Meaghan Choi - Designing Claude Code (and what's coming next)

2026-07-08 - 54 min - source - Read full transcript

Ridd (host)Meaghan Choi

Key insights

Claude Code started as an unsanctioned side project, not a chartered design initiative.

Meaghan saw an internal demo of a slow, clunky CLI pipeline that let Claude edit code directly, got excited, and DM'd her manager saying she wanted to design it even though the team building it didn't think a CLI needed a designer. Her manager said go ahead, and that became her full-time role.

claude-code-origin-story

CLI design inverts the GUI instinct of progressive disclosure toward maximal information density.

Because developers expect to see exactly what a tool is doing and because the models weren't fully trustworthy in the early days, Meaghan designed Claude Code to disclose far more information upfront than a graphical interface would, since there is no equivalent of hiding detail behind a click in a terminal.

claude-code-origin-story

The most novel technical shift, not just the interface, was Claude reading and writing an entire file base directly.

Before this, developers copy-pasted code into a chat and pasted the output back manually. Seeing Claude make an edit in place, with full context of the existing codebase, was the moment people realized the tool was fundamentally different, and the team optimized hard for engineering that first moment of surprise.

claude-code-origin-story

Artifacts emerged bottom-up from people already texting each other ad hoc Claude-generated HTML files.

Before Artifacts existed as a product or even had a name, designers, engineers, and data scientists inside Anthropic were independently asking Claude to output HTML because it was fast to generate and iterate on, then manually screenshotting or sending the raw files to each other for review. The product formalized that workflow rather than inventing a new one.

artifacts-as-prototyping-workflow

Artifacts collapsed Meaghan's entire design-review loop into a single iterative, shareable prototype.

Her standard flow: get a feature request, have Claude generate an artifact instead of doing manual design work, iterate on it live with stakeholders via DM feedback, then once it reaches mid-fidelity have Claude rebuild it against the real component library so it graduates from an HTML artifact into an actual pull request.

artifacts-as-prototyping-workflow

The right amount of interface craft depends on whether you're testing a mental model or shipping a finished product.

Meaghan argues that early-stage exploration should be fast and rough because polishing the wrong shape of a product wastes effort; craft investment should wait until the underlying mental model is validated, and designers need explicit judgment about which phase they're in.

ai-native-design-practice

Claude Tag reframes AI collaboration from many disposable single-player sessions into one persistent, shared, organizational Claude.

Instead of spinning up separate Claude instances per person or task, Claude Tag is a single entity the whole organization interacts with simultaneously, holding shared memory and context across every channel, codebase, and decision, with its own credentialed access to tools like GitHub, Google, and Asana rather than relying on individual users' MCPs.

claude-tag-organizational-ai

More than half of Meaghan's shipped code now originates from Slack conversations with Claude Tag, not an IDE.

She describes sending a one-line UI fix request in Slack DM and Claude Tag finding the relevant repository code, pushing a branch, opening a draft PR with a before/after screenshot, waiting for CI to pass, requesting engineer review, merging, and reporting back, with no further action from her.

claude-tag-organizational-ai

The core emerging designer skill is discernment about when to hand off execution versus when to do the deep thinking yourself.

Meaghan argues Claude is not good at telling you your idea is the wrong shape; it will confidently execute whatever it's asked. The designer's job is recognizing when a request needs shape-level discussion rather than execution, and she says she sees many designers get this backwards.

ai-native-design-practice

Design quality should become a shared organizational responsibility, not something design alone gatekeeps.

Rather than blocking work to enforce polish, Meaghan 'lowers the floor and raises the ceiling' by having Claude automatically stack a nicer alternative PR on top of an engineer's PR as an iteration, and by teaching Claude the standing questions she asks engineers so it can raise quality bars without stopping progress.

future-of-the-designer-role

Not every interface should become dynamic; deciding what stays fixed versus adaptive is a core design judgment call.

Meaghan pushes back on the assumption that customization is always good: surfaces like login screens and billing settings should stay deliberately stable and reliable, while other surfaces benefit from being non-deterministic. Designers now own deciding which is which and how to visually signal the difference to users.

future-of-the-designer-role

Because building is now nearly frictionless, discernment about what's worth building is becoming the scarcer skill than execution itself.

Meaghan warns that ease of building tempts teams into product bloat: not every possible feature or function is worth shipping just because it's cheap to generate, and judgment about what to build (and what to kill) is one of the biggest unmapped design challenges ahead.

future-of-the-designer-role

Companies

Anthropic - Meaghan's employer since late 2024; she is the founding designer on Claude Code and now design lead, also involved in Claude Tag and Artifacts.
OpenAI - Referenced as the dominant, well-known player when Meaghan joined Anthropic in late 2024, when almost nobody outside tech had heard of Anthropic.
Figma - Meaghan still uses Figma for fast layout iteration (e.g. auto layout) when reordering many settings is faster there than prompting Claude.
Jitter - Podcast sponsor; motion design tool the host recommends for animating design work.
Dessn - Podcast sponsor; design/prototyping tool whose new 'Surfaces' feature lets designers prototype directly on top of an existing production interface in the cloud.
Slack - Meaghan's primary interface for working with Claude Tag; she runs most of her thinking and execution work through a dedicated Claude channel in Slack rather than an IDE.
GitHub - Claude Tag is given its own organizational GitHub credentials so it can open branches, push PRs, and merge on Meaghan's behalf.
Asana - Cited as an example of a tool Claude Tag can be given its own organizational access to, alongside GitHub and Google.

