You Run Agents All Night. Then Get Slop.
How ‘taste’ became the last scarce creative resource.
AI can now write code, produce entire music albums, and design products overnight. The access is real. The speed is real.
And the slop is very, very real.
Scroll through any app marketplace, music playlist, or design portfolio today and you’ll find the same thing: technically competent output that feels hollow. Correct but somehow lifeless. Finished but never actually done.
Here’s what’s happening. The cost of execution has dropped to near zero. You can spin up a feature, a logo, a sample pack, or an entire codebase with a handful of prompts. But the thing that separates a product people love from a product nobody notices? That’s still the same as it’s always been.
Taste.
This is the AI taste moat. And two creators explained it better than any researcher or analyst has: Peter Steinberger, the indie developer behind OpenClaw, on the builder’s side; and Diplo, one of the most commercially successful music producers of the past two decades, on the creator’s side.
Their message is the same even though their worlds are completely different. Execution is now free. Taste is the last scarce resource.
Diplo put his entire career into a single sentence: “I’m in a position now where I can use these tools really effectively because people already trust me for my taste and what I do.”
Steinberger named the failure mode builders keep running into: agents that run all night and produce “the ultimate slop” because nobody with taste stayed in the loop.
Two industries. One thesis.
If you’re still optimising prompts instead of curating taste, you’re building on the wrong side of the moat.
Steinberger’s Warning: The Agentic Trap
Peter Steinberger didn’t mince words.
“You can create code and run all night and then you have like the ultimate slop because what those agents don’t really do yet is have taste.”
That line landed hard across developer circles for a reason. It names something builders have been quietly noticing but haven’t wanted to say out loud. You can automate the output. You cannot automate the compass.
Steinberger went further:
“They are spiky smart and they’re really good at things, but if you don’t navigate them well, if you don’t have a vision of what you’re going to build, it’s still going to be slop. If you don’t ask the right questions, it’s still going to be slop.”
The agentic trap is what happens when you remove yourself from the process too early. You front-load a perfect spec, hand it to an agent, and walk away expecting something coherent to emerge. What you get instead is highly competent output with no internal compass.
Steinberger describes his own process differently. He starts rough. He plays with the output. He lets his next step emerge from what he sees and feels in the current state of the project.
“My next prompt depends on what I see and feel and think about the current state of the project. But if you try to put everything into a spec up front, you miss this kind of human-machine loop. And then I don’t know how something good can come out without having feelings in the loop — almost like taste.”
This is the real insight. Not that agents are bad. Not that AI-assisted development is a mistake. It’s that the feedback loop between human taste and machine output is where the actual quality lives.
Think about it in concrete terms. A solo founder who iterates a feature 40 times in a single evening, each time responding viscerally to what feels off, ships something people actually use. The team that lets agents run unsupervised for 12 hours ships something polished and forgettable.
Agent scaffolding is improving fast. Self-critique loops are getting smarter. But the human taste anchor is still 100% required.
Prompting is the new typing. Taste is the new thinking.
Diplo’s Take on References, History, and the Pop Factory
Diplo is not anti-AI. He made that very clear.
“You’re not gonna win. There’s no fighting AI. You have to just work your best to be the best at it right now. You’re wasting your time.”
He told a story about a young producer he had recently signed. They sat down together, fired up Udio, and made three goth-style beats in five minutes. The producer looked at him and said: “I’m cooked, right?”
Diplo’s answer was honest. “I don’t know, man. Maybe.”
But then he explained his own position. People already trust him for his taste. That trust is the product. The AI is just the fastest production line he has ever had access to.
His edge is not his software. It is not his prompt library. It is two decades of deep cultural fluency, from 1973 Jamaican dub to 2006 Baltimore club to 2008 blog-house and beyond.
It is the references. It is the history. It is the accumulated signal that lets him know, in five seconds, whether something is worth pursuing or not.
The tweet summarising his interview captured it cleanly: “Taste and references matter A LOT. That is what makes me good at using AI because I know how to prompt something specifically. You have to have knowledge. You have to have history.”
