“Taste is trainable” - Head of product at Notion (Max Schoening)

Why does it matter? Because the traits that made product people irreplaceable are being repriced — and most don't know it yet.

Max Schoening has been a PM at Google, ran design at Heroku, shipped code as a design leader at GitHub, founded two companies, and is now head of product at Notion. That unusual path turns out to be exactly the right vantage point for this moment. What he sees from there is both clarifying and uncomfortable: the bottleneck in product work has quietly shifted from skill to agency, from craft to simulation, from execution to taste — and most teams are still optimizing for the old bottleneck.

• Taste is trainable — it's literally the ability to run a mental simulation predicting how a specific group will react to an idea, and it compounds with reps the same way a model improves with training data.
• PMs and designers should code not to ship features but to understand the medium — especially agent loops, which cannot be meaningfully designed from the outside.
• Agency has replaced skill as the scarce resource, because AI commoditized skill; the people who thrive are those who treat the world as malleable.
• Every great product has one tiny core superpower, and adding features to compensate for a weak core has never worked and will become more lethal as AI makes feature-building cheaper.

Taste is a trainable simulation skill, not a personality trait

Run a virtual machine in your head. That is Schoening's literal definition of taste: "given an idea, you can predict for a certain in-group whether they're going to like it or not." The mechanism isn't mysterious — it's accumulated reps with feedback, structurally identical to how a model trains. "You just have to do reps. It's almost like training a model."

He points to Japanese craftspeople painting bowls for decades as the clearest analogy: the taste isn't innate, it's the residue of an enormous volume of iterations with real reactions attached. The designers at Notion who Schoening identifies as having the sharpest taste share two habits: they run side projects where they own the full experience end-to-end, and they are relentlessly trying new tools — the person on the team who has suggested 49 new apps and counting. Full ownership accelerates the feedback loop; constant exposure to other people's ideas calibrates the in-group model.

One structural implication he draws: Notion's conference rooms are named after objects of lasting craft — the first typewriter, the Macintosh, a Porsche 911. The purpose is deliberate friction: when you're sitting in that room and you actually look at what the room is named after, "nothing I'm doing amounts to this." Environmental pressure as a rep accelerator.

The corollary he raises — and doesn't fully dismiss — is that taste itself may not be the permanent human moat people assume. If the loop is input, idea, reaction, adjust, that is backpropagation. "I'm not so sure" the last human advantage is taste, he says. Which makes building it urgently, not leisurely, the actual takeaway.

PMs and designers should code to understand agent loops — not to ship CSS tweaks

The origin story matters here. When Schoening joined Notion, the team was designing chat interfaces in Figma. He invokes Brett Victor's talk "Stop Drawing Dead Fish": a static image of a chat interface is the dead fish. You have to feel the AI. So he and two designers built the worst possible playground — a small, LLM-friendly codebase, optimized not for production but for one-shottability — and moved all chat prototyping there.

That playground was explicitly a ramp. The goal wasn't to ship from it; it was to get people past the fear of the terminal so that coding became "just chatting." Designers and PMs who went through it are now, with improving model capabilities, also contributing to the production codebase — to a lesser degree, but the trend line is visible.

The sharper principle Schoening articulates: "I actually don't care at all whether designers write code that lands in production. The reason I like thinking in code is because it forces you to consider the medium." He frames the real choice as between a PM who can tweak UI details in Codex versus a PM who deeply understands how agent loops work. He would take the latter every time — "and the only way that you can actually get to understanding agent loops is if you build them in the material that they're made of, which is currently code."

At Notion, designers are now using the terminal. Not the GUI wrappers — the terminal itself, because Schoening encourages it, knowing they will eventually pull on other threads and understand more of the substrate. The goal is not to turn designers into engineers. It's to make sure nobody is drawing dead fish.

Agency is the new scarce resource — and most people who think they have it, don't

Before AI, skill gaps were a socially acceptable excuse. "I will never be able to do this because insert skill issue." That sentence no longer holds. When an AGI-adjacent model can close most skill gaps on demand, the thing that actually differentiates people is whether they act without permission — agency.

"I don't think agency is very evenly distributed in the world," Schoening says flatly. The people who will do great are those who understand the world around them is malleable. The ones who will struggle are those still asking what it means to be a PM, what it means to be a designer, what their job description permits.

Two examples from Notion make it concrete. Brian Leven blurs engineering and design by default — but his real demonstration of agency is that he became Notion's number one recruiter, going out to find people the org needed without being asked, because he wanted to affect change regardless of how it happened. Eric Lou asked Schoening directly: if you started a startup, would you hire me? Told "not in the first 10, I don't need a PM," he said "I'm going to work on the skills so you'd hire me in the first five" — then moved from writing PRDs to Figma to building prototypes in code, redesigning his own role by degrees.

Schoening's practical heuristic for developing agency: make things. Tinker. A home-cooked meal counts. The treadmill of making builds the reflex that the world is something you can change — "you could just change things" — which is the actual cognitive foundation of agency, not organizational cunning.

Every great product has one tiny core superpower — adding features to compensate for a weak core never works

GitHub's core is the pull request — the idea that anyone can suggest a change and you can see it. Heroku's was git push heroku master: one line, and something on your laptop has a URL. Dropbox's was the menu bar icon that synced so reliably you could use it as a proxy for whether you had internet — because it was better at detecting connectivity than the Mac itself. Notion's is blocks and slash commands. Figma's is the seamless blend of real-time collaboration with the design tool itself. Multi-touch on the original iPhone.

