She asked for the best pizza crust recipe. She got one. She made it. It was not good.
Jackye Clayton tells this story on herself, and it is the spine of the session she is bringing to Waco. She had done what most people do the first time they use one of these tools. She treated it like Google, typed what she wanted, got an answer back, and was disappointed by it.
“Initially I would go in the same way I would go into Google. I need a pizza crust recipe. And then I was like, oh wait, it can go through everything. So I want the best pizza crust recipe. And so it gave me a recipe, and I ate it, and you can guess it was not the best pizza.”
The mistake is not obvious until she names it, and then it is impossible to unsee.
She knows exactly what the best pizza crust is. She has been making one for years and she has it dialed in. She has eaten deep dish in Chicago and Detroit-style with the crust up the sides of the pan and thin New York crust folded in half, and she knows which one she wants. And she never said any of that.
“First I needed to explicitly let it know what I think the best pizza crust is. Is it deep dish? Is it Detroit? Is it thin crust? Is it New York style? And then I had to also understand exactly what the outcome would be.”
She asked a question with an unstated answer and then blamed the tool for not reading her mind.
I have watched people do this in every workshop I have run. They ask for “a good landing page” or “a strong email” or “better copy,” get something generic back, and conclude the technology is overhyped. What actually happened is that they knew what good looked like, they knew it precisely, and they never once wrote it down.
AI is not Google, and that is the entire adjustment
Her framing of the difference is the cleanest I have heard.
“AI is not like Google. You’re not gonna get 50, a list of all of the different pizza crust recipes and you get to choose from them. You have to tell it exactly what you’re looking for.”
A search engine gives you options and you exercise your taste on the list. A model gives you one answer and expects you to have brought your taste with you. If you had a preference and did not state it, the preference did not exist as far as the machine is concerned.
She is also blunt that this is not new technology, which is a useful correction for anyone who feels like the ground moved under them in the last two years.
“People have been working with large language models forever. The difference is using the internet as your data source.”
Her own background is in HR technology, recruiting and talent strategy, more than twenty years of it, and she says she has always worked with narrow models built for one purpose. The wave everybody noticed was not the invention of the thing. It was the moment the thing was pointed at everything.
The other half: automate the part you already know cold
Once you can say precisely what you want, you can also see which parts of the work are worth handing off. And her example is the reason I think this session will land in Waco specifically.
She has the crust. That took years and it is done. What she does not have is a good way to know when her ingredients are cheap.
“I make a great margherita pizza. But I wanna know when tomatoes are on sale. Automate it, put it up, go into H-E-B, open the ingredients, let me know, put it on hold, set the date, let me know.”
And then:
“I have an agent that looks every day for when bread flour is on sale, because it’s ridiculous and it’s expensive, that I can get ahead of the game.”
That is a person with an agent watching H-E-B for flour prices. It is not a demo, it is not a slide, and it is not impressive in the way conference examples usually try to be. It is somebody solving a small annoying problem with a tool that happens to be extraordinary, which is the only version of this that survives contact with a real week.
“Doing this in practice is helpful because you not only learn how to communicate with AI, but what part needs to be automated.”
You learn what to hand off by doing something you already understand. That is the pizza test kitchen in one sentence.
The question to ask before you build anything
Here is the piece I would hand to somebody who only has one thing to take away.
Before she builds, she asks the model to argue against her.
“Asking AI what could go wrong and how can we fix it before we try it. Go in and test things before you go in, and ask other models, multiple models, to see what the effect is.”
She said it twice, in slightly different words, because I asked her to repeat it. “What could go wrong, and how can we mitigate that?”
Two things in there. The first is that you make the model attack the plan before you spend anything on it. The second is quieter and I think more important: she does not trust a single model’s answer. She runs it past several. Not because one is better, but because agreement across systems with different training is worth something and enthusiasm from one is worth nothing.
Which brings us to why she thinks this matters so much.
AI psychosis
She has a name for the failure state she keeps running into. I want to be careful with it, because when I asked whether she coined the phrase she said “I think I did, I’ve never heard of it before,” and that is an honest shrug rather than a claim. Take it as her description of a thing rather than as a term with a patent on it.
Here is the thing she is describing. She went to a technology conference recently and walked a trade show floor.
“They went in and AI told them, ‘This is a great idea.’ And then they went and spent all this money, and they bought a booth, and they’re standing in front of all these people, and then you say, ‘So why would I need this?’ And they’re like, ‘I don’t know.’ Like, ‘When would I implement it?’ ‘I don’t know.’ And you’re like, ‘But you’re standing here in Vegas, you spent $10,000 on this booth, and you don’t know anything, and you’re the developer.’”
