What Is Left to Decide When AI Can Build Anything?

For most of the history of software, the hard part was building. Turning an idea into a working product took teams, months and budgets, and the scarcity of execution shaped how product people spent their days. As AI takes on more of that execution, the scarce resource moves upstream. When a system can produce whatever it is asked to produce, the question that matters is what it should be asked for.

That upstream question has a name: product taste, the ability to know what should exist before it does. It covers three decisions that sit before any build begins: what to make, why it matters, and exactly how it should feel. Each of those decisions used to be blended into execution, corrected along the way by the friction of building. As that friction shrinks, the decisions stand on their own, and their quality shows directly in the result.

Taste is often treated as a gift, a matter of temperament that some people happen to have. A more useful view treats it as a discipline: a trained judgment with parts that can be named, signals that can be read, axes that can be measured, habits that build it and traps that erode it. Seen this way, taste becomes the most teachable advantage a product person can hold in an era of automated execution.

Taste as trained judgment

The first move is definitional. Opinion is a statement about what someone likes. Taste is a claim about what will work, made by someone whose judgment has been shaped by sustained exposure and practice. The difference lies in accountability: a taste claim can be tested, argued and defended against other trained eyes, and it can turn out to be wrong.

That property, falsifiable judgment, is what makes taste improvable. Someone who says a flow feels confusing is offering a view. Someone who says the third screen asks for a decision before the user has any reason to make it is offering a claim others can check. The second kind of statement compounds over time, because each test sharpens the next judgment.

Trained taste also tends to arrive before explanation. Experienced product people often sense that something is off well before they can say why, a smell test built from many earlier encounters with similar problems. Taste also works as compression: a single decision, such as what a product chooses to leave out, can reveal an entire mental model of who the product serves and what it is for.

A related distinction separates what users ask for from what they need. Requests describe symptoms in the user's own vocabulary. Needs describe the underlying problem. Taste is the faculty that reads through the first to reach the second.

Reading taste in the wild

If taste is a judgment, its traces should be visible in the things it produces. They are. A product carries its makers' decisions on its surface, and a trained reader can find them before consulting a single review score.

The first thirty seconds of onboarding are one such surface. What a product asks for first, what it explains and what it trusts the user to discover all signal how well its makers understand the person arriving. Intentional restraint is another: the features a product declines to include often say more about its judgment than the features it ships.

A changelog, read the way a critic would read it, shows what a team chooses to fix, extend or leave alone over time. A product's own copy, the words on its buttons and in its empty states, offers a direct window into how clearly its makers think.

These signals turn taste from a private sense into shared evidence. Two people can look at the same onboarding flow and argue about what it reveals, and that argument is exactly the kind of exercise that trains judgment.

Three axes and a high bar

Shared evidence invites measurement. A workable framework rests on three axes that together describe most great product experiences: clarity, restraint and surprise.

Clarity is how quickly a person understands what the product is and what to do next. Restraint is the discipline of including only what earns its place. Surprise is an unexpected moment that the user immediately appreciates.

Satisfaction metrics capture part of this picture. A score such as NPS measures whether people are content, and delight calls for a different kind of evidence. The most demanding test is simple to state: would people miss the product if it disappeared? A product that clears that bar has become part of how someone works or lives.

The axes also separate taste from polish. Visual refinement can make a product beautiful while leaving it confusing or overloaded. Beauty and taste overlap, yet they remain distinct qualities, and a product can have one in abundance and lack the other.

How judgment gets built

A trained judgment needs training material. The habits that build product taste start with a consumption diet: deliberate attention to what to read, use and study. That diet reaches well beyond software. Film, architecture and food all teach how experiences are composed, paced and edited, and broad aesthetic taste feeds directly into product taste.

Consumption becomes training through structure. A structured teardown takes a product apart to understand why each decision was made. The habit applies equally to products that are loved and products that went wrong, since well-understood mistakes teach as much as successes.

Writing adds a further discipline. Putting a judgment into words forces it to become specific enough to defend, and a specific claim is one that can be tested.

Pace matters too. Shipping fast trains taste because each release produces feedback that tests the judgment behind it. Perfectionism slows that loop and leaves judgments untested for longer, and an untested judgment stays an opinion.

The specification as the unit of taste

These habits gain weight as AI takes over execution. When a system can build what it is told, the instructions become the ceiling of the product. Prompt quality becomes product quality: the precision of a specification shows exactly how clearly its author understands what should exist.

The new job is writing specifications so precise that AI has nothing left to improve on the author's behalf. Alongside that sits a second judgment: deciding when AI output is good enough and when it falls short of what the product requires.

Both judgments run inside the taste loop, a repeating cycle of human judgment, AI execution and human curation. Execution in that loop grows cheaper and faster with each generation of tools. The judgment at either end of it remains the human contribution, which is why product taste works as the last defensible moat.

Levels and traps

Taste operates at different altitudes, and each one is a distinct skill. Feature taste governs a single interaction. Product taste governs the coherence of the whole. Company taste governs what an organization chooses to build at all, and founder taste, often studied through figures such as Jobs, Ive and Chesky, sets the standard others work toward.

Within teams, taste propagates or fades depending on whether it is made explicit. A room where nobody holds it becomes a taste vacuum, where decisions settle on whatever is easiest to agree on.

That tendency points to the most common failures. Consensus taste settles on the option nobody objects to. Trend-chasing builds for the moment at the expense of permanence.

Feature creep adds until clarity disappears. Treating user feedback as taste confuses a valuable input with the judgment that weighs it. The power-user trap optimizes for sophistication at the cost of clarity for everyone else.

Each trap is legible, which means each can be noticed and corrected.

Judgment as the durable skill

Taken together, these ideas describe product taste as a discipline with a full structure: a definition, observable signals, measurable axes, methods of development, a specific role in AI-assisted work, distinct levels and well-mapped failure modes. Each part can be studied and practiced on its own.

The practical consequence is a shift in where product people invest their effort. As execution becomes abundant, the scarce skill is knowing what should exist, why it matters and how it should feel, and stating it precisely enough for anyone, human or machine, to build. That skill compounds with every judgment made, tested and refined.