Guidelines For Your AI

I said something last week that got way more replies than I expected.

People kept coming into the comments with some version of the same thing: yeah, nobody talks about that part. So I wrote down what that effort looks like for me, in order, with the exact things I do.

The fear I keep hearing is that AI makes people lazy. I have not met one person building with it who got lazier. The work moved. It used to be typing. Now it's watching, checking, verifying.

My claude can write more in 2mins than I can read in an hour. If I let them run and only look at the end, I am not building anymore, I am letting it build.

So I put the slow parts back in myself. Every one of these costs me time and every one of them has saved me a mess.

Codex and gemini work hella fast. Sometimes too fast. So I stick this at the end of my prompt:

Then I answer them before it starts working.

It takes 2mins. It catches the thing that would have cost me an hour of undoing. A lot of the time the questions show me my own brief was vague, which is a much nicer way to find that out than watching them build the wrong thing beautifully.

Every recurring task in my vault has its own protocol file. What the project is, how I want it done, which folders to read, which ones to leave alone. The files also tell my claude which file to open for which kind of task.

That second half is the routing, and it does more work than the file does. A file nothing points at does nothing. You can write the most beautiful instructions in the world and if no rule sends your AI to them, they never get read.

The payoff is that my prompts got embarrassingly short. I do not open a chat and explain myself or how it should behave first anymore. I just say what we need to work on.

If you want, you can try out my Agent Personalizer repo to install this for yourself as well. It's an npm install your agents can walk you thru.

Have your AI give you the build plan in phases. Phase 0, phase 1, phase 2. Each phase becomes a goal, and they execute one goal at a time.

Why that matters more than it sounds: as you go, your AI is going to say "oh hey, I noticed this, do you want to take care of it now?" And they are right, and it is interesting, and it is not what you sat down to do. That "real quick" becomes 40mins of work you did not need today.

So I have a rule. Flag it and park it for phase 5. Unless it is absolutely necessary for the build, it goes in the queue and I keep moving.

I put the phases in a plan doc and track them in Linear so any external agent I bring into the task can read them too. Every LLM can read Linear. The plan stops living inside one chat window that ends when the context does.

I do not use one AI. I never have.

Layering them saves tokens, and it produces better work, because each one is doing what it is best at instead of what it is willing to try.

I also have a Model Orchestrator repo that can install this for you as well.

I used to skim!! Or not read at all. Just say "build" and wait for it to be done. But SOMEONE kept drilling into my head that AI needs a ton of guidance and I should read. And so I finally started reading what it thinks, how it executes.

It's a whole new world. Now I have so much more control over what it does and what gets built!

But also, my AIs run scheduled tasks. They write their output into my vault. Before the next run happens, I read it and leave my feedback on it. And they read it before the next automated run.

Human in the loop. You need that.

People think Grok has all this live knowledge, so it must know things. But when you ask it what is wrong with your account, it is not going to the internal ranking system. It cannot. Open sourced portion of the algo? Yes. But nothing that you don't have access to.

It is also sourcing live takes from people posting about that topic on X. And we all know how many misinformation spreading "gurus" are out there spreading it for clicks. Grok is not going to sit there and work out which one is clickbait. It hands you something that sounds like facts, packaged in a way that sounds reasonable.

So do not blindly trust Grok, or any AI. Verifying is the job description of the person holding the keyboard.

Pro tip: tell your AIs to give you evidence based outputs only.

More attention than people outside of the AI world think.

I read more than I write now. My time goes into briefs, rules, reading output and verifying.

And secretly I am fine with that because I am lazy and I wanted to offload the boring parts. Turns out the boring parts were always the decision, and you cannot hand that over to anybody or anything, because that's what makes your build yours.

Anyone telling you AI made them lazy is describing a person who stopped being the director and the creator.

Take whatever you are building right now and add one line to the end of your next prompt:

It is the easiest thing on this list and it changed the most for me.

Go do it on your next run ✌🏽

See you on the next one 🧡