Taste as Moat

In 2023, building a working internal tool took a senior engineer a week. In 2026, with Cursor driving Claude Opus 4.7, it takes an afternoon. The bottleneck stopped being typing speed. It became the question: of the forty things we could ship this week, which two are actually worth shipping? That question has no syntax, no test suite, no model that scores it. It is taste, and it is now the rate-limiter on output.

What taste actually is

Taste, operationally, is the ability to predict — fast and without explicit reasoning — which of many possible artifacts will hold up under contact with reality. A buyer with taste looks at 200 fabric swatches and pulls eight that will sell. A producer with taste hears 40 takes and keeps two. A PM with taste reads a backlog of 60 tickets and kills 50. The pattern is the same: the work is mostly subtraction, and the subtraction is non-verbal.

Rick Rubin built a four-decade career as a producer largely by saying "no" more than anyone else in the room. He doesn't play instruments. He doesn't engineer. He sits on the couch and says that one, or not that one, until what remains is the album. His 2023 book The Creative Act is essentially 400 pages refusing to explain the mechanism, because the mechanism is the refusal.

Why this is the AI-era operator skill

Three forces converged in 2024-2026 to make taste the binding constraint:

Era	Bottleneck	What scaled the operator
Pre-2010	Capital, distribution	Access to factories, shelves, ad budgets
2010-2022	Engineering throughput	Hiring, frameworks, dev tooling
2023-now	Deciding what to make	Taste, judgment, ruthless prioritization

When the marginal cost of producing a thing falls toward zero, the marginal value of not producing the wrong thing rises sharply. An engineer in 2020 shipping a bad feature lost a week of their own time. An engineer in 2026 shipping a bad feature using AI loses a week of their team's onboarding, support, and rollback. The cost of shipping the wrong thing went up even as the cost of shipping went down.

The two diamonds are where humans still matter. Everything else is automatable.

Where it shows up

Procurement and merchandising. A buyer who can hold 30 trends in working memory and reject 28 outperforms one who places safe bets on all 30. The concept quick response retail playbook depends entirely on the buyer's confidence to commit small and edit fast.
Investment. Charlie Munger's "sit on your ass" investing: most of Berkshire's compounding came from a few decisions, not many. The taste was knowing which businesses to ignore for decades.
Software products. Steve Jobs killing the Newton, the original iPod's 1,000-song promise as the only feature, the iPhone shipping without copy-paste. Each was a "no" that, in retrospect, was the product.
Music. George Martin trimming "A Day in the Life" to its final form. Brian Eno's Oblique Strategies card deck is a tool for generating taste under deadline pressure.
AI tooling. A team running concept claude code as operator leverage without taste produces 10× more mediocre code. With taste, they produce 10× more decisive code. Same tool, opposite outcomes.

How taste is built — the part nobody likes

The honest path is volume plus feedback. Rubin listened to roughly 10,000 hours of music before producing his first album at Def Jam. A buyer develops their eye by sitting through 50 seasons of sell-through reports. Karpathy's intuition for what would scale in transformer training came from training hundreds of variants by hand at OpenAI before GPT-2.

There is no shortcut, but there is a multiplier: deliberate exposure to the very best work in your domain, with feedback on your guesses before you see the outcome. The form is: look at the thing, predict its fate, then check. Repeat for years. This is the loop concept deliberate practice describes, applied to judgment instead of motor skills.

The trap most senior people fall into: they stop predicting and start asserting. The feedback loop dies. Taste calcifies into preference, which is taste without learning. A producer who hasn't been wrong in five years has stopped having taste — they have a brand.

What's contested

Whether taste is transferable across domains is genuinely open. Rubin produced rap, rock, country, and metal at the top of each — suggesting the meta-skill is general. But Jony Ive's taste in industrial design didn't transfer cleanly to Apple Watch software, and Elon Musk's taste in rockets has not transferred to social platforms. The evidence cuts both ways. The honest position: taste is partially transferable, but the transfer rate depends on how much the domains share underlying structure (rhythm, restraint, signal-to-noise), and most operators overestimate the transfer.

Also contested: whether AI itself will eventually have taste. Today (May 2026) no model reliably tells you which of two product specs will succeed in market. Models can produce — they cannot yet refuse well. That gap is the moat. Whether the gap closes is the bet every founder is implicitly making.

Why this has to do with other realms

The Inditex playbook (see concept quick response retail) is taste expressed as supply chain: Zara's edge is not making clothes fast, it is the discipline of not committing to a season until the buyer sees what is selling. The fabric is the substrate, the no is the product. Similarly, Hermès' Birkin economics in concept scarcity as strategy rests entirely on the house's willingness to say no to revenue — to leave money on the table because the brand depends on the refusal. Across fashion, music, software, and finance, the operator who compounds is the one who treats subtraction as the act, not the prelude.

An open question

If taste is built by volume plus feedback, and AI tools collapse the volume side of the equation, do junior operators in 2030 develop taste faster (more shots on goal) or slower (no longer paying the cost of bad shots)?

Key sources

The Creative Act by Rick Rubin (2023) — the producer-as-curator argument, written in 78 short chapters that themselves model subtraction.
Poor Charlie's Almanack by Charlie Munger (2005) — taste as the discipline of staying inside your circle of competence and saying no to almost everything.
Steve Jobs by Walter Isaacson (2011) — the Apple case study for taste as product strategy, especially chapters on the original Macintosh and the iPhone.
to verify: the Andrej Karpathy talks (2023-2024) on what scaled at OpenAI — his account of taste developed by hand-training hundreds of model variants.
Brian Eno's Oblique Strategies (1975, with Peter Schmidt) — taste as a deck of constraints, the first published "tool for taste."

Abhishek's take

The thing nobody told me when I started leaning hard on Claude in 2024 was that my throughput problem would disappear and a harder problem would replace it — a queue of forty viable ideas every Monday, and the same five working days. The constraint moved from can we build it to should we. I now spend more time killing well-formed plans than writing new ones, and that subtraction is the work that actually compounds. Rubin had it right: the person in the room saying no the most, with the most discrimination, is the one shaping the output.

Where I've used this

Running an AI-augmented buying floor means I generate ten times more analysis than I did three years ago. The discipline that actually moved numbers wasn't producing more — it was the weekly review where I delete most of it, keep three insights, and act only on those. The pipeline got cheaper; the taste got more valuable.

Abhishek