How to Avoid AI Slop

My Medium friends can read this over on Medium.

You know AI slop when you see it. It's not an AI problem, though. Input output. Put something bad in, like a generic prompt, and something bad comes out.

Nevertheless, the feeds are filling up with it. I looked at how far that has gone in a separate piece, and the numbers aren't pretty.

And still, I use AI every day. For research, for drafts, for the boring automation parts. So how do I use it for publishing myself?

For me, it comes down to three steps. No secret prompt you need to buy for $29, I promise.

Ideas

The slop starts before the first word is written.

Ask a model for "10 blog post ideas about productivity" and you get the same ten ideas everyone else gets. The AI model is not lazy. It's just that thousands of people typed that same request today, into the same models, trained on the same text. The generic is kind of built into that system.

So the real currency now is ideas.

The idea comes from my own work, a comment under my last post, a search query my stats show me, Google Search Console results, or the thing that annoyed me on a Tuesday. Then I bring that idea to the AI, not the other way around.

This detail changes a lot!

The more of the thinking you hand over, the more generic the result gets.

AI is useful for developing an idea. Poking holes in it, finding the numbers, structuring the argument. But asking it to have the idea is where the problems with AI "slop" begin.

But it's not all...

Rules

Out of the box, a model writes in its default voice. Which is everyone's voice. Which is no "good" voice at all. It sounds generic. It should. Makes sense.

So the second step is telling it who you are as a writer.

Not in the prompt, every time, but in one place it always reads. Claude calls that a skill, ChatGPT has custom instructions and CustomGPTs, Gemini has Gems. The name doesn't matter much. It's a file with your rules. Rules for voice, style, tone, wording, forbidden phrases, and more.

Mine has grown over months. A lot. Banned words (no "delve", no "seamless", and in my case no em dashes at all). How long my paragraphs run. How I hedge, and what I don't claim. Phrases I don't like and never want to see again. Every time an edit annoys me twice, the rule goes in those files. And yes, those are many skill files by now.

Those files do more for the output than any model upgrade I've tried so far. I went deeper on this in How to Write Human with AI.

A habit, not a one-time setup. And it does take some adjusting.

Editing

Then the last one.

Read the whole thing before you publish it. Many people don't do that. They automate… and leave.

But at least once, at least quickly, read the thing.

Cut what you wouldn't say out loud. Check numbers and claims (with another AI if you want), because models still make things up with full confidence. Add the things only you can add like a joke, a fitting emoji, the detail from your own week, things like that.

Ten minutes change a lot. The difference between "AI wrote this" and "I wrote this, with AI" can be massive. And that shows.

Downsides, of course

Time, first.

Three steps take some time. Slop is one prompt and a couple of minutes. That is the appeal of slop. We all get that. If you publish twenty posts a day, you can't do much of these 3 steps. The output will show it.

The rules file is never finished. Mine still grows, and some tells slip through anyway. That's annoying.

Also, you need a voice to describe. If you have never written without AI, there is nothing to put in the file yet. Then the file starts almost empty and fills up as you edit.

The Bottom Line

The models write fine. Better every few months, probably.

What they can't do is have your idea, know your voice, or care whether the thing is true. Those three are still on us. I wrote a while ago that having something to say is the job now. These steps build on that.

It's not much work. Less than before AI. But still work. And work that many don't want to do.

If you're not already, follow me on Medium.

Get Superwriter