自主性比职衔更要紧：Notion 产品负责人谈 AI 时代怎么做产品

Schoening says he would not take credit for designers and PMs at Notion now writing code, but he tells the origin story. When he joined, the team was designing chat interfaces in Figma, and he cites Bret Victor's talk "Stop Drawing Dead Fish": a static image of a chat does not let you feel the AI. Two designers and Schoening built a deliberately rough playground, a small LLM-friendly codebase, and moved their chat prototyping there.

The main Notion codebase is not very agent-friendly after a decade of patterns, so the playground was made as unscary and as one-shottable as possible. Once people got past the fear of the terminal, it became just chatting. He sees it as a way of getting people on the treadmill, and as models improve the same designers and PMs also contribute to production code to a lesser degree.

He is wary of vibe coding because the quantity of software has risen but its quality, in his view, has not. Asked by host Lenny Rachitsky about the balance with strategic work, he says he does not care whether designers' code lands in production. He would take a designer or PM who understands how an agent loop works over one who can merely tweak styles, because you can only learn agent loops by building them in code, the material they are made of.

What separates the people who do well, Schoening suspects, is the same thing that has always separated founders from non-founders: agency. It used to be easy to say you could never do something because of a skill issue, but when a capable model puts skills at your fingertips, what matters is whether you act. He says agency is not evenly distributed, and that people who see the world as malleable will do better than those asking what it means to be a PM, designer or engineer.

He offers Notion examples. Brian blurs engineering and design and is also something like the top recruiter, going out to find people the organisation needs. Eric once asked whether Schoening would hire him in a hypothetical startup's first ten people, was told a PM would not make the first ten, and set about building skills to be in the first five, moving from long PRDs to Figma to simply building prototypes. Schoening tells himself to "drive Notion like it's stolen".

Schoening worries that merged roles could cost the industry its specialists. Using a hardware analogy, he notes that 3D-printed prototypes show their layer lines, while the long road to manufacturing for 100 million people is engineering, which is largely missing from talk about token counts and feature counts. On the design side, anyone can now assemble a usable interface, but craft and delight still need specialists.

To cultivate agency, he quotes the Steve Jobs line that the world is made by people no smarter than you. The common thread he has found is making: tinkering, cooking, tweaking a chair. He cautions against treating agency as a way to circumvent a bad boss; start by making things, and people notice once you get good.

Schoening defines malleable software as software that works closer to the interest of the people who use it than the interest of the company that makes it. He says that as an Apple fan he does not want software designed only in an ivory tower, and compares it to a home where you cannot rearrange the living room. Apps glue together interface and data ownership into a square on your phone, so a small change is usually impossible.

The other extreme is running your own Linux distribution, which gives malleability but also constant yak shaving. For him it comes down to whether you have ownership of your computing life. Individual homemade tools are a form of malleability, he says, but they need a platform that encourages it, and he likes communal tools, which is the problem he works on with Notion colleagues, including Jeffrey Litt.

Lenny asks about a Dieter Rams video that Schoening pins on his profile, in which the designer scornfully walks past famous chairs. Schoening, who is German, finds it hilarious but says the point is that design should be useful first and beautiful second. A good test of usefulness is whether you can change and tweak the thing, and he cites *How Buildings Learn* and the idea that the best homes adapt to their occupants over time.

On the claim that people will build their own tools instead of buying Salesforce or Notion, Schoening says it would be silly to claim nothing changes. Much of 2010s SaaS was a fancy form around a spreadsheet, less malleable than the spreadsheet and valuable because it guided people. But he has tried rebuilding Notion in a weekend and says most people do not want to maintain a full software stack.

He compares it to buying steak in styrofoam at Costco instead of hunting, and cites the line that software is like a garden that needs tending. What you pay for as a service is maintenance and specialists thinking hard about a problem. Lenny adds that Slack is heavily used even inside AI companies, and Schoening points to the sheer complexity behind something like a Slack notification flowchart, which only comes with real users and scale.

