How Anthropic, Every, & Ramp design with AI

2026-06-02 - 40 min - source - Read full transcript

Ridd (host)Meaghan ChoiDan ShipperBradley Zipper

Key insights

The first milestone in any org's AI transformation is giving designers direct access to the real production codebase, not a segregated sandbox or forked repo.

Meaghan argues that maintaining a separate playground repo just means maintaining two codebases that drift out of sync, and that a sandbox can't replicate the real data endpoints, internal tooling, and organizational context a production codebase gives you. Proximity to what end users actually see is proximity to real product influence.

org-ai-transformation

The second, more uncomfortable milestone is designers letting go of direct control: features now ship in V1-V3 form without a designer touching every pixel.

Meaghan frames this as symmetric to the ask engineers already made of designers (let us into your code) - now designers have to extend the same trust to engineers shipping design-adjacent work, provided the org builds automated checks and design-system guardrails.

org-ai-transformation

Models now get most design and code tasks to a 'seven out of ten' baseline automatically, which frees designer time for craft rather than eliminating the need for craft.

Bradley describes this as a shift from spending most of a project's time getting to a passable result to spending most of the time adding intention on top of an already-decent starting point - but warns the freed time creates a slippery slope where a designer can end up endlessly polishing a single feature instead of moving to the next highest-leverage task.

designer-ai-workflow

The strongest signal of whether an org is actually transforming with AI is whether the CEO and executive team are personally using the tools daily, not whether they've stood up an AI working group.

Dan says three years of lip-service AI initiatives largely failed, while the orgs that succeeded had leadership literally opening Claude Code and building things themselves, because managing a team that lives in these tools all day is not something you can develop intuition for secondhand.

org-ai-transformation

The Opus 4.5 / GPT-5.3 release window around November-December 2025 marked a step change from reviewing every line of AI-written code to shipping PRs without looking at the codebase at all.

Dan describes going from a year of cautious code review to suddenly shipping PRs across Every's products without knowing the underlying codebase, calling the models 'much smarter' and 'much more independent,' though he notes this occasionally produces genuinely bad PRs.

designer-ai-workflow

Deciding where to spend design time now hinges on a forward-looking question: is this worth polishing if the product might not exist in its current form six months from now?

Meaghan says Claude Code's team explicitly weighs whether time is better spent on deep, exploratory future-shaping work (which models can't yet do) versus incremental ticket polish, since the team is operating at a research frontier where the final shape of the product is still unknown.

designer-ai-workflow

Polish and care are now a shared responsibility between designers and engineers rather than defaulting onto designers just because they can now ship code.

Meaghan got direct feedback from her own engineering team that she was spending too much of her time manually reviewing and polishing PRs herself; the fix wasn't working harder but recognizing that engineers, now empowered the same way designers are, should also own polish and craft.

org-ai-transformation

Removing administrative overhead, not just task execution, is one of the clearest markers of AI fluency inside an org.

Bradley describes never once manually scheduling a research call in four months at Ramp - an internal system auto-books and pre-briefs four customer calls a week - freeing him to spend his time synthesizing what customers say across the whole product instead of managing logistics.

ai-fluency

AI-fluent operators split into two distinct skills: building systems to harness the surge of work non-specialists can now produce, and using the tools to build things that were never possible before.

Dan says the first skill requires designing systems and processes to absorb a 'huge glut' of design-adjacent work now coming from non-designers rather than dismissing it; the second favors curious, multi-disciplinary people (he cites a colleague who was a professional composer and baker before becoming a technical agent-builder) who treat the tools as a playground for things that couldn't be built before.

ai-fluency

Formal shadow-pairing, an hour a month watching a teammate work their actual task, was adopted specifically to counter the isolation of spending entire days talking to a model instead of a person.

Meaghan notes that workflow habits are hard to articulate out loud (she struggled to explain her own process for a live demo) but become obvious when observed directly; the Claude Code design team cycles this monthly, borrowing a practice long normal in engineering pair programming.

knowledge-sharing

Public Slack agents, not shared skill libraries, are currently the most effective mechanism for spreading AI technique across an org, because watching someone else prompt in public is itself a knowledge-transfer event.

Dan says an internal skill library suffers from staleness and personalization problems (skills work best customized to the individual), whereas encouraging public-channel use of Slack agents lets people see others' prompts and copy patterns organically; at Ramp, Bradley describes their bot 'Cody' evolving past Q&A into teaching other agents new patterns and publishing audio and video recaps of what the team is learning.

knowledge-sharing

As AI absorbs more of the UI-design layer, the designer's real work shifts a level deeper: deciding what parts of a product should be held fixed versus made flexible and personalizable.

Meaghan predicts brand and fundamental design systems will remain a human-taste domain (guided by prompting, but still expert-crafted), while personalization becomes the dominant new surface; the harder design decision becomes things like keeping a login screen fixed while letting a dashboard flex, which she calls 'fundamental UX' operating one layer down from where it used to.

future-of-design-value

Companies

Anthropic - Meaghan's employer; she is the design lead behind Claude Code and Cowork and runs onsite trainings teaching design orgs how to adopt Claude and AI workflows.
Every - Dan Shipper is CEO; the company builds an internal Codex-based email agent (Kora) and other AI-native tools, and Dan describes his own workflow living almost entirely inside Claude Code and Codex.
Ramp - Bradley Zipper's employer as a design engineer; Ramp runs an internal Slack agent named Cody and a research-scheduling system that auto-books customer calls for designers.
Devin - Cited by Dan as an early example of a cloud-sandbox coding agent, contrasted with Claude Code's approach of running directly on the user's own machine with full local access.
Victor - A third-party Slack agent product Every uses alongside their own internally built agents, mentioned as having recently raised a large funding round.
Destin (Desen) - Mid-episode sponsor read for a design tool, Surfaces, that lets designers prototype directly on top of an existing production interface pulled from prod.
Paper - Mid-episode sponsor read for a Chrome extension (Snapshot) that copies live website components into Paper's canvas as editable HTML/CSS layers for rapid Claude-assisted iteration.
Inflight - Sponsor read for the host's own feedback tool, newly in open beta, for sharing prototypes and collecting structured feedback at AI-workflow speed.

