“Context instead of control. Part of my job is making sure people have the context they need to do their jobs well.”
Elizabeth Stone, CTO of Netflix
●↑
This is the method engineers use to get reliable results from models, inspired by the work of Andrej Karpathy.
A few words about him, because his story is why he’s worth listening to.
Who is Karpathy? (Skip if you know him)
Karpathy was born in Slovakia in 1986. At fifteen, he emigrated to Canada with his family. In Toronto, he ended up in Geoffrey Hinton’s classrooms, the neural network pioneer, and realized that this was the path for him. At Stanford, during his PhD, he created the university’s first deep learning course and taught it himself: it became one of the most popular courses on campus. And to find out how good machines were at recognizing images, he competed against them, judging each photo by eye, to see how far a human could get.
He’s one of the founders of OpenAI, the company that would create ChatGPT. And for five years, he led AI at Tesla, where his team developed self-driving technology.
In 2024, he could have worked anywhere, but instead he founded a school, Eureka Labs, and published free lessons on YouTube that have taught millions of people how these models work.
And, in the ultimate move, this year he agreed to work at Anthropic.
That’s why I listen to him. He built these systems, put them in millions of cars, and spent years teaching anyone who wanted to learn, for free. When someone like that explains how he works, it’s worth taking notes.
I watched it twice because halfway through there’s something very interesting: engineers have just rediscovered, in their own way, what people who work toward goals have known for years.
Karpathy proposes a three-level method for working with AI without getting ridiculous answers. Because ridiculous answers do happen, and when they do, someone always tries to look clever on LinkedIn by posting a screenshot. Too bad that screenshot doesn’t prove AI is stupid. It proves that the person posting it hasn’t yet understood what they’re talking to.
Three Levels for Strategic Prompts
Level 1: Phrase the initial prompt request well
AI is brilliant, but it doesn’t know your company. Everything you get depends on how you write the initial prompt request.
Remember the car wash example? If you ask, “I need to wash my car, the car wash is 50 meters away, should I walk or drive?” the most advanced models answer that you should walk because it’s close.
I repeated the test with Claude, Gemini, Grok, and ChatGPT while writing this note, and it seems the bug has been fixed.
Why did they all get it so spectacularly wrong? Because AI is great at things that can be measured and calculated, and even though a human could infer the missing information, the only metric it was given was distance, so all its reasoning revolved around that number.
AI doesn’t know the obvious context of your life and your company, and above all, it doesn’t know that it doesn’t know it. Only you have that context: either you write it into the request, or it doesn’t exist to the machine.
In fact, every time you ask it to “create a report,” you’re writing a prompt at the same level as the car wash question.
Creating a report isn’t an objective. The objective is the decision that report needs to guide.
Read that sentence again.
It’s the same mistake you (maybe) make with people, too.
If you opened your to-do list, I bet you’d find something like this:
- Write follow-up emails
- Call with the client
- Presentation tomorrow
Written this way, these tasks only make sense to you, the person who wrote them, and only today: in three days, you won’t know what they’re about. You can’t hand them off to a colleague, a supplier, or an AI. They stay attached to you.
In MAKE PROGRESS® there’s a tool designed specifically for this problem: Micro OKRs. Take the task and rewrite it to include three things: exactly what’s needed, a number, and what successful completion looks like.
Look at the difference:
- “Write follow-up emails” becomes “get confirmation from the client for the next step {describe what it involves} by 6 p.m. today”
- “Call with the client” becomes “project roadmap approved by the client during Thursday’s call”
- “Presentation tomorrow” becomes “10-slide presentation with 3 pricing scenarios, approved by the team, ready by 9 a.m. tomorrow”
Same work. Different words. But now anyone reading it understands what to do, how much work it takes, and when they can say it’s done. And “done” can’t be a matter of opinion.
This is the lingua franca of execution. Your colleague understands it, and so does the person who started yesterday. AI understands it too: a Micro OKR is already a well-written request, with no prompt-writing course required.
Level 2: explain how to self-check the result
Karpathy uses a metaphor I really like: when we work with AI, we’re not building animals that can be made to do what we need just by raising our voices—we’re summoning ghosts.
