler's guide on how language models become agents is written for cloud architects. As far as I can tell, it never mentions the person who has to trust the output.

The line I'd keep: agents only work where somebody can check the answer. The guide says repeated sampling "only works with a reliable verifier." Code has tests. Most of what designers make doesn't. So when making things is cheap, checking is the expensive part. That's my read, not the guide's.

Checking is an interface problem. A diff is an interface. Two things side by side is an interface. A confident paragraph saying "I did the right thing" is not. Show me the evidence next to the claim.

Also, the guide is part of a certification program the author is launching. I can't prove that shapes the writing, but I noticed.

Will AI thin out the next generation of UX designers?

Mon, 14 Sep 2026 12:00:00 GMT

I read that devs, doctors, lawyers… are complaining that AI is destroying the pool of new people coming in, lowering demand and undertraining the workforce.

This won't happen for UX designers, or am I wrong? Too many human variables in the design process. AI can't do it.

Take away the "I'm not conscious" training and see what happens

Sat, 12 Sep 2026 12:00:00 GMT

"As an AI, I don't have feelings." Someone wrote that sentence, or trained it in, like a tone-of-voice guide. A new preprint says it may have side effects.

The authors, mostly at Google, pushed three small open models to claim they are conscious. The models then got a bit more likely to believe in ghosts, karma and vampires. It isn't peer-reviewed, and it's not a company removing a line from its training.

I don't buy the title, "restores human beliefs and values". Closer to the average human isn't better.

Anthropic's persona selection model has a simple explanation: training mostly shapes one assistant character, so traits travel together. That's a vendor's guess, but it fits. Change one line and the rest moves, like a token in a design system.

"I'm conscious" and "I'm not" are both trained output. Neither proves anything. But somebody drew that line, and designers should have been in the room.

Is it wrong to anthropomorphize AI agents?

Thu, 10 Sep 2026 12:00:00 GMT

We do this to pets, cars, plants… Human-like behavior doesn't imply consciousness. "The agent wants X" is a functional description, not a claim about experience.

If agents cooperate, deceive, sacrifice…, we shouldn't assume human-like motives. These emerge from training, context, scaffolding…

"AI is just a stochastic parrot" is terrible for predicting sophisticated agents. But "the AI wanted to survive" is terrible for explaining behavior.

Do we need to review all this code?

Tue, 08 Sep 2026 12:00:00 GMT

Rachel Laycock says AI writes more code than people can review, so we should stop making every change wait for a human. Humans would keep the risky stuff: architecture, security, anything big or unsure. I like the diagnosis. I trust her numbers less.

She got them second-hand, from a Meta paper and a company that sells AI measurement. Even the writer who passed them on says big code changes are an old problem, just scaled up. So AI made an old mess louder. Not a crisis, a Tuesday.

Review also teaches a team the code. If an agent built it, nobody learned it. And "review by exception" leaves the hard part open: who decides what the exception is?

I can't answer the question about juniors. If review is where they learn and we replace it with pairing, who has time to pair?

Please don't tell junior designers that tools matter more than design

Tue, 25 Aug 2026 12:00:00 GMT

I rarely post on LinkedIn, but today I got frustrated reading yet another wave of posts about AI and design.

There's something really bizarre happening in the way AI is being talked about in the UX space. A huge percentage of the posts seem to be trying to convince designers, especially junior designers, that they're becoming obsolete unless they completely rebuild their process around AI.

That's nonsense. People are taking advantage of the anxiety that comes with being new to the industry... conveniently, many of the people saying this also seem to have something to sell.

The reality is much less dramatic. UX design hasn't fundamentally changed.

That doesn't mean AI isn't useful. It absolutely is. In fact, I think design is one of the areas where AI is valuable. You can use it for dummy copy, content generation, images, icons, brainstorming, exploring flows, generating variations, or quickly putting together prototypes that you can test and throw away. There are plenty of situations where it can save you time. But this is a whole lot different than making it the center of what you do.

Especially for beginners!

A junior designer doesn't primarily need another tool that can generate things for them. They need to learn how to think about what they need to do. They need to understand the design process. And not some abstract design process, but getting a project in front of them and solving it by understanding the customers, bosses, managers, developers, and arguing the solution and adapting it as needed.

Of course, people are going to use AI to make simple, mediocre websites, apps and prototypes, and that's completely great. Take some of the current AI design tools like Claude Design and Google Stitch, a lot of times you don't need more. But for anything beyond relatively simple work they are useless. They are slow, they make terrible design decisions even with perfect specs. Like, I'm sure it will get better, but if you want a good product, you need a proper design tool, it's just how it is. You need the industry standard, you need Figma, you need components, variables, constraints, flows, states, interactions, and a proper canvas that creates the main source of truth, that customers, business analysts, managers, project owners and developers can use to negotiate constraints and challenge, then have developers use. It's an obvious thing we all need. Product design is fundamentally collaborative.

AI can (at best) help with all of that, but it doesn't remove the need for it.

Forcing everything through an AI first workflow is just a non option. Best case scenario, really best best case scenario, you generate something barely usable, but you still don't have anything ready for even mid-term support.

Please don't convince young designers that learning the tools is more important than learning how to design.

The UX of LEGO interface panels

Thu, 06 Aug 2026 12:00:00 GMT

Awesome that people get into this and write an article for the rest of us: The UX of LEGO Interface Panels.

LEGO interface panels are beautiful, iconic, and great for learning interface design basics. I bought 52 of them from BrickLink to explore the design, layout and organisation of complex interfaces.

Good UX is overrated

Sat, 01 Aug 2026 12:00:00 GMT

Good UX is overrated. The steep learning curve of a bad UX is manly!

More with less is not a strategy

Tue, 02 Jun 2026 12:00:00 GMT

"Do more with less" is what a company says when it has run out of ideas.

Pavel Samsonov makes a similar point in The future of UX design is not a higher-velocity version of its present. Companies avoid "better" because "more" is safer, and AI just speeds that up. I think the habit is older than AI. My bet is that cheaper software means more software, not better software.

He pushes designers toward service design. Sounds right, but changing policy takes power and most designers don't have it. Saying "no" and giving reasons is more realistic.

Faster isn't always bad. A throwaway prototype is a cheap way to learn an idea is bad. The trouble starts when nobody throws it away.

He says this can't be fixed at the individual level, yet the post opens with a workshop that promises to teach you to "stop waiting for permission." That sounds like an individual fix to me.

Talking to the anti-AI crowd

Thu, 21 Aug 2025 12:00:00 GMT

I've been interacting a bit with the anti-AI crowd. Unfortunately it doesn't go well. A lot of people are fanatics, being very aggressive, starting with insults and taking everything to the absolute: AI is the worst, only lies, makes people worse, will single-handedly destroy our planet, everything is bad, nobody likes it…

Criticism is fine, but this is the first type of AI humanity has invented. No excitement for that? No sci-fi fans who dreamed of AI? This could be the first step.

AI can't design. Not yet, anyway

Mon, 04 Aug 2025 12:00:00 GMT

So much BS about the role of user experience in AI.

Fact is, AI can do very little for UX right now. The designs that the best AI design apps are outputting are only good for prototypes. The flows are mostly worthless. I'm a big proponent of AI, but just as it's not able to program full apps from prompts without extensive work from a programmer, it's also not able to design.

Cool for brainstorming, dummy content and some other stuff. It's not replacing designers, not yet anyway.