Where Should a Rule Live So It Can't Be Skipped?

Have you ever stuck a note on the fridge that says "take the trash out Thursday," and then watched Thursday come and go while the note just sat there, being right?

Most of us have. The note isn't wrong, it just has no way to make anybody do anything, so it slowly turns into wallpaper. What usually fixes it is parking the trash can right in front of the back door, where you have to step around it to leave.

Hold on to that trash can, because it turns out to explain a lot of what's in this episode.

Hi! I'm Bob Michaels, and this is Evolving the Web. Every two weeks I talk through the essays I just published, plainly, as one story. This time there are four essays, and they all come back to a single habit: put each job and each rule where it can't be skipped, and put each fact about you where it can't disagree with itself.

Let's start with who gets to write. In my system, HALO, a single kind of AI model writes every sentence that gets published. Other models plan, check and report, but they never write the words you end up reading.

Why bother? Think about a teacher who lets a student grade their own test. The student isn't a bad person, but you'd still raise an eyebrow at all those perfect scores. So I keep the writer away from the judge.

The checking side is pretty thorough, too. It opens every link, compares every quote and every contact detail against the source, and flags any passage the writer didn't produce. And here's the part I like best: a checker can describe a change that's needed, but only the writer gets to make it. It's like an editor who can scribble notes in the margin all day long but has to hand the page back.

There's one more thing about the writer. It gets my style rules before it starts, and if it can't receive them, it doesn't run at all.

I had a standing rule that every project should start with a written plan. Sounds sensible, right? Well, when I looked back, none of my engagements had actually started with a plan. My instructions asked for a summary at the end and never once said "plan first," so the plan just didn't happen.

It gets better. The tool that was supposed to catch this was never called, and it always reported success anyway. That's a smoke detector with no battery that still blinks a cheerful green light at you. So I rebuilt it so it can actually fail, and I tied it to a stop that blocks a session from ending until the check passes.

Dashes taught me the same lesson. I'd banned them in my instructions, and the writing did use fewer of them, but the count never got all the way down to nothing. Asking nicely only gets you so far, so now a check in code catches the stragglers.

Then there are the helper agents. When the main session hands off a task, the helper starts with none of the main session's instructions. Picture a temp who shows up on day one and nobody gave them the onboarding binder. So the content rules now live in the same file installed in every project, and the system compares those copies exactly.

Here's where I landed. When a rule depends on somebody remembering it, it's really just a request. A control is code that can refuse the work. That's the trash can in front of the door.

Now let's take that same idea outside, onto the open web, and point it at your business.

AI answer engines build a picture of your company from your website, your profiles, your listings and your markup, which is the behind-the-scenes code that tells search engines what your page is about. When those sources disagree, the engine has to guess.

It's a bit like asking a few friends what your cousin does for a living. If one says nurse, one says teacher, and one says "something with boats," you walk away with a pretty fuzzy cousin.

The search companies are fairly clear about what they want. Google tells sites to use their real site name in their organization markup, and Bing's guidelines ask for clear, consistent names.

Now, you might hope that adding markup is the magic button. A study of pages that added it found no big jump in how often AI cited them, so markup alone won't fix this. And to be fair to you, nobody I found has measured name consistency directly. The case rests on how the engines say they work and on clearing away contradictions they'd otherwise have to sort out.

So here's what I'd do this week. Pick a single name and a single sentence about your company. Find every place they show up. Make them all match. Then ask the AI assistants the same questions and see what comes back. It's a little tedious, honestly, but so is having your cousin described as "something with boats."

That brings us to money, which is where putting things in the right place got very real for me.

Agent work spends tokens, the little units you pay for, in three places: looking things up, doing the work, and reporting back. Since August, I've had a standing decision that looking things up runs on the smallest model.

Then I did a research run and actually checked where the money went. I'd guessed wrong. The cost sat in how much input two of the vendors were taking in, a spot I hadn't been watching at all.

The bigger surprise came when my expensive agents did their own research. They ran through a whole session's allowance doing it. That's paying your head chef to drive to the grocery store and wander the aisles. These days the cheap agents go find the facts, and the expensive one stays in the kitchen and writes.

Reporting back got trimmed, too. The finished work goes into a file, and the reply is just a short note saying it's there. And if a cost reading ever comes back as nothing at all, I treat that as a bug until somebody proves otherwise. Free is lovely, but in my experience it usually means the meter isn't hooked up.

So that's the thread. The rules that matter most get a place where they can't be skipped, every job goes to the model that fits it, and every fact about you lives somewhere it can't argue with itself. Move the trash can to the door, and Thursday takes care of itself.

If you'd like the longer versions, all four essays are up on bobmichaels.ai, and I'd love to hear which one sounds most like your own fridge note.

I'm Bob Michaels. Thanks for hanging out. See you in two weeks.