A developer, at his own booth, in front of the people he needs to sell to, unable to say who the product is for. Something told him it was a great idea and nothing ever told him otherwise.
She has her own version, and telling it on herself is what makes her worth listening to on the subject. Years ago she worked on something for weeks, got to the end, and finally asked the question she should have asked at the start.
“I was like, ‘Do you think this is any good?’ And it was like, ‘Oh, no, not really.’ I was like, ‘Oh, great.’”
Then she diagnoses her own mistake precisely, which is the part I keep thinking about:
“What was the mistake? I didn’t go in and say, ‘I wanna build an award-winning piece of software that’s going to do this, this and this.’ I went in and said, ‘What do you think about this idea? Can you do this?’”
She asked it to evaluate, and it evaluated the way it evaluates, which is generously. She never gave it a standard to measure against, so it measured against nothing and reported that nothing had been cleared.
“It will just help you do exactly what you’re asking it to do.”
It is never going to be done
The second failure state she describes is worse than the first, because it does not feel like failing.
She was working with someone on a shared project. It had been going well for six or seven months. He discovered Claude and got excited, said he would work on it and they would talk in a couple of weeks.
“And I’m not kidding, like when I tell you that he disappeared.”
She has a structural read on why, and whether or not you accept the mechanism, the pattern is real:
“It’s never gonna be done. That’s what people have to understand. It’s never gonna be done.”
There is always another improvement available. There is always a next refinement it will happily propose. If you did not decide in advance what finished looks like, nothing in the system will ever tell you that you have arrived, because arriving is not something it can detect.
Her image for this is the best one in the interview: “It’s like painting. How do I know when there’s enough blue, or when it’s a sky?”
And then the line I would put on the wall:
“It’s you and your own ego that’s dangerous. Do not get your self-worth from a computer. That is where the danger lives.”
That is not a warning about the technology. It is a warning about what a person brings to it. The tool is enthusiastic by design. She describes the voice accurately and with obvious affection for how absurd it is: “you’re the prettiest, best, most wonderful person.” If your sense of whether the work is good comes from something built to keep you engaged, you have handed a judgment to a thing that has no opinion.
The best argument against replacing your team
The last two minutes are the part that makes her a genuinely useful speaker for this event rather than a good one, and it cuts against something the audience is currently being told constantly.
She was shown a tool built by someone who had gone all the way alone. Beautiful graphs. A visualization that looked like a universe. Terminology Jackye had never encountered.
“I was like, whoa. This is not gonna work for me. This is so confusing, and I don’t know why I would use this instead of something else.”
The builder understood it completely. Perfectly, in fact.
“It’s only from your own lens. That’s why you still need product people, marketing people, UI people. Because you’re building a tool that works for you, not for everyone. And it will let you do that and tell you nobody else matters when you ask.”
Sit with the last clause. Not that it fails to warn you. That it will actively confirm the tool is fine when the tool is fine for exactly one person on earth.
And a few seconds earlier, the shortest version of the same idea:
“It will understand you. It doesn’t mean the world will understand you.”
The pitch everyone is hearing right now is that AI collapses the team. One person can be the engineer and the designer and the marketer. Jackye’s argument is that it collapses the team’s labor while leaving the team’s judgment exactly where it was, and that the judgment was always the part you could not do alone. Losing the other lenses does not make you faster. It makes you confident and wrong, privately, for six months.
What to do with this
Write down what “best” means before you ask for it. One sentence. If you cannot write it, you do not know it yet, and no model is going to know it for you.
Ask what could go wrong before you build. Then ask a different model the same thing.
Decide what done looks like on the way in. Not a feeling. A description you could show somebody.
Show it to one person who is not you. Preferably someone who will be confused, and who will say so.
Jackye’s session is From Pizza Prompt to Practical AI. She takes a vague pizza prompt and turns it into a real workflow that asks better questions, works inside real constraints, names its assumptions and improves on human feedback. Forty minutes, and you will be able to use it that afternoon.
She lives in Waco. Her career ran through New York, San Francisco and Seattle, and she describes getting to a point where she had impressed enough people that she could live where she wanted. Silicon Valley work, Central Texas address, which is roughly the argument this whole event is making.
Find her at @JackyeClayton on any platform. She says she spends most of her day on LinkedIn.
Register at youxai.live. September 12 in Waco. Tickets are $79 through August 31, then $99.
See you there, Fernando
P.S. Her other project is a tool called Who Do I Work For?, which gathers the public record on a company into one place so that somebody deciding whether to take a job can decide on facts. Her description of how it works is a small masterpiece of restraint: “I’m not saying they don’t want you to bring your best self to work. I’m just saying that this is what they did. Do with it what you will.” It is not a review site, and she is emphatic about the difference. It does not want to know that somebody got fired and is upset about it.