Brian Lovin - How to level up with AI as a designer

2026-04-14 - 55 min - source - Read full transcript

Vid (host)Brian Lovin

Key insights

The real design surface has shifted from static Figma screens to the agent harness itself.

Brian says as soon as he realized he couldn't design half of AI-native features in Figma, what he was actually designing was the harness that lets an agent do longer tasks and verify its own work. Static mocks can't simulate slow, imperfect, clarifying-question-asking model behavior, so he built prototype playground - a shared code base - to actually feel what using the AI product would be like.

ai-native-design-process

Notion has rewritten its agent harness roughly every 6 months since Brian joined, and he expects that pace to continue.

Each rewrite strips away old assumptions; the current harness is roughly 'give your agent the ability to write scripts and search for things,' but Brian assumes even that simple shape will be obsolete in 6 months, so he treats understanding why a harness works as the actual job rather than memorizing a fixed toolkit.

agent-harness-and-tooling

AI is still bad at last-mile visual fit and finish, so real designers still need to touch code or push Figma-to-code legibility.

Even leading models with extended thinking are, in Brian's experience, bad at visual interaction design. He sees two paths: go into the IDE and edit CSS directly for anything that matters, or invest heavily in naming Figma layers and components to match the code base so tools like Figma's MCP can translate a frame into decent code.

ai-native-design-process

Once a set of primitives is truly polished, AI becomes excellent at reusing and extrapolating from them.

Brian argues the common complaint that 'AI is bad at design' is often really a complaint about unpolished primitives. If you spend real time sweating the details on a good tertiary button or component once, AI can reliably reuse it and even extrapolate reasonable variations, like scaling it for a different context.

ai-native-design-process

Multi-model cross-review (Opus plans, Codex reviews, or vice versa) is useful but produces inconsistent verdicts.

Brian described running the same security-review prompt through two models on his Shuri side project: one flagged ten issues, the other called it security theater and said none of them mattered. He calls the inconsistency 'the pretty exhausting part of using AI today' and is skeptical that any fixed ordering of models and skills reliably solves it yet.

agent-harness-and-tooling

People who already knew how to code have a meaningfully better time with AI because they know the right upstream vocabulary.

Since models predict from the tokens you give them, precise input language steers output quality. Programmers already have words like 'durable workflow,' 'background job,' or 'queue parallelization' that let them describe intent precisely, while non-programmers lack that vocabulary and get worse results even with the same tools.

prompt-craft-and-legibility

Simon Last's 'simpler and dumber' prompt is Brian's single most-used prompt, run about 20 times a day.

The full prompt is: 'let's step back and think really hard, how can we make this simpler and dumber while still achieving our goals.' Brian applies it after every plan, every bug fix, and every new component, treating it as a default check against AI's tendency to over-engineer.

prompt-craft-and-legibility

Obsessing over job titles like designer, PM, or engineer will hurt people as AI blurs the boundaries between disciplines.

Brian argues these role boundaries are dissolving: designers should be able to ship code to prove an idea works, PMs should be able to design around their own specs, and engineers should care about visual component systems. The people who move fluidly between these types of work will do better than people anchored to a fixed title.

designer-identity-and-titles

Every company should treat their own product category as potentially obsolete and design for AI legibility now.

Just as it would be reckless for a note-taking or issue-tracking company to avoid examining its future relevance, Brian says it's reckless for individual designers to avoid the same self-examination. Notion shipped a CLI specifically so everything doable through its API becomes legible to AI agents, betting that APIs and CLIs will remain relevant even as UI paradigms shift.

designer-identity-and-titles

Heavy AI usage is not sustainable at a 24/7 pace, and Brian predicts a broad burnout correction within a couple of months.

He compares the current moment to the post-Claude-3.5/Sonnet-3.5 release bender he went on with his fiancee last year, prompting from their phones on vacation out of fear of wasting time, until he hit a wall mid-year. He expects most people currently obsessed with 24/7 AI usage to hit the same wall soon, absent a major new model capability release.

sustainable-ai-adoption

The ability to stay focused and undistracted for two hours a day is becoming a real competitive advantage.

Brian frames this as the counterintuitive outcome of the AI moment: as tool-switching and prompting-everywhere culture creates constant distraction and context-switch exhaustion, the people who can protect blocks of calm, focused time will out-execute people who stay maximally plugged in.

sustainable-ai-adoption

Media referenced

Lenny's tweet on designer growth - article - referenced as the trigger for a discussion on designers being the only role not currently experiencing headcount growth, correlated with slowing PM growth

Companies

Notion - Brian's employer since January 2025; the episode centers on his work on app builder, Notion agent, custom agents, and Notion workers
Campsite - the startup Brian worked at before Notion, referenced as the point where Cursor tab completion was 'getting good' but he remained skeptical of AI
Vercel - their AI SDKs helped Brian execute tools and return structured outputs while building prototype playground
Jitter - episode sponsor; released image-to-video AI generation
Dessen - episode sponsor; lets teams prototype directly in their own code base and design system without opening an IDE
Shuri - Brian's personal side-project read-it-later app, originally built on top of Notion, now standalone with its own API, CLI, and MCP
Sentry - error-reporting tool Brian has Claude hooked up to for debugging Shuri from user bug reports
Supabase - backend Brian uses for Shuri that Claude can query for account information when debugging
Axiom - log drain Brian uses so Claude can replay exactly what happened leading up to a Shuri failure
Cursor - coding tool Brian used heavily last year before Claude Code; still reaches for its Composer 2 fast mode for front-end pixel polishing
Conductor - Brian's preferred middle-ground tool: lets him see the code while multitasking across several chat-driven agent tabs at once
Linear - cited alongside Notion as a company that might plausibly declare 'issue tracking is dead' given how fast the ecosystem is changing

