How AI is reshaping the design role at Anthropic and beyond

Executive overview

The traditional design process — research, diverge, converge, mock, prototype — is dying. Engineering velocity driven by AI tools has outpaced it. Designers can no longer lead with beautiful mocks; they must shift toward enabling execution and pointing teams in a shared direction.

The design role is splitting into two modes: supporting fast-moving implementation, and setting near-term vision (now 3–6 months, not 5 years). Time spent mocking has dropped from 60–70% to 30–40%, with the remainder moving into pairing with engineers and direct implementation.

Designers who thrive now are those who let engineers cook, consult in real time, and help create coherence — not those who gatekeep process.

How the design process is changing

• Engineers shipping via AI agents has forced design to change, not any internal design-led reckoning
• The "trust the process" gospel — research, diverge, converge — is no longer viable at pace
• Design work is now stratified: (1) supporting implementation and execution, (2) setting near-term direction
• Vision work has compressed from 2–10 year decks to 3–6 month prototypes that point teams in a direction
• Mocking was 60–70% of the job a few years ago; it is now 30–40%
• The gap is filled by pairing with engineers and direct implementation (polishing, last-mile work in code)
• This shift is spreading beyond frontier AI companies — the "Don't Trust the Design Process" talk resonated broadly, though a portion of the industry is still resistant

Day-to-day work as a designer at Anthropic

• A significant portion of time is just staying current: model updates, internal prototypes, research team directions, Slack channels
• High information density means tracking what is coming is itself a useful design input
• Remaining time splits between: near-term vision work (3-month horizon), jamming with engineers on active builds, and direct implementation (polishing in code alongside engineers)
• Traditional elements — user research, prototyping, Figma mocking — still exist but occupy a smaller share
• Figma remains valuable for exploring 8–10 directions simultaneously and micro visual details; coding tools are too linear for exploratory divergence
• IDEs may now be most useful for designers and PMs rather than engineers, who have moved to agents

Maintaining quality while shipping fast

• Label early releases honestly (e.g., "research preview") — sets correct expectations
• The promise to users is: we will iterate visibly and respond to feedback
• Trust degrades when something ships early and then nothing improves; it holds when teams visibly ship fixes
• Building trust through speed — responsiveness to feedback is the quality signal, not polish at launch
• Claude Co-Work shipped externally in 10 days of final build, but that followed extensive internal prototyping of sub-interactions (to-do list formats, question widgets, onboarding, etc.)

Where human judgment still matters

• AI will get better at taste and design; holding on too tightly to that as a human advantage is a mistake
• The persistent human role: deciding what gets built and being accountable for that decision
• Many hard parts of building software are interpersonal — resolving disagreements about scope, prioritisation, tradeoffs
• AI can surface data to inform a decision, but someone still needs to be accountable for pressing go
• Parallel: radiologists using AI — the human is still needed to sign off, not because AI can't read, but because accountability requires a person

The future of design–engineering interface

• Chat and terminals are not the final UI paradigm, but they are not going away — they opened an "infinite flexibility" mode that baked-in UIs could not offer
• Claude's interactive widgets (weather, stocks, structured questions) show demand for clickable UIs alongside conversation
• Trend: more UIs will be generated by models on demand rather than hand-coded per use case
• Talking as an interface scales across all intelligence levels — this is why conversation has been so durable as a paradigm

The legibility framework for designers

• Framework by Evan Tana (SPC): map founders and ideas on a 2x2 of legible vs. illegible
• Highly legible founder + legible idea = already being built; not novel
• Most interesting opportunities: illegible ideas with energy around them that nobody can fully articulate yet
• Designer's role at a frontier lab: spot illegible ideas, understand what the energy is, translate them into form — through UX, storytelling, or prototype
• Example: an internal prototype called "Claude Studio" was dense and confusing but had strong internal energy; the skills framework and Co-Work's to-do/context display emerged from it

The three hiring archetypes

1. Strong generalists — block-shaped (not just T-shaped); 80th-percentile competent across multiple skills; already spans PM, engineering, and design thinking
2. Deep specialists — T-shaped with an unusually long stem; top 10% in one area (e.g., near-engineer-level technical depth, or world-class visual design)
3. Craft new grads — early career, wise beyond their years, humble, blank-slate learners; not burdened by entrenched process; often overlooked in favour of senior hires

• All three archetypes need resilience and willingness to change methods and tools
• For new grads: build real things, share them, find a community (e.g., Socratica model)
• For mid-career designers: use coding tools as part of your toolkit even without going deep on React; the abstraction layer will keep rising

On management and IC work

• Jenny moved back to IC at Anthropic deliberately — to stay close to how the work is changing
• Going back to IC revealed how much the role had shifted; managing without this would mean managing a process you don't fully understand
• Recommendation for design managers: take an IC rotation, similar to how engineering managers often rotate onto projects
• Rustiest skill returning to IC: receiving critical feedback regularly — presenting work is a vulnerable exercise
• The "low leverage task" reframe: senior leaders who dog-food products, file bug reports, and push small PRs appear to do low-leverage work but actually create high-leverage culture and team trust
• Middle management will persist but needs to combine direction-setting with people management — pure people management is not enough
• Psychological safety + high standards are not in tension; safety makes it easier to hold high standards without fear

Building high-performing teams

• Comfort level that allows teams to gently roast each other and their manager is a signal of genuine psychological safety
• Psychological safety alone is insufficient — pair it with visibly high standards
• The balance: teams know you are always there for them and know you will not fire them on a whim, and they also know you want excellent work
• Framework parallel: radical candor — caring deeply + challenging directly