A friend of mine and I both spend about 12 hours a day inside AI tools. When I sent him some of my automated notes and articles, he said: "These look good. But I can tell it's AI."
One line. That was the wake up call.
The Tells You Probably Haven't Noticed
Once you spend enough time with these models, you start seeing the patterns everywhere. The way they structure arguments is too clean. Too logical. Every sentence flows perfectly into the next.
They say "you" and "everyone" when a real human would say "I experienced this specific thing." They never trail off. Everything is too perfect.
That perfection is the actual problem. And most people building AI content pipelines have no idea.
Why "Perfect" Gets You Caught
Real humans are not consistent. We contradict ourselves. We ramble. We write "mannnn" when making a point. We leave things unfinished and circle back.
AI does none of that naturally. It's been trained to be clean, structured, and logical. Great traits in a technical document. Death for anything that's supposed to sound like a real person talking.
So I built the slop filter. Banned phrases hardcoded across every pipeline. No "Here's the thing." No "What this means is." No dramatic fragments. No rhetorical setups that read like a LinkedIn ghostwriter template.
Engineering Imperfection on Purpose
Banning phrases is the surface layer. The deeper move: deliberately engineer human imperfection into AI output.
Think about UGC video. When a real person films a product review, the camera is shaky. Slightly off center. Lighting is inconsistent. That messiness is what makes it feel real. Brands learned to manufacture that messiness on purpose. They make it look worse to make it feel more authentic.
Same principle applies to writing. Tell the model to write the way you'd message a friend at 2am. Give it your specific verbal tics, your sentence length patterns, the words you overuse. Let it be imperfect in the exact ways you're imperfect.
ZeroGPT and similar tools catch obvious AI writing today. Once you engineer the imperfection layer properly, those detectors lose confidence fast.
How the Stack Actually Works
Here's the breakdown for builders. The voice layer is not a single prompt. It's a three layer architecture:
Layer 1: System prompt (static persona config). Hardcoded identity. Voice notes, banned phrases, sentence rhythm preferences, examples of good and bad output. Never changes between runs. Think of it as approximate fine tuning through instruction.
Layer 2: Content type router. Short form, long form, email, and social comments each get separate sub configurations injected at runtime. A monolithic system prompt trying to handle all formats is the root cause of flat AI output.
Layer 3: Live correction loop. The live prompt is intentionally minimal, sometimes a single sentence. All heavy lifting happens upstream. When output is wrong, you correct with specifics and fold that correction back into Layer 1 or Layer 2 on the next iteration.
In practice this looks like a growing JSONL or markdown file per voice, versioned and iterated over weeks of active use.
Slop Filter Implementation
Here's what a minimal slop filter config looks like:
{
"banned_phrases": [
"Here's the thing",
"What this means is",
"At the end of the day",
"Gamechanger",
"Dive deep",
"In today's fast paced world"
],
"banned_structures": [
"Rhetorical question opener",
"Dramatic single word paragraph"
],
"voice_anchors": [
"Casual but precise. Like texting a founder friend.",
"Short declarative sentences for emphasis. No dashes.",
"Personal examples over abstractions."
]
}
This config gets injected into the system prompt on every run. Banned phrases are also checked in a post generation step. The LLM gets a second pass to self audit before output is accepted. Anything matching the filter triggers a regenerate with an appended note: "You used a banned phrase. Rewrite avoiding it."
The Counterintuitive Truth About Prompting
Most people think prompt engineering is about adding more detail. More context. More instructions. That's wrong most of the time. More detail often confuses the model more. Less and more precise usually works better.
The real leverage is not in the prompt you write right now. It's in the prepipeline. The system prompt. The hardcoded rules. The structure you've built before you type the actual request. When that layer is set up correctly, the live prompt can be a single sentence.
Why Correcting Beats Regenerating
Most people hit regenerate when they don't like the output. Regenerating resets. Correcting trains. Those are completely different operations and the difference in output quality is not subtle.
When it makes a mistake, say: "Don't do that again. Do it this way instead." You correct it. It learns. You adjust.
The model that sounds like you is not configured in a settings panel. It's built through active correction over time, the same way you'd onboard any new hire who needed to learn your communication style. You don't hand them a style guide once. You correct them in real time until they get it.
Most people never do this because they're in generation mode, not training mode. That's the gap.
Are you running a monolithic system prompt or split configs per content type for your AI pipelines?
Top comments (0)