design process
Jenny Wen (head of design for Claude Cowork at Anthropic) argues that the traditional design process — treated as "gospel" (research, diverge, converge, repeat) — is effectively dead, not because designers chose to abandon it but because AI coding agents let engineers move far faster than that process can accommodate; design has split into fast execution-support work (done via jamming and last-mile implementation in code) and much shorter-horizon (3–6 month) vision-setting, while humans remain necessary mainly as the accountable decision-makers about what actually gets built.
Problem framing: the classic design process taught as gospel (research → diverge → converge, "trust the process") is dead because designers no longer have time to make polished mocks before engineers, running many concurrent AI coding agents, have already shipped something.
Root cause: this is not a designer-led choice but a forced downstream effect of engineering velocity — engineers themselves struggle to keep up with their own agents — so designers shift from blocking/gatekeeping toward enabling, guiding, and connecting after the fact.
Consequence for time allocation: design work has split into two modes — supporting execution (mocking dropped from ~60-70% to ~30-40% of time, replaced by "jamming"/pairing and direct code implementation) and setting vision/direction (compressed from 2-10-year decks to 3-6-month prototype-driven direction), because the pace of change makes both long documents and far-out planning obsolete quickly.
Justification for less upfront mocking: nondeterministic AI models can't be fully represented in a clickable prototype or exhaustive mockup, so real execution with real models and real users surfaces use cases and produces a better product than pre-specified design.
Trust strategy: ship early under an explicit "research preview" label with known flaws stated upfront, then build and maintain user trust through visible, rapid, public iteration on feedback rather than through pre-launch polish.
Claim about AI capability: AI will keep getting better at taste and judgment, and designers may be over-clinging to "taste" as uniquely human, but a human must still be accountable for deciding what gets built — especially because AI can weigh in on but can't resolve disputes between people about what should or shouldn't ship.
Hiring implication: given this shift, Wen now favors three archetypes — strong "block-shaped" generalists, deep specialists (often highly technical, top ~10%), and adaptable, humble "craft new grad" early-career hires unburdened by legacy process — over conventional experience-weighted hiring.
Organizational implication: designers can act like internal VCs, using a legibility lens to scout illegible-but-energetic internal prototypes (e.g., the "Cloud Studio" prototype that seeded both Claude's skills framework and Cowork's UI patterns) rather than only executing defined briefs.
Leadership claim: in this fast-changing environment, effective leadership combines psychological safety with high standards (echoing Radical Candor), including counterintuitively doing nominally "low-leverage" hands-on tasks personally to build trust and product fluency.
Classic design process (research → diverge → converge) — The traditional methodology designers are taught — research/discovery, then diverge and converge repeatedly, summed up as "trust the process" — which Wen says is now effectively dead for AI-native teams. Apply: Historically applied by running a full discovery-to-mockup cycle before building; Wen argues teams should instead compress or skip stages of this cycle when engineering moves too fast for it.
"Don't Trust the Design Process" (talk thesis) — Wen's Berlin conference talk arguing the classic diverge/converge design process is obsolete. Apply: Use it as a prompt to re-examine whether a team's design rituals still fit its actual pace of engineering change, rather than preserving them by default.
"Last mile" design implementation — Designers doing hands-on polish/implementation directly in real code alongside engineers, rather than only handing off specs. Apply: Use coding tools (e.g., Claude Code inside VS Code) to make the final visual/interaction adjustments yourself after an engineer ships a first version.
"Research preview" labeling — Explicitly labeling an early, known-imperfect release (e.g., Claude Cowork) as a "research preview" to set user expectations upfront. Apply: Ship a rough version publicly flagged as preview/early, stating known flaws, instead of withholding it until polished.
Building trust through speed — Wen's principle that user trust in an early, imperfect release is maintained by visibly and rapidly iterating on feedback afterward, not by pre-launch polish. Apply: After shipping something rough, publicly respond to and fix feedback quickly (e.g., same-day fixes) so users see continuous improvement.
Design-system-in-code pointer — Directing engineers (and the AI agents they use) to the actual design-system code rather than dictating exact specs, since AI-written code doesn't reliably follow the design system on its own. Apply: Point collaborators to the design-system source and let them (or their coding agent) reason from it directly instead of issuing pixel-level instructions.
Principle-based feedback — Giving engineers the underlying reasoning/research behind a design choice instead of dictating the specific decision. Apply: Explain why a design choice matters (e.g., citing user research) so the engineer can extract and reapply the principle elsewhere themselves.