Techniques and frameworks

Claude Tag - An organization-wide, persistent Claude that all employees interact with as a shared entity (not per-user sessions), with its own credentials/tool access and a shared memory layer across every channel and codebase.
Artifacts - Anthropic's shareable, live, version-controlled output format that grew organically out of engineers and designers texting each other ad hoc HTML files generated by Claude.
Execution vs. shaping discernment - Meaghan's core designer skill framework: knowing when to offload a task to Claude for execution versus when to slow down and do the 'thinky' shape-defining work yourself.
Stacking PRs - Meaghan's tactic for raising quality without gatekeeping: she has Claude automatically stack a polished alternative PR on top of an engineer's PR as an iteration, rather than blocking their work.
Fixed vs. adaptive UI - Meaghan's framework for deciding which surfaces (login, billing) should stay deliberately stable versus which should be dynamic/non-deterministic.

Summary

Meaghan Choi, design lead for Claude Code at Anthropic, walks through how she went from spotting an internal demo of a clunky, 45-minute CLI-editing pipeline in late 2024 to becoming the founding designer on what became Claude Code, a product nobody thought needed a designer at all. She describes CLI design as an inversion of typical GUI instincts: instead of progressive disclosure, developers expect maximal information density, and the biggest design win was making the tool feel like a conversational back-and-forth rather than a mechanical print-and-respond loop. The real breakthrough, she says, wasn't visual but mechanical: Claude reading and editing an entire file base directly, eliminating the copy-paste-back-and-forth workflow everyone had been suffering through.

A large part of the conversation covers Artifacts, which Meaghan describes as emerging bottom-up rather than being designed top-down. People across Anthropic were already asking Claude to generate HTML files and texting them to each other for review before there was any product wrapper or even a name for it. Her own workflow now runs almost entirely through artifacts: she has Claude generate a prototype from a feature request, iterates live with stakeholders via DM, and once the shape is validated, has Claude rebuild it against real components so it graduates into an actual pull request. She frames the underlying design judgment as knowing when speed and rough iteration matter more than craft (early-stage mental-model testing) versus when polish is worth the investment.

The most striking segment covers Claude Tag, an organization-wide, persistent Claude that the whole company interacts with as a single shared entity rather than as disposable per-session instances. It holds shared memory and context across every Slack channel, codebase, and decision, and can be given its own credentialed access to tools like GitHub, Google, and Asana. Meaghan says more than half of her shipped code now originates from Slack DMs to Claude Tag: she sends a one-line fix request, and it finds the code, opens a draft PR with a before/after screenshot (a preference it learned once and now applies automatically), waits for CI, requests engineer review, merges, and reports back, with no further input from her.

Pushed on whether this amounts to designers "outsourcing their thinking," Meaghan argues the opposite: automating execution frees her to spend time on genuinely hard, ambiguous problems, like how to explain a shared organizational AI entity to users without breaching security boundaries. She distinguishes tactical execution work from "existential" brainstorming sessions with Claude, where she treats it like a product partner in open-ended exploration rather than asking for a specific output.

The episode closes on where design authority goes in an AI-native world. Meaghan argues design quality should become a shared organizational responsibility rather than something design alone gatekeeps, describing how she has Claude automatically stack polished PRs on top of engineers' work as suggestions rather than blockers. She also pushes back on the assumption that everything should become dynamic and customizable: some surfaces, like login screens and billing settings, should stay deliberately fixed and reliable, and deciding which is which is now a core design judgment call. Her closing framework for designers who want to thrive: stay fluid and comfortable with ambiguity, develop discernment (a proxy for taste) about what's worth building at all, and take full ownership of shipping rather than treating execution as someone else's job.

Notable Quotes

"Knowing when to use these tools for execution and offload and knowing when you need to go deep and do the thinky work is a skill set, and I think that's one that I see people doing wrong." - Meaghan Choi

"The moment you notice the tool is the moment we failed. It should be totally out of your way." - Meaghan Choi (quoting a Claude Code team designer)

"I would say more than half of my code is done like this now." - Meaghan Choi

"When you can build everything, should everything be built? Just because it can build and just because it can be a function doesn't mean it needs to be." - Meaghan Choi

"We're 1% in this journey of what the shape of AI should be." - Meaghan Choi