Product Field Work All writing
WritingDesign and experience

Taste is the moat

When anyone can generate a competent interface in minutes, competent stops being impressive. Customers can smell a clone, a bolted-on feature and lazy AI output almost instantly. What they notice, and remember, is a product that feels like one coherent thing, made by people who decided what to leave out.

The problem

Ask an AI for a “modern SaaS app” with no direction and you get the average of every SaaS app it has seen. The result is fine. It’s also indistinguishable from thousands of others, and people have learned to spot it.

The deeper problem is inside the product. When each new feature costs an afternoon, it’s easy to keep adding them, each reasonable on its own, until the product becomes a drawer of features. But users, especially in enterprise tools, don’t think in features. They have a job to do. They want to get in, get it done, and get out. Every feature that doesn’t serve that job is something they have to look past.

1995Hitachi researchers testing 26 ATM layouts found people rate better-looking interfaces as easier to use: the aesthetic-usability effect
log₂(n)Hick’s law: decision time grows with every added choice. Each feature on a screen is a small tax on everyone who didn’t need it
“Nobody decided”the message generic AI design sends: the same fonts, gradients and rounded cards on thousands of products
How I’d measure taste, as far as it can be measured
  1. Steps and choices in each core job, which should fall over time
  2. Time to get in and out for the top five jobs
  3. Consistency violations found by automated review against the design skills
  4. Features removed or merged, not just added
Why I wrote it

I’ve built products where every feature was defensible and the whole was still confusing. With agents, that failure can happen ten times faster. It’s also the first time in software history that there’s truly no excuse for anything less than the best-designed, most usable version of a product.

The ideaDesign journeys, not features. Make every addition earn its place by what it improves and what it removes. Encode the team’s taste, its principles, system, journeys and anti-patterns, into the skills every agent builds with. And use unlimited refinement to raise the bar, with a person owning the final call.

1The average, and a decision

Here is what “no decision” looks like next to a product where someone chose every detail for a specific customer. Turn on the markers to see the fingerprint.

Same brief, two results

Brief: “A tool for coordinators to manage today’s requests.” Left: the statistical average. Right: designed for the job.

✨ Now with AI
Supercharge your workflow

The all-in-one platform to manage everything, powered by AI.

Get started free →
⚡Lightning fast
🤖AI-powered
🔒Secure by default

The average: competent, generic, about nothing in particular

RequestsTue 14 Oct · North region
Today · 23 open
4 due before noon · 2 waiting on a customer
Replace badge reader, Building C9:30
Access for new contractor11:00
Update visitor list, lobby13:00
Quarterly key auditFri
Parking permits, marketing teamFri
Start nextAssignUpdate 3 selected

A decision: the job is the interface, and nothing competes with it

1Purple-to-indigo gradient on the logo, headline and button
2A “Now with AI” badge that says nothing about the job
3A headline that could describe any product
4Three identical rounded cards with emoji icons
5The same default font as everyone else
6A marketing page where the user wanted their work

2Journeys, not features

A coordinator doesn’t open the product to use a feature. She opens it to close out today’s requests before lunch. Every feature is judged by one question: does it make that trip shorter and clearer, or does it add a detour?

One job, two products

Job: close out this morning’s five requests. The same capabilities, bolted on one by one versus woven into the journey. Illustrative.

Users don’t remember features. They remember whether the tool got out of their way.

3The subtraction test

Every addition should answer the same questions before it ships, and one of them is always “what does this remove?” A feature that serves the job and takes something away is a win. A feature that serves no job, no matter how clever, is a cost everyone pays.

Run a proposal through the test

Pick a proposed change to the coordinator’s list screen. The deltas are my estimates of the effect on the core job.

4Taste, encoded

If agents build most of the screens, the team’s taste has to live somewhere they can read it. Not as adjectives, which the model averages, but as specific decisions: the system, the journeys and their step budgets, the voice, the anti-patterns to avoid, and examples of good and bad. Every correction a designer makes more than once belongs in the skill.

This is where designers’ work moves. Instead of drawing each screen, they author and tend the taste that every screen is built from.

What goes into a design skill

A sketch of one team’s design skill. Pick a file to see what it holds.

5No excuse for anything less

Refinement used to be rationed. Polish was the first thing cut when a deadline got close. Now an agent can critique a screen against the team’s rubric, fix what it finds, run it past the customer panel, and do it again, many times, before a person looks at it. The bar isn’t “good enough to ship.” It’s “the best version anyone could make.”

A screen, pass by pass

Illustrative scores against a five-part rubric: consistency, hierarchy, clarity, accessibility and craft. Press play.

A person owns the final callRubrics catch errors. Only someone with taste can say “this is right,” or “this is technically fine and still wrong.”
Designers tend the skillsEvery repeated correction becomes a rule, an example or an anti-pattern, so the next screen starts where the last one ended.
A weekly journey reviewWalk the top jobs end to end, as a customer would. Features are reviewed in the journey, never alone.
Steps per core jobFor the top five jobs, tracked release to release
Choices per screenVisible actions a user has to consider
In and out timeHow long the most common task takes end to end
Consistency violationsFound by automated review against the skills
SubtractionsFeatures removed, merged or hidden by default
Repeat correctionsDesign feedback given more than once, which should approach zero

Choices

Decisions, not adjectives“Modern and clean” produces the average. Specific choices, written down with reasons, produce a product.
Every addition names a subtractionIf a feature can’t say what it makes shorter, simpler or unnecessary, it waits.
Journeys are the unit of designFeatures are reviewed inside the job they serve, with a step budget that doesn’t grow.
Where this can go wrong

Taste can become one person’s preference dressed up as principle, which is why it has to be written down, explained and open to challenge. Distinctive isn’t the same as unusual: in enterprise software, familiar patterns help people move fast, and breaking them needs a reason. Beauty can’t rescue a broken workflow; attractive design makes people forgive small problems, not big ones. And accessibility isn’t a style choice. It belongs in the rubric from the first pass.

Sources

  1. Noqta, June 2026: the “AI slop” look and why defaults compound into sameness.
  2. Superdesign, 2026: the average look, and the fix of a reusable design spec rather than a better prompt.
  3. AI Engineer World’s Fair 2026, “The missing layer: design taste in AI agents”: save repeated design corrections into a skill.
  4. Shopify, UX laws: Kurosu and Kashimura’s 1995 aesthetic-usability study, and Hick’s law.
  5. Jon Yablonski, Laws of UX: Hick’s law and removing what doesn’t serve the user’s goal.
  6. Anthropic, Skill authoring best practices: specific, tested skills that explain the reasoning behind each rule.
  7. “Defeating Feature Fatigue,” HBR 2006: more features win the sale and lose the customer.