Claude widgets — Interactive UI components (e.g., for weather, stocks) that let Claude elicit input or display tactile content within the chat interface. Apply: Layer clickable/tactile widgets on top of the chat paradigm for cases where seeing/touching content is more efficient than typing.
Multi-channel chat access (WhatsApp/Telegram/SMS) — Making Claude reachable through existing messaging apps rather than only a dedicated interface. Apply: Integrate an AI assistant into channels people already chat in to lower the friction of using it.
2x2 leverage framework — An early-management-training framework sorting tasks into "only I can do," "anyone can do," and "low leverage," conventionally used to justify dropping the low-leverage bucket. Apply: Audit your task list against the three categories and normally cut the low-leverage tasks to focus on higher-leverage work.
Deliberate low-leverage leadership tasks — A counter-move to the leverage framework: senior leaders personally doing nominally low-leverage tasks (dogfooding, filing bug reports/PRs, personal gestures like Mike Krieger submitting PRs) because doing them personally makes them high-leverage acts of trust-building. Apply: Occasionally do hands-on, unglamorous work yourself (test the product, file a bug, make a personal gesture) to build deep product fluency and signal to the team that no task is beneath you.
Radical Candor — Referenced management concept combining caring personally about someone with challenging them directly. Apply: Pair psychological safety (no fear of being punished/fired) with maintaining high standards so feedback can be delivered and received directly.
"Roasting the leader" as a safety signal — Wen's informal indicator of psychological safety: whether reports feel comfortable joking about or imitating the leader's own habits/phrases. Apply: Treat team members feeling free to poke fun at you as a positive sign of trust and lack of fear, rather than a discipline problem.
T-shaped / block-shaped / F-shape (E-shape) skill models — Frameworks describing skill breadth vs. depth: T-shaped (one deep skill, broad base), block-shaped (broad AND deep across several skills), and F/E-shape (multiple T's combined, per Marc Andreessen) — used to describe generalist vs. specialist hires. Apply: Evaluate design candidates against these shapes: hire block/F-shaped generalists when roles span design, PM, and engineering, or T-shaped deep specialists (top ~10%) when a specific deep skill is needed.
"Craft new grad" hiring archetype — A hiring category for early-career people who seem wise beyond their years, humble, and fast learners unburdened by legacy process. Apply: Seek junior candidates who've already built real things on their own and can adapt quickly, rather than screening mainly for years of experience.
Socratica (build-and-showcase model) — A program/movement at Wen's alma mater centered on building projects and showcasing them like a science fair. Apply: Emulate it by building actual projects and sharing them publicly with a community as a way to develop and demonstrate design/craft skill.
Crits (design critique sessions) — A recurring design ritual of sharing work for group critical feedback. Apply: Keep doing regular crits — including while in an IC role — to stay sharp at giving and receiving critical feedback.
Legibility framework (Founder × Idea) — A VC framework (credited to Evan Tana of SPC) crossing founder legibility and idea legibility, where a legible founder with an illegible idea marks the most interesting/overlooked opportunity zone. Apply: When scouting ideas, treat a fully "legible" (easily understood) idea as a signal it's probably not novel, and instead look for illegible ideas attached to credible people.
Designers acting as internal VCs — Applying the legibility framework inside a company: scanning internal channels for prototypes that are confusing but have real energy/enthusiasm behind them, and helping shape them into products. Apply: Monitor internal experimental work (e.g., Slack, prototypes) broadly, and when something feels illegible but has genuine internal excitement, dig in and help develop it — as Wen did with the "Cloud Studio" prototype that fed into Cowork and the skills framework.
Skills framework (markdown instruction files) — An Anthropic/Claude concept, extracted from the "Cloud Studio" internal prototype, where markdown files instruct Claude how to perform a task. Apply: Write markdown files describing a task/skill so Claude can follow them when doing agentic work.
Introspection via AI-archive synthesis — Feeding a personal notes archive (interview notes, 1:1 notes, memos) into Cowork and asking it to synthesize a rubric or memo from it. Apply: Point the assistant at a folder of your own accumulated notes and give it a specific synthesis prompt (e.g., "read this and craft a memo/rubric for how I assess X") to surface patterns in your own thinking.
The retreat from mockups isn't purely about speed pressure — Wen frames it as structurally necessary, since AI models are nondeterministic and a clickable prototype literally cannot represent all their possible states.
