Nobody Writes Skills by Hand Anymore: Meaghan Choi, Designer at Anthropic

Anthropic Claude Code design lead Meaghan Choi said at the Dive Club Live design meetup in New York that she no longer writes Claude Code skills by hand, telling the audience, "Nobody writes skills by hand anymore. If someone tells you they do, they're lying. Everyone just prompts for them." In a 12-minute talk posted June 4, 2026, Choi demonstrated her /prototype skill, which by default generates five versions of a feature into an HTML file, and said she now asks Claude to pick which prototype to build and explain why rather than choosing herself. Choi, who joined Anthropic in fall 2024 and has led design for both Claude Code and Cowork since January 2026, also noted a ready-made /design command that arrived in Claude Code in August 2026 on the Pro, Max, Team and Enterprise plans.

Let me start with a joke. At Dive Club Live, a design meetup in New York, Meaghan Choi was on stage showing her /prototype skill. It builds several versions of a future feature. A skill in Claude Code is a saved instruction that Claude pulls out when the task calls for it. In the middle of the demo, Meaghan says: I didn’t write it. Nobody writes skills by hand anymore. If someone tells you they do, they’re lying. Everyone just prompts for them. Meaghan Choi, Dive Club Live talk, video from June 4, 2026 It’s a joke, and it’s about one skill. But the habit behind it is real: Meaghan orders her working tools from Claude. Here I’m with her. My own skills, for building websites, video or images, I also got by asking Claude. Not a single one was written by hand from scratch. At Anthropic, Meaghan Choi https://www.linkedin.com/in/meaghanchoi/ leads design for Claude Code and Cowork, another of the company’s products. In a 12-minute talk the video came out on June 4, 2026 she showed how Claude takes a feature from rough options to a finished code change. And the decision about what ships, she keeps for herself. If you’re curious about Claude Code for designers, this is a rare case: the person showing it is the one who designs the tool. Below I go through her moves one by one: first what Meaghan says, then, on a separate line, my take. Meaghan joined Anthropic in fall 2024 as design lead for Claude Code. Since January 2026 she has been leading design for both products Claude Code and Cowork . She was designing command-line tools long before AI and loves the terminal. But she’s quick to add: You don’t have to work this way: everything I show today, you can do in the desktop app too Meaghan Choi, Dive Club Live talk, 2026 My take: for a designer, this is the most important line of the intro. A lot of people are scared of the black window full of commands, and they shouldn’t be. I work in Claude Code every day, and only in the desktop app, in the Code tab. Meaghan’s tip: if you have several Claudes running at once, give each task its own worktree. It’s an isolated copy of the project, so tasks don’t get in each other’s way. She runs the whole demo on an online whiteboard for diagrams, something like a simplified Miro. If you want a board like that to try the same exercise on, there’s one on mozgi.io https://mozgi.io/board/?lang=en , along with a big catalog of prompts and agent skills. She gives Claude the task in one line: “add autocomplete”, meaning a hint for how to finish a word while you’re typing text on the board. “That’s it. No design specs,” she says. No mockup, no brief: Claude comes up with what it could look like on its own. Then she calls /prototype. By default the skill makes five versions of the feature, packs them into an HTML file, shows a preview and iterates a few times. And here’s what she now asks for differently: I used to say: make prototypes, and I’ll pick. Now I say: make prototypes, then pick yourself what you’ll build, and explain why Why she does it is simple: she’s curious what Claude will come up with. In the demo she lets it search the web. In a real project she’d also have it read Slack, Google Docs, team discussions and data from BigQuery. Next comes a scene that some retellings of this talk get wrong. The options are ready. Meaghan says she usually looks at the previews herself, and sometimes Claude asks her to pick. Then she asks the room: anyone have a preference? Someone in the audience says: “Number two.” She takes number two. Which option Claude would have picked is never said on camera. If you don’t feel like building your own skill, there’s a ready-made /design command. It arrived in Claude Code in August 2026, on the Pro, Max, Team and Enterprise plans. Claude lays out a canvas with editable mockups: you pick one, refine it and ask Claude to build it. And if you want something close to Meaghan’s /prototype right now, there’s a community skill called prototype https://mozgi.io/skills/emilkowalski-prototype/?lang=en in the mozgi.io catalog: it builds several genuinely different UI variants behind a visual picker. From there, according to Meaghan, Claude implements the chosen option, tests it, tweaks the styles and puts up a PR with a screenshot. A PR pull request is a finished code change that gets reviewed before it goes into the product. Meaghan no longer watches Claude work. She looks at the result: Usually I look at the PR Claude opened, with a recording of how the feature works Claude checks itself with Claude in Chrome. It’s a browser extension: Claude opens the page on its own and looks at what came out. When the change touches the interface, Meaghan says, this is the best way to let Claude check its own work. The docs even have a scenario made for designers: build an interface from a Figma mockup and compare it against the mockup in the browser. Claude doesn’t ask her permission for every step: I’m always in auto mode, and everyone at Anthropic is always in auto mode Auto mode means a second model reviews Claude’s actions instead of you. Since August 14, 2026 it’s on by default on Pro, Max and Team, and since late September, on all plans. As the docs honestly say, it doesn’t guarantee safety. The demo ends with Claude recording a GIF and getting ready to open the PR, but the twelve minutes run out first. My take: I’m taking the screenshot or recording in every change. For a designer it’s the most useful part of the whole chain: you see what the feature really looks like without digging into the code. But browser checks have a ceiling. Claude will notice a broken layout. It won’t tell you whether it looks good. Meaghan explains why herself. Claude, unfortunately, like most language