Listen to this post: Apple Podcasts | Spotify | YouTube
A five-part series on working with the robots. Part 1, Part 2, Part 3
Your job is to keep diverging. The robots’ job is to converge. The space between, where you take what’s tangled in your human head, and hand it off as something the robots can run with—that’s the prompt.
There’s a popular image floating around called “The Anatomy of an o1 Prompt”—Goal, Return Format, Warnings, Context Dump. Color-coded sections. It’s a clean diagram. It looks like a recipe.
Templates like that are fine, but we’ve come a long way since o1. They’re not how I prompt. Not anymore.
I’m a human, and I’m working to be more human all the time.
What I’ve found, after building dozens of these things, is that good prompts get reverse-engineered, not architected. You don’t sit down with a blank page and a template and write the perfect prompt. You have a long, messy conversation with the AI about the actual problem. You go in circles. Eventually, you end up somewhere useful. Then you turn around and ask: if we had to have that whole conversation again, how would we shortcut it?
That’s the prompt.
It’s the conversation distilled.
How Prompts Get Built: Fitness & Finance
Fitness: The first version of my Centaur Fitness prompt was a lengthy conversation between me and Gemini. I told the AI about my current exercise plan. What I’d done in the past. Who I read and admired—Greenfield, Berkhan, Sisson, Patrick, MacKenzie. What my goals were. I asked what the latest science had to say about these ideas. I shared what I genuinely enjoyed doing, even if it wasn’t optimized.
Eventually we landed on a weekly framework that felt right. Simple. Maintainable. Mine.
Then I asked a question that turned the conversation into a tool: If we had to do this all over again, what’s the shortest interview that would have gotten us to the same place?
I tested that on the other LLMs. A good prompt will get reasonably close answers from different robots.
The conversation, with the dead ends removed, becomes a prompt worth using and sharing.
The header at the top—You are an elite health and fitness consultant… interview me—is assigning a profile, a personality, instructions for how to run the conversation while keeping the end in mind.
What’s left, and what guides the interview, are the questions that mattered. Once we found them.
In a way, I didn’t write this prompt. I tested it. I edited it. I finished it.
Finance: Years ago I built a scoring system for stocks. An idea I heard from David Gardner at Motley Fool. Then I started bolting things on. The importance of free cash flow, from Bezos’s early shareholder letters. Then, free cash flow per share, from Buffett and Munger. Chris Mayer’s 100 Baggers gave me the “twin engines”—growing earnings and an expanding multiple.
Years of small edits, updates to my thinking, gave me this scoring system, hardened into something I trusted.
Every quarter I’d refresh the scores by hand. I’d color code—red, yellow, green—in Excel. Which businesses were worth buying? And more importantly, which businesses, no matter the turbulence, should I be holding? When a stock was in the tank and every instinct said get out, I’d look at the score and think: I’m not selling that. It’s a 10-point stock.
The robots didn’t change the system, they automated it.
But the real benefit wasn’t speed. It was an upgrade I hadn’t been able to crack on my own.
Why?
Math.
A 10 point stock is a great business. But what should we pay for it? Price is important. Charlie Munger said, “It’s far better to buy a wonderful company at a fair price than a fair company at a wonderful price.” But my system only covered the wonderful company bit.
I’d spent years hacking together formulas to try and find a price, a value, I believed in. But again, math is hard work for us divergent types. Less hard with a convergent robot friend sitting next to you.
I wanted to take an industry standard, estimate and discount future cash flows, and then apply a margin of safety based on my scoring system.
A 10-point business earns a bit more wiggle room than a 7-point business. As they should.
The quality score isn’t just a conviction rating anymore. It’s also a dial that shows how much I’m willing to pay for it.
The fitness prompt, I might not touch again for a year. We’ll see how fluffy I get. The stocks run every quarter, on every position I own. Different jobs get a different cadence, but they both require the plus.
Better together.
What Makes A Good Prompt
After building a bunch of these—fitness, stocks, an editor for my writing, a coach prompt for my baseball players’ swings (yes, the robots can see video now), a cover-art prompt for the blogcast—a few patterns hold up.
Light beats heavy. Code can be heavy or light. Prompts are the same. A bloated prompt with twelve sections of preamble doesn’t outperform a tight prompt that names what matters and gets out of the way. It just uses more tokens, runs slower, and gives the AI more ways to misinterpret. The best prompts I have are very direct.
Interview first, output later. The fitness prompt puts the AI in ask mode before answer mode. For me, the interview, one question at a time, pulls more out of me. It surfaces injuries, habits, desires. It connects ideas from the guru you read, to the science you don’t, to the goal you’ve been thinking about but hadn’t put into words.