Cowork's viral "built in 10 days" framing understates the real timeline: individual interaction patterns (to-do lists, multiple-choice prompts, onboarding) had each been separately prototyped internally for a long time beforehand; the 10 days was an assembly-and-ship sprint, not the whole build.
A single ambiguous, hard-to-explain internal prototype ("Cloud Studio") ended up seeding two separate shipped concepts — the Claude "skills" markdown-instruction framework and several of Cowork's UI patterns (showing Claude's plan/to-dos and working context).
Wen's stated signal for psychological safety is inverted from the usual assumption: it's a good sign, not a discipline problem, when reports feel safe enough to "roast" or imitate the leader's own habitual phrases.
Wen deliberately performs nominally "low-leverage" tasks (dogfooding, filing bugs, personal gestures) as a leader specifically because doing them herself converts them into high-leverage trust-building acts — a direct counter to the standard "eliminate low-leverage work" management framework she cites having been taught.
Internal Slack at a fast-moving AI lab is framed as a primary competitive information source — "the best AI news is probably internally... in the Slack" — implying public reporting lags well behind what insiders already know.
Even as engineering workflows move away from IDEs toward command-line/agent-driven work, Wen suggests the opposite may hold for designers, who still benefit from direct micro-edits (e.g., changing a CSS class by hand) that are faster than prompting an agent for small stylistic tweaks.
Wen uses Cowork on her own personal notes archive (years of 1:1 notes, memos, and interview notes kept in iA Writer) to surface an articulated hiring rubric she says she couldn't have consciously written herself — using the AI as a tool for introspection on her own past writing rather than an external task tool.
«This design process that designers have been taught, we sort of treat it as gospel. That's basically dead.»
— 00:00
«You as a designer actually like do not have the time to make these beautiful mocks anymore.»
— 00:05
«Few years ago, 60 to 70% of it was mocking and prototyping. But now I feel the mocking up part of it is 30 to 40%.»
— 00:15
«You're better off not blocking that, letting them cook.»
— 00:23
«How do keep up with all our agents? There are seven agents who are constantly running.»
— 00:32
«At the end of the day, someone has to decide what is actually going to get built and what actually matters. Someone still needs to be accountable for the decision.»
— 01:03
«I have opinions about it now, but ask me in like 3 months and it might actually change even more.»
— 08:39
«The ability to execute, try something out and try it with like real data and a real user kind of mindset in the product — I think that does lead to a better product.»
— 11:50
«Our Slack is a gold mine.»
— 15:17
«That is — the best AI news is probably internally, if you're ever at one of these companies, in the Slack.»
— 15:33
«I think the best design happens when you're able to just like throw a bunch of ideas at the wall and curate and just like push yourself to come up with a bunch of these different directions.»
— 20:53
«You have to show that you are continuously shipping and improving it because I think the way that you really lose trust around quality and releasing something early is if you release it early and then nothing ever happens.»
— 26:00
«A lot of the hard parts of building software are actually like not building it... it's like, oh, you and some other person disagreeing about what should go into this feature or what shouldn't go into this feature... AI can weigh in, but it can't necessarily solve this dispute between you and somebody else.»
— 29:04
«I think we undervalue just how good it'll get at this stuff.»
— 30:34
«Whatever we had, let's put it out there and then let's go out there and iterate from there, which is what we're doing.»
— 42:57
«It's like Claude, but Claude is really good at taking all your garbage and then turning it into something nice.»
— 45:44
«The best kind of like craft new grad folks I know are just like people who just like use the technology, build actual things, don't feel like limited by... how little experience they might have.»
— 50:57
«if both the founder and idea is like super legible, the idea is probably not that novel.»
— 62:38
«how can designers almost be more like VCs in this way internally.»
— 65:39
«hey, I have a bunch of interview notes and a bunch of notes related to design craft in this folder. Read it and then help me craft a memo rubric for how I assess craft in interviews.»
— 76:05
Reception
Comments are overwhelmingly critical and skeptical, pushing back hard on the claim that design process is 'dead,' with many accusing the video of clickbait, hype, and conflict of interest, though a few isolated comments are positive or neutral.
A candid first-person account of how one design leader's day-to-day has changed inside a frontier AI lab, most valuable for its specific process, hiring, and leadership details; its headline claim that "the design process is dead" is a broader generalization than the transcript itself fully supports, which is exactly what a large share of the audience pushed back on.

77:25