Models Are Commodity. Taste Is the Moat.

There's a paper going viral this month from the University of Melbourne. The headline finding: a frontier model with the full procedure stuffed into its system prompt outperforms the same model orchestrated through LangGraph on every domain, every metric.

The agent-framework crowd is in defense mode. Half my timeline is declaring agents dead. The other half is declaring engineers dead.

Both are wrong, but the conversation reveals something more interesting: the real bottleneck in AI products isn't the model, the framework, or the orchestrator. It's the human pointing the model at the right problem in the first place. And that human is, increasingly, not an engineer.

The two camps

I see roughly 100 AI startup pitches a year as a GP at Hello World Investments. Every one of them eventually reveals which camp the founder is in.

Camp 1 says: "we have better models." Better fine-tuning. Better RAG. Better retrieval. A clever new agent harness. A novel use of a specific frontier capability.

Camp 2 says: "we have better judgment about what models should do." They talk about which 1% of failure modes will gut user trust. They talk about the difference between a clinically correct answer and an emotionally correct one. They show you the features they killed, not the ones they shipped.

Camp 1 will lose.

I'm not being contrarian for the sake of it. I've watched four Camp 1 startups in the last eight months get reduced to features by a single OpenAI release. The model layer is moving too fast to defend. Every moat someone draws at that level is one product update away from being filled in.

Camp 2 doesn't have this problem. Because the moat isn't a technical artifact — it's a sequence of thousands of small, opinionated decisions about what a real human, in a real moment, actually needs from this product. Models commoditize. Decisions about humans don't.

What taste actually means in AI products

"Taste" sounds soft. It isn't. In an AI product, taste is the specific skill of knowing:

1. Which 1% of failure modes will destroy user trust. A health app that's wrong about your step count once is annoying. A health app that's wrong about your bloodwork once is dead. Engineers optimize the 99%. Taste tells you which 1% to obsess over.
2. When the right answer is no answer. Most AI products are bad because they answer everything. A scared user asking a health question doesn't always need information — sometimes they need a path to a real human. Knowing when to defer, escalate, or stay silent is a product decision that lives nowhere in the model.
3. What to leave out. Every feature you ship has an opportunity cost in attention. The hardest skill in AI product work is killing features that work. Not because they fail — because they're not what the product is about. Producers learn this on every album: a great song that doesn't fit the record gets cut.
4. Tone before logic. I've shipped Healify features this year where the model logic was correct and the tone of the response was wrong. Users don't separate those two things. They experience them as one. A clinically perfect answer in the wrong voice is a wrong answer. This is the single biggest thing engineers underweight and creatives can't unsee.
5. When to break your own rules. Every AI product needs a system prompt. Every system prompt has assumptions. The art is knowing which user moments break those assumptions and need a different response entirely. That's a judgment call that no benchmark will ever measure.

None of these are engineering skills. They're creative skills disguised as product skills.

Three Healify examples

Healify is an AI-powered health platform built on LangGraph agents. We're 18 months into shipping it in production. Here are three taste decisions that mattered more than any model choice:

We killed a feature that worked. Our nutrition agent could generate detailed meal plans from your bloodwork results. Technically impressive. We shipped it, ran it for a month, and killed it. Why: it took the most emotional moment in the app (seeing your bloodwork) and immediately turned it into a logistics task. Users didn't want efficiency in that moment — they wanted to be seen. We replaced it with a single line of plain-language reassurance and the meal plan moved to a separate flow. Engagement went up.

We rewrote the system prompt four times in a week. Not because the logic was wrong — because the agent kept sounding like a therapist with no skin in the game. Users could feel the corporate tone the moment something serious came up. The fourth version stripped out every hedge, every "consult your doctor," every safety platitude that wasn't actually doing safety work. Trust scores went up. The clinical content didn't change at all.

We let the designer debug the agents. Last week we hit a bug three engineers couldn't see. A designer spotted it in thirty seconds — the agent's tone was wrong before its logic was. The model was returning correct information in a voice that broke the product's whole vibe. This is now a process: every agent change goes through tone review before it ships. Nobody on a traditional team would have written this into the workflow.

These aren't decisions a model can make. They're decisions a creative would have made instinctively, and an engineer would have benchmarked into the ground.

Why creatives have the unfair advantage

If you spent ten years making music, writing screenplays, designing brands, directing films, building games — you trained one specific muscle that almost nothing else trains: the muscle of *"is this good?"*

You did it under deadline. You did it with feedback that wasn't always logical. You did it knowing that a track that's 99% right and 1% off is still a track that fails. You learned to feel the difference between technically correct and actually good. You learned that the user — the listener, the viewer, the player — is always right about how they feel, even when they can't articulate why.

That muscle is what AI products run on now.

The model gives you infinite drafts. The founder picks the right one. If your entire career was building the draft instead of judging it — if you were a senior engineer or a quant or a researcher who shipped what the spec asked for — that's a hole in your kit no amount of compute will fill.

You can hire engineering. You can rent compute. You cannot hire taste. Taste is the residue of a thousand decisions made under pressure with no right answer. It's path-dependent. And it's exactly what creative careers manufacture at scale.

This is why I keep seeing the same pattern as a GP: the AI founders winning aren't the ones with the most impressive technical backgrounds. They're the ones who spent ten years in some other room learning to feel what's good before they wrote a line of code.

If you're a creative sitting on an AI idea

Three things.

One: stop saying you're not technical enough. The barrier to building software collapsed in 2024. Cursor, Claude Code, the agent stack — these tools were specifically designed for people whose best skill isn't writing code. Your job isn't to out-engineer engineers. Your job is to bring the judgment they can't.

Two: the moat is in your hand already. Every "is this good?" you've ever asked while making something else is preparation for this. Don't translate that instinct into engineering language and lose half of it in the process. Build the product the way you'd produce a track or direct a scene — with opinions about tone, omission, and what the audience is actually feeling.

Three: the next 24 months are the window. The current model layer favors people who can sense what to build for a real human. The next layer might favor different skills again. Right now, the unfair advantage you have is widest. Use it.

If you want the exact questions I run every Healify feature through — packaged as a 6-page playbook with three real Healify examples for each — grab the free PDF.

If you want a focused hour to pressure-test the idea with someone who's made the leap, I have a consult slot open most weeks. The link's on the homepage.

The next great AI founders won't be engineers. They'll be the people who already know what good feels like — and now finally have the tools to ship it.