"All the great products have something tiny that is a superpower," Schoening says. "One tiny core that is so exceptionally good." The anti-pattern is the loop he's seen kill products repeatedly: if I just add one more thing, it will finally be great. "That never works."

He lived the failure version himself. In 2014 he was building a Notion competitor. They spent enormous time perfecting the editing experience — markdown folding, features that now live in Obsidian. The week they were about to get a term sheet from True Ventures, Notion pivoted from website building to document collaboration, and the investor cited a conflict. But the real lesson wasn't the timing: "The first version of the Notion editor was terrible. You couldn't even select between two blocks. But it turns out it didn't matter." The core of blocks and flexibility was right. The polished editing experience his team built was working diligently on the wrong thing. They then went down the death spiral — kept adding features asking "is it good now?" — and the answer was always no, because the core wasn't good.

With AI making it cheaper than ever to add features, this trap will claim more products faster.

The SaaS apocalypse is exaggerated — the 'as-a-service' part is exactly what people don't want to rebuild

Decompose the acronym. A lot of SaaS in the 2010s was, as Schoening puts it, "a very fancy form around a spreadsheet" — it guided people to fill out the form correctly, which is actually less malleable than a spreadsheet. That part is genuinely vulnerable to AI. But the "as a service" component is different: "I don't think most people actually want to maintain the full stack of software."

He tried rebuilding Notion in a weekend for himself, just to pressure-test his own frustrations with it. He doesn't think people actually want that. "Software is like a garden. You need to tend to it, and the thing you pay for in the as-a-service is the maintenance and a bunch of specialists thinking really hard about a problem." That doesn't go away when AI gets cheaper.

Anthropic — a company that could plausibly build anything — runs on Slack. Nobody there is rebuilding Slack. The reason is that Slack's notification decision flowchart, the one with dozens of conditional branches for how to deliver a single notification, only exists because of real users, real scale, and decades of understanding the customer. You don't get that from a weekend vibe-coding session.

Schoening's prediction: software moves back toward the general-purpose tools of the '90s — word processors, spreadsheets, FileMaker Pro — but those will still be delivered as a service, and specialized tools for security and deep domain problems will persist. The opportunity isn't to kill SaaS; it's to make it more malleable. Notion's own data point: a journalist tweeted that thanks to Notion AI she finally understood and used Notion. AI as a tutor unlocking a general-purpose platform is the model, not AI as a replacement for needing the platform at all.

AI is getting better at coding at an exponential rate — progress in every other domain is mostly coding principles applied elsewhere

"It's very clear at least empirically that models are getting better at coding at some exponential rate right now. I don't think that's changing." What Schoening is not impressed by is progress in any other domain. Writing, reasoning, other knowledge work — he's not seeing the same curve.

His read on announcements of breakthroughs in non-coding domains: "You just applied coding principles to this domain." Which is valuable — but it means the underlying engine is still the same thing getting better at the same thing, not a generalized intelligence expansion.

The implication he draws is structural: "Software is eating the world" accelerates. If the cost of encoding business practices in code approaches zero, there will simply be a lot more code, running in a lot more places. HR teams at Notion are already automating things they previously had to bug an engineering team to build. Finance, marketing, legal — each function that learns to encode its own processes in software is benefiting from the one curve that is genuinely compounding.

He's also skeptical of the frontier intelligence race for most knowledge work applications. The retina display analogy: after you can't see the pixels, making them smaller doesn't help. At some point models reach "good enough" for most tasks, and the optimization switches to cost, speed, and local deployment. Cursor, Intercom, and Notion itself are all experimenting with fine-tuned or self-hosted models for exactly this reason — not because frontier intelligence stopped mattering, but because good enough unlocked the ROI calculation.

The first 10% of any project is now free — but the last 10% is still 90% of the work

"The first 10% of every project are now free. That's how I would describe it." It takes almost no effort to build the first version of a startup — or even a version 0.8. But Schoening is direct about what hasn't changed: "the last 10% are still actually 90%." Getting from a demo to something reliable for 100 million people is the same engineering problem it has always been.

The physical metaphor he reaches for: a hardware startup making first enclosures with 3D printing. You can see all the layer lines. It is visually obvious this is not the shipping product. Software prototypes don't have visible layer lines — which is why so many people mistake them for being closer to done than they are.

What has changed is the cost of exploration. You can now send off 10 parallel experiments and see which assumption was right, where before you would have committed to a waterfall and found out much later. At GitHub, the team's mantra was "demos not memos" — give me something to react to. Now that bar is dramatically lower. "Here's the version of the product — what if we did it this other way? Here's that version."

The watch-out Schoening raises about vibe coding: the amount of software has increased, but reliability hasn't kept up. Quality — that "Apple-esque machined unibody aluminum engineering" — remains the hard thing. Cheap exploration is a gift for running more shots on goal. It is not a shortcut past the engineering required to make any of those shots actually count.

The product world is quietly splitting into two kinds of people — and the divide will widen fast

Every structural shift Schoening describes points toward the same underlying fork: those who treat the world as something they can change, and those who wait to be told what their role allows. When skills were scarce, the second group had a defensible position. That position is gone.

What's coming isn't a world where everyone codes — it's a world where every function encodes its own processes, where the people who move between disciplines without asking permission reshape the ones who don't, and where the tiny core of a product matters more, not less, as AI makes surrounding noise cheaper to generate. The makers win. The describers of what makers should do face a much harder decade.

Drive it like it's stolen.

---

Topics: product management, AI tools, malleable software, agency, taste, SaaS, designer-engineer convergence, agentic products, Notion, prototyping, career development