What Product Taste Actually Is

In my last post I argued that the bottleneck in product work has moved from building to judgment. I used the word taste for that judgment. The more I’ve sat with it, the more I think we under-describe what we mean, and then wonder why the skill feels slippery.

Taste is not vision. Vision points at a horizon and asks people to walk toward it. Taste looks at something in front of you (a prototype, a flow, a pricing page, an AI-generated option) and predicts, before users tell you, whether it will land. One is a bet on the future. The other is a read on an artifact now.

Taste is also not preference. Preference is what you like. Taste is what you expect users to need, return to, or pay for. The two get confused constantly, especially by experienced PMs, because preference arrives dressed as judgment. “This feels off” can be a calibrated signal or just your aesthetic. The only way to tell is contact with reality: ship, watch, update the model in your head.

Why this matters more now

What makes taste scarce is speed. When building took months, weak judgment hid behind process: discovery, alignment, roadmaps, reviews. All of it bought time for the right answer to emerge socially. When you can generate fifteen variants before lunch, that process can’t keep up. Someone still has to look at the pile and say: these twelve die, these two get sharper, this one ships. That someone is exercising taste whether they name it or not.

AI didn’t remove the need for that someone. It multiplied the pile.

In a recent interview on The Knowledge Project, Shopify CEO Tobi Lütke named a failure mode that should feel familiar inside any product org right now: the “slop grenade.”¹ Someone asks a model for output, doesn’t really check it, and tosses the result at a colleague to unpack. The sender feels productive. The recipient inherits a decompression job. Sometimes they feed it into another model just to survive the email. Work didn’t get faster. Ambiguity got redistributed.

I keep coming back to that image because it clarifies what taste is for in 2026. Taste is the refusal to throw the grenade. It’s the few minutes you spend deciding whether an artifact is worth someone else’s attention, including your own future attention when it ships to users. When generating text, code, specs, and UI is nearly free, the scarce act is not production. It’s filtration with accountability attached.

Lütke’s broader point lands where my last essay did: machines can inform decisions, but they can’t own them. Only a person can be wrong in a way that carries consequences. Taste is how that ownership shows up day to day. Not as a grand strategy offsite, but as a hundred small calls (keep, revise, kill) made before the pile becomes someone else’s problem.

Taste is compressed experience

I don’t think taste is mystical. It’s compressed experience. Every cycle of saying “I think this will work,” putting it in front of people, and seeing what happened updates a private model of what good looks like in your domain. High volume with stakes attached compounds. Low volume with no stakes (endless scoring in private, no kill decisions, no user contact) mostly trains confidence, not accuracy.

There’s a risk worth naming. Taste built only on AI outputs makes you fluent at ranking LLM style. Taste built only on competitor screenshots makes you fluent at copying. Taste built only on your own instincts, never recalibrated, decays into taste theater. And taste that never has to absorb a slop grenade you almost threw yourself never learns the cost of skipping the check.

The muscle stays honest only when it keeps colliding with users who don’t share your priors, and with colleagues who shouldn’t have to clean up your unfinished thinking.

A posture, not a framework

So when people ask how to get better at product taste, I increasingly think the honest answer isn’t a framework. It’s a posture: treat every candidate solution as a hypothesis about human behavior, make the call visible, and let reality edit you. Don’t outsource the judgment step to the next person in the thread.

The PMs who will matter next aren’t the ones with the cleanest docs or the longest AI-assisted briefs. They’re the ones whose private model of what good looks like keeps getting sharper, who will kill their own favorites when that model says so, and who refuse to confuse “I generated it” with “I decided it.”

The harder question I’d love your take on: the last time you almost threw a slop grenade (or caught one), what did it teach you about your own judgment under speed?

1. Tobi Lütke, “The Skills AI Makes More Valuable,” The Knowledge Project (September 2026). ↩