Taste engineering is the next skill after prompts, context and loops
TL;DR
Prompting got us an answer. Context engineering got the model to answer like it actually knew our business. Intent engineering made us specific enough to get what we meant instead of what we literally said. Loops let the agent check its own draft against that intent and keep going until it landed. Every one of those stages solved a version of the same problem: execution. None of them solved the harder problem, deciding whether the finished thing is actually good. That call is taste, and it is where the frontier is moving now.
What changed once execution stopped being the bottleneck?
Three years ago the bottleneck was getting a model to produce anything useful at all. We solved that in stages.
Prompt engineering got us a usable answer if we asked well enough. Most people are still stuck here, collecting prompt templates like recipes. Context engineering came next: pulling the prompt's raw material out of the prompt itself and into permanent memory, files, and file structure the agent could reach on its own. At AIOS this is the step where we prime the system with a client's context, or our own, before it does anything. It is genuinely useful. Most disappointing AI output is still a context problem wearing an intelligence costume.
Then we hit the next wall. Full context, sharp prompt, and the output still missed, because we were vague about what we actually wanted and let the model guess. So we got specific: intent engineering, naming the exact outcome instead of hoping it would be inferred. Once intent was sharp enough to be judged, we could loop it, have the agent check its own draft against that stated intent and keep refining until it actually matched, instead of stopping at the first attempt.
Four stages, one thread running through all of them. Each stage made the agent better at building the thing. None of them told us whether the thing was worth building, or whether the version it landed on was the right one.
What is taste engineering?
Taste engineering is the discipline of applying human judgment, elegance and restraint to what an agent produces, and knowing when to accept it, reject it or send it back. It is not a prompting technique and it is not a loop condition. It sits above both.
A loop can check whether an output hits the criteria you gave it. Only a person can look at the result and recognise whether it is actually good, the way a good editor recognises a strong sentence or a good architect recognises a room that works. Call it elegance, restraint, editorial judgment, whatever word fits your trade. The instinct that looks at ten competent options and picks the one that is genuinely right is taste.
See where AI fits in your business. Free.
A 45-minute audit. We map the highest-value automations and what they're worth in time and money. No pitch, no pressure.
Why does taste matter more once agents can build almost anything?
Because the constraint has moved. When execution was hard, a mediocre idea, well built, still stood out. Now an agent can execute almost any idea competently, so competence stops being the differentiator. Everyone has access to the same models, the same context engineering, the same looping. What nobody can hand you off the shelf is the judgment to know which output deserves to ship.
This is art, this is not art. That call has always been human, and it is becoming the whole game.
A business that automates its workflows and outsources its taste ends up with a hundred outputs that are all fine and none that are memorable. The ones that pull ahead are the ones where a person with real judgment is still standing at the gate.
What does taste actually look like in practice?
Less dramatic than the word suggests. An agent drafts ten variations of a landing page headline. All ten are grammatically sound, on brief, technically correct. Taste is the person who reads all ten and picks the one that will make someone stop scrolling, and can say why the other nine, though correct, were flat.
It looks like reviewing a workflow automation before it goes live and asking not "does this work" but "does this feel right for how this business actually operates, or did we just build what was easiest to automate." It looks like a client report an agent assembled perfectly to spec, that a person still reads end to end before it goes out, because spec and judgment are not the same test.
At AIOS, that is the human approval gate we hold before anything client-facing or irreversible ships. Not because the agent got a fact wrong. Because taste is a different check than correctness, and no amount of looping replaces it.
Can taste be taught, or is it just instinct?
Some of it is instinct, and some people will always have a sharper eye than others. But most of what reads as natural instinct is pattern recognition built from looking at a lot of finished work, good and bad, and noticing what actually separates the two. That part can be built on purpose.
Look at more finished work in your field than you currently do. Get specific about naming why something works instead of just feeling that it does. Keep a running list, mental or written, of the small choices that separate what you would ship from what you would not. The agent will keep getting better at execution. It will not develop your taste for you. That has to stay a human muscle, exercised deliberately.
Is taste engineering the last stop?
Probably not. Prompting looked like the whole story until context engineering exposed what it was missing. Context looked complete until intent exposed how vague we still were underneath it. Intent looked like the answer until loops showed what specificity could do once the agent kept checking its own work. Taste engineering is where the frontier sits today. It will not be the final word either.
What does look durable is the split it draws. Execution, building, checking, refining, is heading toward the agent almost entirely. Judgment, deciding what is actually good and why, is heading toward the human almost entirely. Wherever the frontier moves next, that split is worth holding onto.
What to do this week
- Pick one AI output you approved this week without really looking at it. Go back and judge it properly. Would you actually ship it, or did it just clear the bar of technically fine?
- Put your sharpest eye, not your deepest technical skill, at the approval gate. Taste is not a technical role, and the two are not the same person by default.
- Ask "is this good" as a separate question from "is this correct" on your next agent output. They are not the same test, and only one of them can be automated.
- Take ten outputs from the same prompt or loop and rank them. Notice what you are actually judging when you do. That is your taste, made visible.
Where to from here
Book a free AI audit and we'll show you what's worth augmenting first in your business, and what isn't.
Live with passion & AI,
Brett
Running an event? Put practical AI on your stage.
Keynotes and workshops that send business owners home with a plan they can use Monday morning. No hype.
Frequently asked questions
What is taste engineering in the context of AI?
+
Taste engineering is the human skill of judging whether an AI agent's output is actually good, not just technically correct, and choosing, rejecting or refining it accordingly. It sits above prompting, context and loops, because none of those can tell you whether the finished result is the right one.
How is taste engineering different from prompt engineering?
+
Prompt engineering is about the instruction you give before the work starts. Taste engineering happens after the work is done, when a person decides whether the result is genuinely good. You can write a strong prompt and still need real judgment to know if what came back is worth shipping.
Why did we move from prompts to context to intent to loops before landing on taste?
+
Each stage solved a different execution problem. Context engineering gave the model your business's information. Intent engineering made you specific about the outcome. Loops let the agent check its own work against that outcome. Once agents got reliably good at execution, the open problem left was judging the result, which is taste.
Can taste engineering be automated with more loops or better prompts?
+
No. A loop can check an output against criteria you defined in advance. Taste is the judgment that decides which criteria matter and recognises quality that was never written into the brief. That call has to stay with a person.
Is taste just a fancy word for quality control?
+
It overlaps, but quality control checks for correctness and compliance. Taste is a judgment about elegance, restraint and whether something is genuinely good, the kind of call an experienced editor or designer makes that a checklist cannot capture.
Can someone develop better taste, or is it fixed?
+
It can be built. Most of what reads as natural instinct is pattern recognition from studying a lot of finished work, good and bad, and getting specific about what separates them. Looking at more examples and naming why something works is a deliberate way to sharpen it.
Brett is a four-time founder (Darra Tyres, Gladfish, EzyTrac, Anaboo) and the operator behind AIOS, Anaboo's AI Operating System. He writes from inside the build, installing AI in his own businesses first and reporting back what actually moves the numbers. Based between Singapore, the UK and Australia.