Techniques and frameworks

Production-codebase-first design access - Meaghan's first milestone for any org's AI transformation: give designers direct access to the real production codebase rather than a separate sandbox repo, since a forked playground repo just becomes a second codebase to maintain and lacks real data endpoints and tooling.
Seven-out-of-ten baseline reallocation - Bradley's framing that AI tools now get most tasks to a 'seven out of ten' baseline automatically, which is meant to free up time for craft and care on the remaining 30 percent rather than being treated as a finished result.
Shadow-pairing practice - The Claude Code design team schedules an hour a month where designers watch each other work on their actual current task, because workflow habits are hard to describe verbally but visible when observed directly, and because working with models all day is otherwise isolating.
Public-channel Slack agents - Both Every and Ramp encourage (without mandating) that people use their internal AI agents in public Slack channels rather than DMs, so other employees can observe real prompts and techniques and pick up new patterns by osmosis.
Two-bucket AI-fluency framework - Dan's mental model for what separates strong AI-fluent operators: building systems to harness the glut of work now producible by non-specialists, versus using the tools to build things that were never possible before.
Auto-scheduled customer research - Ramp's internal system pops four customer research calls onto Bradley's calendar each week pre-briefed on what he's already working on, eliminating the administrative overhead of manually scheduling and prepping research calls.

Summary

This episode is a live panel recorded in New York, bringing together Meaghan Choi (design lead for Claude Code and Cowork at Anthropic), Dan Shipper (CEO of Every), and Bradley Zipper (design engineer at Ramp) to compare how AI-native organizations are actually restructuring design work. Host Ridd frames the conversation around organizational transformation: what milestones signal that a company is genuinely changing how it builds, versus paying lip service to AI. Meaghan opens with two uncomfortable but foundational shifts: designers need direct access to the real production codebase (not a segregated sandbox, which just becomes a second codebase to maintain), and designers need to get comfortable with features shipping without their direct touch on every pixel, in the same way engineers had to get comfortable letting designers into code.

A recurring thread is how the "seven out of ten" baseline that AI tools now produce changes time allocation, not eliminates the need for craft. Bradley describes this freed time as double-edged: it lets people put more care into fewer things, but also creates a slippery slope toward endless, low-leverage polishing. Meaghan connects this to a frontier-research mindset at Anthropic, where the team explicitly asks whether something is worth polishing given the product's final shape is still unsettled, and describes getting direct feedback from her own engineers that she was over-investing her time in manual PR polish rather than higher-leverage work - reframing craft as a shared responsibility across design and engineering, not a designer's sole burden now that they can ship code.

Dan Shipper's contributions center on organizational signals and a personal account of the Opus 4.5 / GPT-5.3 release window (roughly November-December 2025), which he describes as the moment he went from cautiously reviewing every line of AI-generated code to shipping PRs across Every's products without knowing the underlying codebase at all. He argues the single strongest predictor of real organizational transformation is whether the CEO and executive team are personally in the tools daily, since managing a team that lives inside AI agents all day isn't something leadership can develop intuition for secondhand. He also offers a two-bucket framework for AI fluency: building systems to harness the surge of AI-produced work from non-specialists, and using the tools to build things that were never possible before - the second favoring curious, multi-disciplinary people over narrow specialists.

Knowledge transfer inside AI-native orgs comes up as a genuinely unsolved problem that all three guests are actively experimenting with. Meaghan describes a monthly shadow-pairing practice on the Claude Code design team, an hour spent literally watching a teammate work their actual task, adopted partly to counter the isolation of talking to a model all day and partly because workflow habits are hard to explain verbally but obvious when observed. Dan and Bradley both point to public-channel Slack agents as the most effective diffusion mechanism currently available: watching someone else prompt in a shared channel is itself a form of learning, and internal skill libraries tend to go stale or need per-person customization anyway. Bradley's account of Ramp's Slack bot "Cody" is a highlight - it evolved from simple Q&A into teaching other agents new patterns and publishing its own audio and video recaps of team learnings, becoming a genuine internal culture artifact rather than just a utility.

The panel closes on where design value goes next. Meaghan's personal (explicitly not company) view is that AI will be capable of most fundamental design work within the year, pushing human designers to focus on brand and systems-level taste (still expert-crafted, even if AI-assisted) and, increasingly, on deciding what parts of a product should stay fixed versus become personalizable and flexible - a decision she calls "fundamental UX" that now sits a layer beneath where UI design work used to live. The overall tone across the episode is one of genuine excitement paired with humility: multiple speakers stress that AI has only really solved two problems so far (search and coding), and that everyone in the room is still early enough in the shift that yesterday's workflow can be obsolete by next week's model release.

Notable Quotes

"You need to be more comfortable letting go of design. And that means that a lot of features can go out without you." - Meaghan Choi

"The main thing that I always look at is what is the CEO doing, and maybe more broadly what is the executive team doing... that's not outsourcable." - Dan Shipper

"We have really only solved two use cases... Search and coding. There's so much else out there that to assume we're anywhere close to the end right now is like we're just not." - Bradley Zipper

"I don't want your login screen to change every single time you log in, but you might want your dashboard to be flexible as well. The decision of what to keep fixed, what to be allowed to customize... that's like fundamental UX." - Meaghan Choi