Using AI like the top 0.1%
Seven AI prompting techniques from Alicia Lyttle's masterclass, the account security settings to change before you use any of them, and what all of it looks like when your working material is test data.
By Gillian O'Donnell, yieldHUB · 6 min read · September 2026
I turned 25 in July. One thing on my list was to learn to use AI the way the top 0.1% use it.
Two weeks later I was in a room with 3,000 women in Fort Lauderdale, listening to Alicia Lyttle the Queen of AI.
165 days without writing a single email manually, and more productive than ever.
Then she introduced her AI Chief of Staff, Maximus.
Maximus is not a person.
What follows is her framework, plus what I think it means for those of us who work with semiconductor production and test data.
First: lock your AI accounts down
Alicia made 3,000 people do this before she taught a single technique; two minutes of settings work comes first.
| ChatGPT | Claude |
|---|---|
| Settings → Account info → turn on multi-factor authentication → Data controls → turn off "improve the model for everyone" | Settings → Privacy → turn off location metadata and "help improve our AI models" |
Worth being clear about what this does and does not solve. Turning off model training stops an AI model from learning from your conversations. It does not make a public chat window an appropriate place for customer data, and no settings page will. The rule we work to is simple: general-purpose AI for thinking, a controlled environment for the data itself.
Seven techniques
She gave us the exact prompts and workflows for each of these. The principle underneath them: AI isn't giving you average answers, you're asking it average questions. Every technique below is a fix for that.
Let AI interview you before it does anything for you
Most people hand over a task. The top 0.1% hand over a conversation. Ask the model to question you until it has what it needs, then let it start.
Assign a price tag to the output
"Write me a strategy" gets you $50 work. "Write me the $50,000 version" gets you something else entirely.
Cast a specific character
Not "a consultant." A specific one, with a specific track record.
Don't trust one answer. Convene a panel of experts
One opinion is a draft. Several opinions, argued against each other, is a decision.
Make AI critique its own work
Then make it rewrite based on its own critique.
Run your draft through a second AI before you trust it
The cross-examination is where the quality comes from. Don't let one model grade its own homework.
Change modes
Genius mode exists. Most people never use it.
| One more she added |
|---|
| Tell it to run your prompt now, at the end of your prompt. It stops the model paraphrasing your question back at you. |
She then took it a step further and taught us how to build an AI virtual team: not one assistant doing everything, but a set of agents.
What changes when the subject is test data
These seven make you better at getting good work out of a general-purpose model. They do nothing about the harder problem underneath, which is the state of the data you would point that model at.
Test data arrives from multiple test houses in different formats, with inconsistent test names, changing limits, retests and duplicate touchdowns. An AI given that raw will answer confidently and wrongly, which is worse than not answering. Technique five, making the model critique its own work, is the closest thing on the list to a defence, and it is not enough.
I've written before about whether LLMs can be trusted in semiconductor manufacturing, and about how AI earns that trust. The short version of both: it is earned on the inputs, not the model. This is why our work at yieldHUB starts with cleaning and centralizing data before any analysis, machine learning or otherwise, is allowed near it. Correlate parametric and bin results across sites only once the inputs are aligned. The same discipline applies to prompting: the quality of the question sets the ceiling, and the quality of the data sets the floor.
At yieldHUB, clean, reliable, trustworthy data that acts as a single source of truth is of paramount importance. That is why we approach AI from a reliability lens. We bring our semiconductor test data expertise together with world-class, market-leading data science and AI to deliver digital transformations for our customers. yieldHUB Live is one example, where the results for our customers on their test floors have been remarkable.
You + AI agents = your unfair advantage.
For engineers, I'd add one condition.
Clean data + AI agents is the advantage.
Everything before that is a faster way to be wrong.
