For a while, prompts were currency. I shared them with coworkers in private group chats. Tell it it’s a world-class expert. Ask it to take a deep breath. Finding one that worked felt like picking a lock. The locks changed. The models outgrew the coaxing.
If your main AI skill is a list of clever phrasings, that skill has an expiry date. But those prompts did something for us. Writing one made you figure out what you actually wanted. That part hasn’t changed. The model can’t read your mind. It doesn’t know what good looks like for your report, who it’s for, or what would make you throw a draft away. That lives in your head.
That’s the problem worth solving.
What I mean by a system
Not code. The best systems are simple: saved instructions, a folder of strong examples, a checklist you run before trusting an answer, a second chat that reviews the first.
A sentence helps you once. A system helps you every time.
If you’ve explained the same thing to AI twice, it belongs in a system.
The rules
I call them invariants: rules that hold no matter which AI tool you use, today or three years from now. They follow the order of any task you hand to AI.
The brief: telling it what you want
- Write the scorecard first. Decide what good looks like before you see any drafts, or the drafts will decide for you.
- Name the audience, not the persona. Who reads the page matters more than who the model pretends to be.
- Show one good example. It sets tone faster than a paragraph of rules.
The work: what passes between steps
- Gather first, judge second. Collect the facts, then draw conclusions from only those.
- Give it the real material. Your summary hides the contradictions the model needs to see.
- Tell it when to stop. “Three options, then wait for me” gives you a point to step in.
The check: trusting what it gives you
- Let it come back empty. A check that can’t fail hasn’t checked anything.
- Run it three times. Where the answers split is where you put your eyes.
- Make every claim point to its source. No line to point to, no claim.
These are a few from each part, not the full set. The series builds out the rest.
What you can expect
Each invariant comes with the fix for the prompt in front of you and the way to make it permanent, with before-and-after results so you can judge for yourself.
You don’t need a computer science degree for any of this. You need to know what you want and be willing to write it down once.
When the ground moves or a rule breaks, I’ll strike it in public and show what replaced it.
That’s the whole idea. Systems, not sentences.