His prediction is that tools will become more general and go back toward the 90s model of word processors, spreadsheets and FileMaker Pro, still delivered as a service, alongside specialised tools such as security products. He notes that a journalist tweeted that Notion AI finally helped her understand Notion, since AI acts like a tutor. Schoening concludes that the apocalypse is greatly exaggerated, though things will not stay the same.

Asked what has changed most in his job, Schoening says the first 10% of every project is now free. There is little point in writing a PRD when you can build a janky demo and show what you think should be built. He recalls GitHub's slogan "demos not memos", and says it is now easier to give people something to react to, which builds iteration into the product much earlier.

Even if the first 90% is done, he says, the last 10% is still effectively 90% of the work. What changes is that exploration is cheap: you can send ten agents down ten paths and see whether you were right. He is conflicted about what comes next, believing in durable plain text and code, but wondering whether the future is just chatting back and forth.

He raises two open questions. If inference becomes near-instant, do we return to direct manipulation, moulding code like clay, rather than queuing jobs and wandering off? And is intelligence saturating like a retina display, where past some point people want a different modality and more speed rather than more IQ? He says he is more interested in an exoskeleton than a god in a data center, since society is largely not capped by intelligence.

Notion's policy on token spend, as Schoening understands it, is unlimited, because it is the wrong thing to optimise for while people are exploring something new. He suspects that in six to 12 months many companies will start asking uncomfortable ROI questions. He warns against boasting about token spend, which he likens to boasting about lines of code written in a day.

He nonetheless sympathises with companies that run leaderboards, because he has been surprised by how much prodding it takes to get people to identify the outer loop of their work and build a software factory around it. Roles far from engineering are easy to convince because they suddenly have superpowers, and then you have to ask whether they understand why they cannot merge a PR. For engineers, he echoes Simon Last's idea that any manual intervention in code should feel a bit like a bug, while still reviewing code carefully.

On tools, he is not sure Figma is declining and says he dislikes rivalry framing, since both Figma and vibe coding could rise under Jevons paradox. What surprised him is the terminal: it is scary at first, but once PMs are in Claude Code or Codex they are fine, and he encourages them to use the terminal over GUIs because curiosity will teach them how computers work. Designers also use developer tools such as Conductor.

Asked which role AI transforms next, Schoening gives a self-described hot take: models are clearly improving exponentially at coding, but he is not impressed by progress elsewhere and still dislikes AI slop writing. If the cost of creating software falls toward zero, there will simply be more of it, and software engineering will move into other domains. HR staff automate tasks without bugging engineers, and when labs claim progress outside coding, he says they are applying coding principles.

Lenny adds that a Codex product lead said every winning agent will be a coding agent that builds what it needs, and Schoening agrees that all agent harnesses resemble coding agents. On the cost question, he sketches two futures. If the gap between frontier labs and open-weight models widens, the labs decide what the world looks like, which he dislikes because he hates centralisation. If it does not, more companies like Cursor and Intercom will tune their own models for tasks where good enough is cheaper.

He says we are in a lucky timeline with several competent US labs, and expects layers to commoditise like the cloud wars. He draws on Heroku, which lost to Kubernetes despite better user experience because Heroku said it would replace your ops team while Kubernetes promised to make your ops team superheroes without lock-in. Businesses want choice, he says, and products that make users feel like superheroes win, which is also why automatic code review that roasts you works worse than tools that let you publish your work with Claude.

Asked why the Notion agent is well loved, Schoening says he is his own worst critic but is surprised by how good it is. Notion launched an assistant before ChatGPT, he says, and agents need context rather than narrow pipes to data. For the first time it is obvious why a connected workspace is valuable, and because he thinks of Notion as an operating system, it resembles the Unix environment coding agents live in. He credits Notion's enterprise search and automatic permission handling, and a willingness to try hard things.

