The 10-80-10 Rule: How to Use AI for Marketing Without Losing Your Voice

I spent last week doing fourteen strategy calls with business owners. Every single one of them could explain their business brilliantly for an hour straight.

But ask when they last emailed their list? Quiet.

One woman hadn't sent a newsletter in six months. A guy with 80+ companies using his software had written 27 blog posts in eight years. Another one sent me a 2,000-word positioning document that was genuinely sharp. His website had zero traffic.

These aren't beginners. They know exactly what to do. They just can't get it out the door.

AI was supposed to fix this. For most of them, it hasn't.

Because most people are using AI wrong.

I. Two Camps, Both Stuck

Camp 1: Full automation. Everything goes through AI. Social posts, emails, ads, customer replies. The output is fast. It's also painfully obvious. Same structure, same rhythm, same "In today's rapidly evolving landscape" energy.

You can spot it in two seconds. So can your audience.

Camp 2: Fancy autocomplete. These people ask AI for a headline, rewrite it themselves, and spend the same four hours on one email they always did. Nothing actually changed.

One camp has speed but no soul. The other has soul but no speed.

I was in camp two for months. Generate a draft, hate it, rewrite the whole thing, wonder what the point was.

Then I figured out a framework that fixed it. I call it the 10-80-10 rule.

II. The Framework

You do the first 10%. Your input. Stories, direction, perspective. The stuff only you can provide.

AI does the next 80%. The entire production chain. Not one task. Everything from drafting to formatting to creating variations to checking compliance.

You do the last 10%. The quality check. One question: does this sound like me?

Your voice bookends the process. AI fills the middle.

Simple idea. The hard part is making the 80% actually good enough that your final 10% is just a quick gut check and not a full rewrite.

That's what this article is really about.

III. Why Most People's 80% Is Broken

Here's the thing nobody tells you. If you open Claude or ChatGPT and type "write me a LinkedIn post about marketing," the output will be mediocre. Every time. Not because the AI is bad. Because you gave it nothing to work with.

No context about who you are. No reference for how you write. No information about your audience. No examples of what "good" looks like to you.

You're asking a brilliant writer who has never met you to write in your voice. Of course it sounds generic.

The difference between people who get slop and people who get genuinely useful output comes down to setup. Specifically, three files.

File 1: The Context File

In Claude Code, this is called CLAUDE.md. It sits in your project folder and loads automatically every time you start a session. Think of it as your business brief that AI reads before doing anything.

Here's what goes in mine:

# Grow On Repeat

## About
Dutch entrepreneur, 22+ years experience. Built and sold 3 companies.
$10M+ spent on Facebook ads personally. Father of 3.

## Voice
Conversational, direct, honest, light humor. Grade 8 or lower.
No jargon, no buzzwords, no fake urgency. No em dashes.

## Current Offers
GrowOS (self-serve) and GrowOS Accelerated (white-glove).

## Lessons
When I give you feedback on output, always add it to lessons.md
so the same mistake never happens again.

## Rules
1. Never fabricate stats, testimonials, or story details.
2. Use documented stories only.
3. Be concise.

This is simplified, but you get the idea. Every time AI creates something for me, it already knows who I am, how I write, and what I sell. That alone eliminates half the "this sounds like a robot" problem.

You don't need anything fancy. Open a text file, write down who you are, what you do, how you talk, and what you never want AI to say. Save it as CLAUDE.md in your project folder.

File 2: The Voice File

This is the one that makes the biggest difference. And most people skip it.

A voice file isn't a description of your tone ("conversational and friendly"). Descriptions are useless. AI interprets "conversational" differently than you do.

Instead, paste actual examples of your writing. Real emails you've sent. Real social posts. Real copy from your website. The more examples, the better the match.

Mine has about 30 real emails I've written over the years. When AI writes something for me, it references those actual emails, not a description of my style.

Here's how to build yours in five minutes:

1. Open your Sent folder in whatever email platform you use (Kit, ActiveCampaign, Gmail, whatever)
2. Find 20-25 emails you're proud of (the more the better, but start with at least 10)
3. Copy-paste them into a file called tone-of-voice.md
4. Add a short note at the top: "Write like these examples. Match the sentence length, the rhythm, the vocabulary, the level of formality."

That's it. You now have a voice reference that actually works.

File 3: The Lessons File

This is the secret weapon. It's a running list of corrections.

Every time AI produces something I don't like, I don't just fix the output. I add a rule to the lessons file. Some of mine:

- Never use em dashes. Replace with periods or commas.
- Don't lean on statistics. Use personal experience instead.
- Hooks must reflect what the reader actually feels,
  not clever metaphors.
- Don't undermine authority in openings. No fake struggles.
- Vary structural patterns. Never repeat the same
  subheading style across sections.

The lessons file is cumulative. It only grows. And because it loads every session, every mistake gets fixed permanently. I never give the same correction twice.

Here's the key part: add a rule to your context file (CLAUDE.md) that says: "When I give you feedback on any output, always add it to lessons.md so the same mistake never happens again." This way the system automatically captures your corrections without you having to manually update the file. It compounds.