Techniques and frameworks

Prototype playground - Notion's internal shared code base of AI-native prototypes with a shared component kit close to production Notion, used instead of Figma mocks to actually feel what interacting with AI is like
Designing the agent harness - Brian's reframe that a designer's real job now is designing the scaffolding that lets an agent do longer tasks and verify its own work, not designing static screens
Model cross-review - having one model (e.g. Opus) create a plan and a different model (e.g. Codex) review it, then swapping roles, to catch things a single model misses - though Brian finds the quality of these reviews inconsistent
Making intent legible to AI - naming Figma layers and components to match code base naming so MCP tools and coding agents can translate a design directly into working code
The 'simpler and dumber' prompt - a prompt from Notion co-founder Simon Last that Brian runs roughly 20 times a day: 'let's step back and think really hard, how can we make this simpler and dumber while still achieving our goals'
Weekday-plan, weekend-execute rhythm - Brian's workflow for Shuri: mornings triage overnight bugs via a 'fix issues' skill, evenings sketch feature intent for Claude to plan, weekends are for building against the accumulated plans

Summary

Brian Lovin, a designer at Notion working on AI products since January 2025, walks through how a year of building AI-native features changed his entire design process. He arrived skeptical of AI, having watched early Cursor tab completion and found it unimpressive, and his first project - an ambitious "app builder" that let Notion generate any tool from primitives like databases and AI chat - initially failed because he approached it like a traditional Figma-first design project. Static mocks couldn't capture how slow, imperfect, and clarifying-question-heavy real model behavior actually was. That failure pushed him to build prototype playground, an internal shared code base where designers build real, interactive AI prototypes instead of static screens, which he says permanently changed how he thinks about the job: designing the agent harness itself, not the interface on top of it.

Much of the conversation covers how fast that harness keeps changing. Notion has rewritten its agent harness roughly every six months since Brian joined, each time discarding assumptions that had felt foundational months earlier. He treats his job as staying close to engineers building the harness and understanding why models behave the way they do, rather than mastering any fixed toolset. On tooling specifically, he describes cycling through Cursor, Claude Code in the terminal, and now mostly Conductor for multitasking across several chat-driven agent sessions, while reaching for Cursor's fast Composer mode for front-end pixel polishing. He's skeptical of tribalism between Claude, Codex, and Cursor camps and argues designers should try everything and build real intuition for which tool fits which task, the same way a designer learns when to reach for a frame versus a group in Figma.

A recurring thread is that AI is still bad at last-mile visual fit and finish, so real quality work still requires touching code, or investing in making Figma files legible to AI by naming layers and components to match the production code base. Brian's counterintuitive finding is that once a set of design primitives is genuinely polished, AI becomes excellent at reusing and extrapolating from them, which reframes complaints about "AI being bad at design" as really being about unpolished source material. He also describes the inconsistency of multi-model review workflows: giving the same prompt to two different models can produce opposite verdicts, which he calls the most exhausting part of current AI work, resisting the temptation to believe any fixed process fully solves it.

On language and prompting, Brian explains that since models predict from the upstream tokens you provide, people who already know how to code get meaningfully better results because they have precise vocabulary (durable workflow, queue parallelization) that non-programmers lack. His most-used prompt, borrowed from Notion co-founder Simon Last, is a request to "step back and think really hard, how can we make this simpler and dumber" - run roughly 20 times a day. He also shares detail on Shuri, his personal read-it-later side project, now standalone with its own API, CLI, and MCP, where he's built a debugging loop that lets him paste a user's bug report into Claude, which is hooked up to Sentry, Supabase, and Axiom logs to replay exactly what happened.

The conversation closes on designer identity and sustainability. Responding to Lenny's tweet about design being the only role not currently growing, Brian argues that obsessing over titles - designer versus PM versus engineer - is what will hurt people, since AI is dissolving those boundaries; the people who move fluidly between disciplines will do better than those anchored to a fixed box. He extends the same logic to companies: any business, including Notion itself, needs to seriously examine whether its current shape survives the next few years, which is part of why Notion shipped a CLI to make its API legible to AI agents. Finally, he predicts a broad AI-usage burnout correction is coming soon, drawing on his own experience hitting a wall after a 24/7 AI bender last year, and argues that the ability to stay calm and focused for two uninterrupted hours a day is becoming a real competitive advantage.

Notable Quotes

"As soon as you realize you can't design half of this stuff in Figma, what you're really designing is the harness for the agent to do longer things and verify its own work." - Brian Lovin

"Let's step back and think really hard. How can we make this simpler and dumber while still achieving our goals?" - Brian Lovin (quoting Notion co-founder Simon Last)

"Our obsession with titles is what will screw people over. If you're like, 'But am I a designer or a PM or an engineer?' No, no, no, no. These things are going away." - Brian Lovin

"The ability to focus and be undistractable for 2 hours a day is like a meaningful competitive advantage right now, which is insane." - Brian Lovin