On shipping, Schoening says that when a company is following a hugely successful first act, as at GitHub or Notion, people can get precious. He reminds them that "you can just do stuff" and that users will be more upset if the company stops innovating than if it breaks something, and the team talks about increasing shots on goal. But feature count is a silly metric like lines of code, and he would rather have fewer, very good features.

Quality is something he is still struggling with, and he says the labs are not exempt, citing regressions in their tools. His team's standard is "obviously good": you know it when you see it, like the first iPhone or ChatGPT. His core value is incremental correctness, meaning iterate and get good at iterating, but also consolidate. Notion has six automation primitives, he says, and someone has to reconcile them into the simple core even if the next release slips.

Lenny asks how to build taste, and Schoening says he is not sure he has great taste but watches how others use it. The common ingredient is iteration with feedback, and it takes a long time in any one domain. He describes taste as running a virtual machine in your head that predicts whether a chosen in-group will like an idea, and says you decide who your in-group is and then get good at emulating their reaction.

He compares this to training a model: input an idea, see how people react, and update, which looks like backpropagation, so he doubts taste is the one thing humans have left. Japanese craftspeople who paint the same bowl for years are his example, and he says you cannot speedrun it, only increase the frequency of reps. Designers with high taste, he notes, tend to have side projects they own end to end and are always tinkering with new apps.

He also advises surrounding yourself with tasteful things so your own work feels lacking. At Notion the conference rooms are named after famous objects such as the first typewriter, the Macintosh and a Porsche 911, which reminds him that nothing he is doing amounts to those and he has to do better.

Asked what matters in building a successful product, Schoening says he has contributed to great products rather than built them, and now thinks the team is the only thing that matters. He rejects the idea that the best design or the best engineering always wins. The biggest pitfall, he says, is the loop of thinking that one more feature will finally make the product great, because truly great products have one tiny core that is exceptionally good.

His examples are multitouch on the iPhone, the pull request at GitHub, blocks and slash commands at Notion, and the blend of real-time collaboration at Figma. At Heroku it was `git push heroku master`, which took you from something on your computer to a URL; at Dropbox it was the menu bar icon that synced so reliably that it served as an internet indicator. For years Dropbox widened its surface area while he kept thinking he only wanted the one job.

He also says being first is overrated: Bluetooth headphones were poor until AirPods, and Apple was not the first MP3 player maker. He prizes durability over virality and cites Anthropic, which started later with less funding, as an example of getting it right. On jobs to be done, he uses it as a prompt to ask what users would hire your product for and whether you are honest about what they want, and to stop acting like an employee who made a thing. His example is landing pages: people write in marketing speak, and he asks them to explain the product at a whiteboard as they would to a friend.

On universal basic income, Schoening asks for some grace because he is half joking when he says we already have it and it is called knowledge work. What humans need to live and be content is far less than the hierarchy of jobs suggests, he says, and he does not exempt his own. He expects people to keep inventing reasons to stay in the loop around agents, and says that with AGI he would do the same thing: code for the intellectual challenge, build tools and make the world more malleable.

In the contrarian corner, he argues that inclusivity is not always great: he believes the world is run by group chats of eight people or fewer, and a product that aimed at all 8 billion would upset its first 500 million, since the top of the class wants different things. Under failures, he lists loosening the requirement for designers who can code, not pushing for package management in GitHub Actions, and polishing a markdown editor for a Notion competitor in 2014 while Notion's own editor was rough. He had worked diligently on the wrong thing for too long, he says, though you should not abandon the core idea and should relentlessly iterate.

His closing advice to younger people is not to let the rush or the fear of a last train and a permanent underclass distract them from what they love, though working very hard in your late teens and twenties is fine. He suggests tuning down the amplitude of worry, reading history and tying it back to agency. In the lightning round he recommends *Code* by Charles Petzold, *Tools for Conviviality* and *Seeing Like a State*, and he ends by telling listeners to walk around a human-made environment and realise how much they could make from scratch.