how I use AI to write without producing slop

I'm guessing you've tried getting AI to write for you and it turned out like shi... let me keep this PG, it turned out kind of crappy.

I've been using AI to write for nearly 2 years and it was good, but not great. It kept defaulting to generic statements like:

"It's not just a newsletter. It's a movement."

"That's when everything changed."

That AI slop shows up in nearly everyone's AI-written content. Your audience can pick it up even if they're not paying attention. But the second they notice it, they stop trusting everything else you publish.

I use Claude on almost everything I write now, including this email. And nobody has accused me of being a bot yet because I built a system for removing all the slop and making my writing 100% me.

But I've been able to step up my Claude game after listening to Dan Martell interview on Open Residency podcast. He talked about putting his entire life and his business management practices into master context documents, so AI could improve every part of his business.

Dan said to take your secret sauce and save it as a document. Then give your AI access to it.

A system that becomes a second brain and helps me use all of my expertise to its fullest? That sounded like a genius idea.

It also sounded like a lot of work that might fail like dozens of my other AI experiments so I started small with a master content doc.

The first thing I did may sound dumb, but it's the most effective thing I've ever tried: I asked Claude to teach me how to use Claude.

I typed, "I want you to do research on how to use Claude to create content. I want to know what the features are, how to set them up, and how to make Claude sound like me and use my expertise instead of its own."

Then I gave it my philosophy: Readers want to know what the expert knows, not what Claude knows.

It came back with a full system for me to build. I had Claude turn it into a workbook for you to use, but here's the short version:

Part 1: The knowledge gap flag

AI slop is produced when AI doesn't know something you know. It tries to be helpful by filling the hole with generic filler that could apply to anyone. It is predicting what it thinks an amalgamation of humans would say in a similar sentence.

So I gave Claude a standing rule: if a draft needs an example, a number, or an opinion I haven't given it, it can't invent one. It has to insert a flag instead. [NEEDS EXAMPLE]. [NEEDS NUMBER]. Every flag is a question only I can answer. I answer it, the answer goes into a document, and that hole never opens again (usually).

A draft with five flags beats a draft with five invented generalities. The flags also show me what knowledge needs to be added to my master context doc.

Part 2: The interview process

Claude interviewed me, one question at a time, to build a document that holds what I actually know. I gave it as much of me as I could. My positioning. My client results with real numbers. The common advice in my industry I tell clients to ignore.

Here's the instruction I gave it: "If my answer could come from any competitor's website, tell me and ask the question again." (I stole this from Dan too. He calls it reverse prompting: tell the AI to ask you the questions you don't know to ask yourself.)

Build this however you do your best thinking. For me, it was a 2 hour typing session on Claude's desktop app. Others talk through their whole first version on their phone while on a walk. What you're trying to capture is the way you explain your expertise when someone is pushing back (like having a professor make you defend your term paper).

Every draft I write now pulls from that document instead of from Claude's generic brain.

Part 3: The banned behavior list

I gave Claude the exact constructions that scream "AI wrote this" and told it to rewrite every one as a direct claim:

• "It's not X. It's Y."

• "That changed everything." / "Here's the shift."

• The triple fragments. No fluff. No filler. No soul.

Every draft gets a second pass where Claude audits itself against the list and shows me what it changed.

The result: drafts arrive sounding like me, carrying my stories and my numbers, with flags marking where my knowledge goes that isn't in the master document yet.

I didn't set up this system to save time. I use Claude to organize my thinking, and the draft gives me a foil to work off of.

I know my working style isn't for everyone (or maybe anyone). You probably want to produce more content from your expertise, faster. And this system will help you do it.

TL;DR

Three rules keep AI from producing slop:

1. Forbid hallucination. Knowledge gaps get flagged, instead of invented.

2. Interview yourself. Your knowledge lives in documents Claude must draft from.

3. Ban the tells. List the constructions and audit every draft against the list.

t

PS. I assembled everything Claude taught me into a step-by-step workbook. It's the same setup Claude wrote for me to make my master context docs. It's got everything I use, from the project instructions to the interview prompts to turning your old client call transcripts into a knowledge library. Download it here.