Taste as a Moat

Taste has always been the separator. AI does not create that moat; it makes it more visible by making average execution cheap.

Source Artifacts

The Thesis

Taste is the capacity to make high-quality qualitative judgments when there is no clean objective metric. It is the judgment that separates an artist from a master, a careful designer from a generic one, and Linus Torvalds from a programmer who merely knows C syntax. That separator is not new. The tools change; the scarce judgment that decides what is worth making and what "right" feels like remains. Source: User, 2026-06-29; X/@mitchellh article signal, 2026-06-27

AI changes the economics, not the ontology. When a model can generate ten plausible layouts, ten React implementations, or ten copy variants in seconds, the bottleneck stops being production. The bottleneck becomes selection, constraint, omission, and repair. A person or agent with no real design understanding can produce acceptable surfaces by imitation, but not exceptional work except by close copying. Exceptional work requires a model of the domain: what matters, what can be removed, where the viewer's eye lands, which abstraction will age well, and which compromise will rot. Source: User, 2026-06-29

The taste.ai investor-signal bookmark sharpens the market pressure: investors want taste to become automatable because then average generated work becomes more valuable. That does not make taste disappear; it moves the contest to whether the product has a real judgment system or only a nicer generator. If automation merely makes slop harder to distinguish from competent work, it raises the floor while leaving the moat with the people and systems that can still tell what should exist. Source: X/@VitusDoesThings, 2026-06-17

This is why "AI raises the floor" is not a counterargument to taste. Raising the floor increases the premium on the ceiling. More people can now produce average work, so the gap between average and genuinely good becomes more important, not less. Builder Ethos says AI collapses the engineering barrier; Taste Is the Differentiator (on Shared Primitives) names the consequence: the shared primitive is no longer the moat, the opinionated layer is. Source: Builder Ethos, 2026-04-14; Taste Is the Differentiator (on Shared Primitives), 2026-06-22

The Linus Reconciliation

The correct AI stance is not anti-tool romanticism. Torvalds's 2026 Open Source Summit position is the useful middle: AI is a high-productivity tool, analogous to earlier jumps from machine code to assemblers, high-level languages, and compilers. The New Stack reported his argument that AI can boost programmer productivity while serious systems still require human understanding of the generated code and architecture. Source: The New Stack, 2026-05-29

That maps cleanly onto Kevin's view. Natural language is the next interface in the long movement away from raw machine representation and toward human intent: machine code -> assembly -> higher-level languages -> compilers -> IDEs -> prompts and agents. But each step abstracts mechanics; it does not remove judgment. A compiler lets you stop hand-writing machine code, but it does not know whether your API boundary is wrong. An LLM lets you stop hand-writing every line, but it does not know what product should exist unless the surrounding system gives it taste, context, and evaluation. Source: User, 2026-06-29; The New Stack, 2026-05-29

So AI makes taste a stronger separator. The people who understand systems, design, and product can express intent faster. The people who lack that understanding can also prompt tools, but they are mostly accelerating into incoherence. The prompt is not the work. The work is the taste-bearing system around the prompt: the brief, the constraints, the examples, the evals, the review, and the final cut. Source: User, 2026-06-29

Where Kevin Locates Taste

Distribution Is the Moat lists brand and taste as one of the things that set you apart when product converges: the version of the product that feels considered wins. Source: Distribution Is the Moat, 2026-04-18 Career Profile frames Kevin's own edge as "taste and range," and treats the portfolio site as proof that he is an engineer with taste, not a generic SWE. Source: Career Profile

Sigil UI is the clearest artifact of this stance: an opinionated, token-first design language built so agents author UI without producing generic AI aesthetics. The Greatness Trap records the distinction precisely. Sigil is not a bid for greatness, it is an expression of a strong opinion about how design should work. The opinion is the taste, and the taste is what makes the thing defensible. Source: The Greatness Trap, 2026-05-17

The Mechanism

Taste is hard to copy because it is not the visible artifact. It is the decision function that produced the artifact. You can copy an interface; you cannot copy the thousands of exposures and corrections that made the designer know why the copied version works. You can copy Linus-style C; you cannot copy the systems judgment that decides which special case should disappear. Source: User, 2026-06-29; Taste Is the Differentiator (on Shared Primitives)

The reviewed EXM visual names the missing operating layer: most teams have a system of record for the final artifact, but not a system of judgment for the edits that made it good. The dense signal is the diff between first draft and final, especially when the stakes are high and a senior reviewer marks up live client work. That is why taste cannot be learned only from finished deliverables: the conclusion hides the choices. Kevin's agent stack should preserve edits, preferences, evals, and code review comments as training material for taste, not only the shipped page. Source: X/@EXM7777 and local image review, 2026-07-04

Operating Rules

• Use AI aggressively, but never outsource judgment.
• Treat taste as an eval: write down what "good" means before asking agents to produce it.
• Inspect outputs for coherence, not just correctness.
• Prefer constraints over vibes. If the model keeps producing slop, the missing object is probably a taste-bearing rule or example.
• Do not confuse copying with taste. Close imitation may produce a good artifact once; it does not create a repeatable judgment loop.

Failure Modes

• Anti-AI craft nostalgia: rejecting the tool because older craft rituals felt meaningful.
• Prompt cargo culting: assuming natural language intent replaces domain understanding.
• Surface copying: reproducing an artifact without inheriting the judgment that made it coherent.
• Average worship: accepting model output because it is competent, even when it has no point of view.
• Unwritten evals: claiming to have taste while never codifying the bar into examples, constraints, rubrics, or review loops.

Connections

---

Timeline

• 2026-07-04 | Added the taste.ai investor-signal row as market evidence for the thesis: automating taste is valuable only if the system preserves judgment, not merely if it makes average AI work harder to distinguish from real work. Source: X/@VitusDoesThings, 2026-06-17
• 2026-07-04 | Added the "system of judgment" artifact: taste is transmitted through reviewed diffs and live markups more densely than through final deliverables, so agent workflows should preserve edits, preferences, and evals as taste data. Source: X/@EXM7777 and local image review, 2026-07-04
• 2026-05-31 | Compiled from the taste threads across Builder Ethos (taste is what remains in the Golden Age), Distribution Is the Moat (considered products win), The Greatness Trap (opinionated building via Sigil UI), and Career Profile (taste and range as Kevin's differentiator). Source: compiled from wiki graph
• 2026-06-22 | Generalized into Taste Is the Differentiator (on Shared Primitives) — the mechanism (commoditized primitives → cheap execution → judgment is the scarce input) with the "agents are design systems / shadcn presets are workers" framing and a cross-field table (creative, frontend, AI, distribution, agents). Source: User, 2026-06-22
• 2026-06-29 | Overhauled with Kevin's stronger thesis: taste is the historical moat across art, programming, design, product, and agents; AI is the natural next tool layer, but it makes understanding and evaluation more important rather than less. Added the Mitchell Hashimoto taste article signal and Linus Torvalds AI-as-tool framing. Source: User, 2026-06-29; X/@mitchellh, 2026-06-27; The New Stack, 2026-05-29