models, isn’t good at design yet She adds right away that it won’t always be this way. But for now, she says, deciding what goes into the product has to be her call. And a bit later: anyone can ship now, but that doesn’t mean everything should ship. My take: I agree completely, and for me this is the key moment of the talk. Claude Design was already out by then, but there’s no contradiction. It makes mockups. Which one is any good, a human decides, and that’s exactly Meaghan’s point. If a client ever tells you “the AI will just draw it all,” you can quote the person who leads design for Claude Code. Code, Meaghan says, is already automated. On top of that, Claude does a lot of her work that isn’t about code at all. The most interesting example is a routine: work that Claude kicks off on its own, on a schedule. Her routine scans every repository and finds interface changes. For each one it looks for traces of a designer: in Slack, in Google Meet transcripts, in Google Docs. If there are none, Claude breaks down the change, drafts its own counter-proposal and messages the engineer: go work with a designer on this. How did it end? I had to turn off the DMs: Claude, unfortunately, turned out to be really bad at design Now Meaghan reads the routine’s report herself. When a new model comes out, she just reruns the routine, since the instructions are already written. The result, she says, usually gets better. My take: I’m taking the idea, not the messages to engineers. If even Meaghan switched those off, it’s definitely too early for me. You can build a similar routine without Anthropic’s internal tools. Claude Code has routines and the /schedule command for this: still in preview, on Pro, Max, Team and Enterprise, running at most once an hour. The easiest place to start is a check of a single mockup rather than every repository. On July 8, 2026 the Dive Club podcast came out: 54 minutes of Meaghan talking with host Michael Riddering. It’s a different conversation, and she fills in a lot of gaps. Almost everything goes through Slack. Meaghan writes to Claude in Slack: “remove the shadow.” It finds the style, deletes it and opens a draft change with before-and-after screenshots. Those habits of hers are remembered by Claude Tag. That’s Claude you invite into a Slack work chat as a member of the team. I’d say more than half of my code gets done this way now Meaghan Choi, Dive Club podcast, 2026 A caveat: Claude Tag came out on June 23, 2026 and is only available on Team and Enterprise. And it was right. Her mockup had a gap, and Claude filled it its own way. It offered to redo it strictly to the mockup, even though it thought its own version was better. This is only about a hole in the mockup: Claude wasn’t deciding what the product should be. Claude isn’t great yet at telling you your idea is wrong: it’ll just do what you asked A check built from questions. The questions Meaghan asks about engineers’ changes, she’s gradually turning into an automated check: who is this for, what are we communicating, does it match the design system, has the copy been reviewed. Her advice: a check like that should teach the team, not slow it down. My take: in two months the shift is noticeable. From the podcast I’m taking two things, and both work without Slack: before-and-after screenshots in every change, and her four questions as the base for my own review skill. And since Claude will just do what it’s asked, checking the task itself is on me. Remember how Meaghan set the task? One line: “add autocomplete”, no mockup, no brief. Claude came up with the options itself. But in a real project, she said, she’d also let it read the team’s conversations and docs, so it understands what’s already been decided. In August, Anthropic published a guide about exactly that second case: what to do when the task is big and one line isn’t enough. This isn’t Meaghan anymore. The guide is called “The AI-native SDLC playbook” August 21, 2026 , and it was put together by the Anthropic team that helps customers roll out Claude. The core idea is a file called INTENT.md, literally “what I want.” Before handing a task to Claude, you briefly write down five things: Then Claude asks you questions like a meticulous analyst, and you finish the file together. Where it misunderstood, you correct it. From that file Claude writes a detailed task description, then a plan, and only then the code. The easiest way to think about INTENT.md is as a brief. Except you’re writing it for Claude, not a contractor, and Claude helps you finish it. One important detail: this is advice from the guide, not a button in Claude Code. Claude won’t go looking for that file on its own; you have to hand it over directly: “read INTENT.md and work from it.” My take: I’m taking INTENT.md, but only for big tasks. When several people are discussing a feature, a brief like this saves a ton of rework. For something small like Meaghan’s autocomplete, it’s overkill: one line is enough, and the guide itself says small changes should just be done right away. In short: in Meaghan Choi’s workflow, Claude makes the options, checks itself in the browser and brings back a change with a recording. What goes into the product, a human decides. I’m taking five options where Claude picks first and explains why. I’m taking a screenshot or recording in every change, and the browser check, keeping in mind that it doesn’t check taste. For now I’m not taking messages to engineers sent on Claude’s behalf, Claude Tag, which isn’t on a personal plan, or INTENT.md for small tasks. If you want to try Claude Code as a designer on just one thing, start with how she sets the task. Ask Claude to make several options, pick one first and explain the choice. Then compare it with what you’d have picked yourself. The first time I tried it, my agents got the overall direction right and missed on a few things. That mismatch is exactly what showed me where Claude thinks like me and where it doesn’t. And that’s fixable. If you want to try it today, you’ll find a prototype-style skill, plus thousands of prompts and other agent skills, on mozgi.io https://mozgi.io/?lang=en . And my Telegram channel about AI is t.me/mozgi io https://t.me/mozgi io in Russian . Nobody Writes Skills by Hand Anymore: Meaghan Choi, Designer at Anthropic https://pub.towardsai.net/nobody-writes-skills-by-hand-anymore-meaghan-choi-designer-at-anthropic-cc4c8c2cb26b was originally published in Towards AI https://pub.towardsai.net on Medium, where people are continuing the conversation by highlighting and responding to this story.