An animal—and we are animals—learns by living in the world. It gets hungry, it has ambitions. If you tell a capable person, “Become an expert on this topic in two weeks,” they’ll somehow figure it out, because something inside them drives them. A ghost won’t.
An AI model rapidly replays everything humanity has written: a statistical imitation. It doesn’t live in the world and it doesn’t want anything. Encouraging it, threatening it, or writing “do it better” won’t change the result.
And if it’s missing information, it makes it up, with absolute confidence, because otherwise its reasoning falls apart.
Understanding this difference changes how you work: you motivate an animal, but with a ghost, you have to check that it’s real, like the Ghostbusters did.
To do that, you need to decide in advance what “good result” means and check it against real data, not feelings.
Now go back to the three Micro OKRs from earlier and notice something: the verification is already written into them, in the words themselves.
- “Confirmation from the client by 6 p.m. today”: at 6:01, open your email. The confirmation is there or it isn’t.
- “Roadmap approved during Thursday’s call”: on Thursday evening, the roadmap is approved or it isn’t. No need to ask, “How did the call go?”
- “10 slides, 3 pricing scenarios, approved by the team”: count the slides, count the scenarios, ask the team. Three checks anyone can do in a minute.
When you write a task this way, you’ve already built in how to verify it before you start.
“Write follow-up emails,” on the other hand, can’t be verified because any email will do. And you know what happens? You might actually send it!
If the prompt says what successful completion looks like, AI has a criterion it can use to check its own work before handing it over. It’s not an extra step you take afterward; it’s the way you write the request.
Level 3: build the environment where the work happens
Requests and verification need a workspace that improves over time, instead of starting from scratch with every conversation.
Engineers solve it this way: they write the rules in a document that AI reads before getting started. They organize documents into folders with a clear, predictable, standardized structure. They turn recurring tasks into operating manuals (skills). And they divide actions into three categories: always do, ask first, never do.
Does any of this sound familiar?
It’s a full-fledged operating system.
Does your company have one?
Is there a written strategy everyone can read, with a few measurable priorities, procedures to make sure the same problem is never solved twice, and a weekly meeting to review the numbers and adjust course?
That’s MAKE PROGRESS® (in fact, I copied the paragraph from my book)
Micro OKRs are more useful than ever.
Explore these ideas in the original newsletter.
A person who writes tasks this way immediately improves their work. But for an entire company to speak that language, you need a method: who sets priorities, how often the numbers are reviewed, and where the strategy and procedures live.
Micro OKRs are the language . MAKE PROGRESS® is the grammar that makes it work across a company.
This method for writing prompts is just the last mile, the part where you talk to the machine to move faster.
First come a clear strategy, well-chosen metrics, and a system that holds them together.
Without all that, AI will tell you to walk to the car wash for almost every task. You just won’t notice, because neither of you knows the strategy!
A 120-second exercise.
Open your list and pick the first 5 tasks. Count how many include a number and a definition of “successfully completed.” Then rewrite one as a Micro OKR and give it to someone, or an AI, without adding any explanation.
Imagine doing the same thing across the whole company: written, measurable priorities; checks decided before getting started; procedures and data in the same place. And an AI that reads all of it before answering you.
How much would the quality of what you produce change? Not the quantity—the quality.
That’s exactly what we’ll show tomorrow.
Let’s talk on June 17 at 12:00 p.m.
At 12:00 p.m., a free Leadership Augmented webinar. No list of tools: how to build the three levels that come before any AI—strategy, verification, and system.
Leadership Augmented™ – Business Strategy with AI and OKRs
See you there—I have lots more to tell you!
ALWAYS MAKE PROGRESS ●↑
Antonio
The video that inspired this newsletter
This edition was inspired by “Andrej Karpathy: From Vibe Coding to Agentic Engineering w/ Stephanie Zhan”, a conversation between Andrej Karpathy (one of OpenAI’s founders, who spent five years leading AI at Tesla and is now at Anthropic) and Stephanie Zhan, published in April 2026.
Karpathy explains why vibe coding isn’t enough and what it takes to move to agentic engineering: precise specifications, verification of results, and a work environment that improves with every cycle. In this article, I’ve translated those same three levels into Micro OKRs and MAKE PROGRESS® for people working on strategy, even if they don’t write a single line of code.