Most people pressing the same buttons do not have that layer. They work from descriptions. Diplo works from memory, intuition, and a reference library that took years to build.
And that gap only grows wider as the tools get better.
The 99% with access to the same AI tools still produce “good enough.” The 1% with genuine taste and deep references produce something that feels human-directed, alive, and worth paying for.
Your reference library is now more valuable than your prompt library.
Why References Beat Prompts: The New Moat Explained
There is a simple way to understand why taste has become the moat.
Prompts are tactical instructions. They describe what you want. AI gets better at executing on tactical descriptions every month. By the time you have mastered a prompting technique, a newer model has made that technique mostly obsolete anyway. Prompting is a skill, but it is a depreciating one.
References and taste are something different. They are pattern recognition developed over years of deep attention to what works and why. They are emotional intelligence about what a piece of work needs to become. They are personal weirdness, accumulated preferences, and the willingness to trust your gut over the consensus.
Think of it this way. Prompts are the steering wheel. Taste is the map, the destination, and the reason you are driving in the first place.
The evidence is already visible. App stores are saturated with AI-generated productivity tools that are functionally identical and emotionally indistinguishable. Music streaming platforms are drowning in AI-assisted tracks that technically pass quality thresholds and generate no emotional response. The content that breaks through is increasingly the content that feels like a specific person made a specific decision about what it should be.
Compare the eras. In the 2010s, attention was the moat. Whoever captured and held attention could build a business around it. In the early 2020s, distribution was the moat. Whoever could reach an audience at scale had leverage. In 2026 and beyond, taste is the moat.
The counterpoint is real: models are getting better at simulating taste. They can mimic aesthetic preferences at a surface level. But simulation is not origination. The human who has actually spent years absorbing references, developing opinions, and learning to feel the difference between what is good and what is right still wins. Because they know what the simulation got wrong.
The premium product will always be the one that feels human-directed, even when 90% of the labour is AI.
A Practical Framework for Building Your Taste Moat
The good news: taste is a skill, not a gift. It grows with deliberate practice.
Start with a taste vault
Collect 50 to 100 specific examples of work you love across mediums. Not work that is considered good. Work that actually moves you, surprises you, makes you want to understand how it was made. Pull from this vault when you are prompting instead of describing from scratch. “Make something that feels like this specific thing” beats “make something good” every single time.
Use the Steinberger loop, not the hands-off spec
Do not hand agents a spec and disappear. Set 15 to 30 minute checkpoints. Stop. Interact with the live output. What feels wrong? What wants to become something? Let your next move come from that gut reaction, not from a document you wrote before you saw anything.
Dedicate time to reference building
One hour a week. Pick one artist, era, movement, or product category. Go deep, not broad. Write down what you notice. What surprises you. What you would steal. Share it publicly. Teaching what you find is one of the fastest ways to make it stick.
The one honest question
Before each prompt, ask yourself: do I actually know what this should feel like, or am I describing it from the outside? If you are describing, stop. Find a specific reference first.
So, Are You Cooked?
Back to Diplo’s young producer.
The honest answer is: it depends entirely on what he does next.
If he treats AI as a shortcut to output, he will produce the same averaged content as everyone else using the same tools. If he spends the next two years deep in references, building his palate, understanding not just what sounds he likes but why they work and where they came from, he builds something the tools cannot replicate.
That is the actual play.
AI democratises execution. It makes taste more valuable. The creators and builders who understand this early will build moats that take years to cross.
Peter Steinberger is building software with this approach. Diplo is making music with it.
The tools are not the advantage. The taste behind them is.
Share your own taste moat experiments in the comments or on X. The original interviews are worth your time: How OpenClaw’s Creator Uses AI to Run His Life in 40 Minutes | Peter Steinberger and Why Diplo’s Weirdest Ideas Became the Biggest Songs in the World (Full Interview).
Taste is not going away. It is finally getting the economic value it always deserved.