Constrain the output. Give me a seven-day grid. One line per day. That’s a constraint. Constraints save you from the LLM’s natural urge to over-produce. To over-explain. Without a constraint, the AI gives you a lot of wind. They literally know everything, so they can go on for quite some time.
Forbid the BS. The most important don’t do this clause has nothing to do with the task. It has to do with how the AI talks to you.
By default, LLMs are catastrophically agreeable. Great question! Absolutely! That’s an interesting point! They tell you your idea is brilliant, your draft is great, your business plan has legs, your novel is going to sell. They flatter your premises before answering. And the worst of them walk back their opinions the moment you push.
It’s super annoying.
So tell the AI to knock it off.
No “great question.” No “that’s a fascinating perspective.” Lead with the strongest counterargument. If I’m wrong, say so…
That language gets you an honest read. You stop having to filter compliments and wade through BS, to get your answer.
This is why I keep using Lex for my writing, even though it doesn’t always run on the latest models. The team that built it tuned the personality hard—Lex is mean as hell, but in a useful way. It pushes back on weak sentences. It tells me when something is repetitive or unoriginal. It makes me doubt my own work in the places I should be doubting it.
When Prompts Fail
The fitness and stocks prompts work because the answer is out there, but it’s only yours once you put yourself into it. The machine brings the rigor. You bring the taste—your perspective, your goals, your influences.
Every week I record the audio version of my blog. We call it a blogcast around here (like and subscribe!). The intro needs to set up the piece—a hook, a frame, a reason for the listener to lean in. I built a prompt for this. Decent structure. Clear ask. Examples of intros I’ve written before.
When I run it, the AI gives me something that matches the post. But is a match what we want? A post about freedom gets an intro about the feeling of freedom. A post about decision-making gets an intro about the weight of a hard choice. Coherent. On-topic. And a little bit stale.
The good intros—the ones I use more often, the ones I write myself—pull from outside the post. A book I was reading that week. A movie I half-remember. A conversation with one of my kids. An old quote that surfaces as I start recording. A connection the post never makes explicit but quietly suggests.
Additive rather than repetitive.
That’s the divergent move.
The AI is very good at understanding the post. What I want is what’s next to the post—what associates, what rhymes, what surfaces when a writer’s mind drifts away from the assignment for thirty seconds. And the robots, at this point, still need us humans for that.
Some jobs are “+” jobs. Some jobs the human still does alone.
Vibe Coding
I know, I know. I have a website, so you’d think I was technical. But it’s just not so.
That said, I am now a coder. And I have Replit, Grok, and Claude to thank.
These services can vibe-code for you. Just give them prompts, tell them what you’d like to build, and the little robots go to work.
I’ve built a few web apps that are in beta right now, and might be forever.
- TempRx: Sauna Calculator & Cold Plunge Calculator
- CanIRetire: Retirement Calculator
- Flow Finder (Pomodoro Timer)
- Tabata Timer (in the works)
- Breathwork Coach (in the works)
- HelloRatio—a measure of your city’s friendliness (launching soon)
They start here, and if they earn it, they migrate to their own homes. I run ads and make dozens and dozens of dollars.
But mostly, it’s just fun to make stuff.
You can build just about anything you want. With a place to host the code, you’ve got your own personal App Store. A Swiss Army knife of tools, customized for you, shareable with the world.
I’m not running a frontier lab. I don’t have a billion-dollar budget or a thousand engineers. But the same flywheel is available to me, and to you, on a smaller scale.
I had an idea on a Tuesday. Wednesday I had a working app.
That doesn’t fit any software-development timeline of the past. It only makes sense because I have a robot, the equivalent of a whole team of developers, writing the code for me.
Suck it, CS grads.
But building a site, thinking of an idea, that’s the easy part. Keeping it going? Executing on a vision? That’s work.
TempRx and CanIRetire both graduated to their own sites, and they both have content on those sites. But I let the robots write that content. The posts, the city guides, the scenarios. All AI, and it shows. No plus.
CanIRetire’s calculator—I designed that myself. Argued with the robot over the assumptions and outputs. It has my fingerprints on it, and it holds up. The posts that came after, I handed over completely. I didn’t want to write those. And that is a signal.
Same story with the TempRx directory. I like the idea—a real resource of the best saunas and cold plunges in the world. City by city. But using the robots? I think that directory only comes to life if it’s curated, not aggregated, by someone who got in the water.
A good product is one you want to keep showing up for. One that has meaning. If you are willing to hand off 100% of the project to AI, that’s a clue. If it lacks meaning to you, my hunch is it will lack meaning to everyone else as well.
The robots can ship the app in a day. But they can’t make you care about it in six months.
The handoff requires the plus.
If you enjoyed this post, please share it with a friend.
What do you think?