After a few weeks of this, the output quality jumps dramatically. It's like training a new employee who actually remembers everything you tell them.

These three files are why my 80% works. Without them, you're starting from zero every single time. With them, you're starting from a system that already knows your business, sounds like you, and has learned from every previous correction.

IV. 141 Ads in 30 Minutes (The Proof)

I didn't figure this out in theory. It came from my product launch last week.

I needed Facebook ads. Not five carefully crafted ones. Enough to test at scale and let the data decide.

My 10%: I briefed the system. Target audience, product details, the key messages I wanted to test, my personal angle. About 15 minutes of voice notes.

The 80%: Claude Code produced 141 ad variations. Not just the copy. The images. The compliance checks against Meta's advertising policies. The bulk upload file formatted for Facebook Ads Manager so I could push everything live in one click.

My hands-on time for 141 ads: about 30 minutes.

My 10%: I scanned the output. Killed a few that didn't sound right. Adjusted some headlines. Uploaded.

The results? 67 purchases tracked back to those ads alone, at $29 average cost per sale.

The single best-performing ad generated 15 sales at $7.96 each. What was this optimized, data-driven creative masterpiece?

"I built this system to run my own marketing."

That's it. Just me being honest. My 10%.

And the ads were just one piece. The entire product launch (sales page, video, 16 training videos, emails, ads, delivery, the product itself) took one week. Solo. No team. A year ago, that would have been a month with multiple people.

V. The Daily Practice

The launch was the dramatic proof. But I use 10-80-10 for everything now.

Newsletters. I record a voice note about what's on my mind. AI drafts it in my voice (it has 30 real emails to reference). I review it, change a line or two, schedule it. Five to eight minutes.

Before this? Two to three hours. Not because the writing was hard. Because the gap between "I have something to say" and "this is published" was enormous.

Social posts. I batch a full week in one session. Quick voice note about themes. Platform-specific posts generated for LinkedIn, Twitter, and Facebook. Kill the ones that feel off. Schedule the rest. Maybe 30 minutes for an entire week across three platforms.

This article. I talked through the framework in a series of voice notes over about 15 minutes. AI produced the first draft. I'm reviewing and editing it now. (Yes, I practice what I preach.)

The pattern is always the same. I start it. AI runs it. I finish it.

VI. How It Gets Better Over Time

Here's what nobody talks about.

First time you run a new workflow, your final 10% is more like 20%. You're fixing things, adjusting tone, correcting assumptions.

That's normal.

But the key is what happens next. Don't just fix the output. Fix the process.

"That paragraph was too formal" becomes a lesson: "Never open a paragraph with a dependent clause. Start with the subject."

"That stat felt forced" becomes: "Don't use external statistics. Lead with personal experience."

I call this process feedback (as opposed to output feedback). Output feedback fixes one email. Process feedback fixes every email from now on.

After five runs of the same workflow, my edit rate drops to almost nothing. The system learned.

VII. The Quality Check (What to Actually Look For)

Your final 10% isn't proofreading. It's an authenticity check. Here's exactly what I scan for:

1. The "would I say this out loud?" test. Read any sentence aloud. If it sounds like a blog post instead of a person talking, rewrite it. AI loves complete, grammatically perfect sentences. Real people use fragments. Start with "And." Say "But" at the beginning of a paragraph.

2. Structure that's too clean. If every paragraph is the same length, every list has three items, and every section follows the same pattern, it's AI. Break the rhythm. One-sentence paragraph. Then a longer one. Then a fragment.

3. Generic enthusiasm. "This powerful framework will transform your marketing" is AI. "This saved me about three hours last Tuesday" is human. Specific beats impressive.

4. Missing failures. If everything in the piece went perfectly, it's AI. Real stories include the part where something broke, didn't work, or surprised you.

5. Em dashes everywhere. This is the number one tell. AI uses em dashes constantly. Replace every single one with a period, comma, or colon. (I have this as rule #1 in my lessons file for a reason.)

If you want a more thorough checklist, there's an open-source tool called Humanizer that runs inside Claude Code and automatically catches AI writing patterns. It checks for things like:

• Rule of three (three items in a row, repeatedly)
• Negative parallelisms ("not X, not Y, but Z")
• Vague AI words: "leverage," "unlock," "empower," "seamless"
• Paragraph length variation (should be aggressive: 1 line, then 3 lines, then 1 line)
• Opening energy (does it sound like a person or a publication?)

I run something similar on every piece before it goes out. It catches things you stop noticing after staring at a draft for five minutes.

VIII. Try It Yourself (Copy-Paste Ready)

Here's the simplest version of 10-80-10 you can try right now.

Step 1: Create your context file.

Open a text file. Write this (fill in your details):

# My Business

I run [your business]. I help [your audience] with [what you do].

## My Voice
[Paste 3-5 real emails or social posts you've written here.
These are more valuable than any description of your style.]

## Rules
- Never use [words/phrases you hate]
- Always [preferences you have]
- My sign-off is: [however you close emails]
- When I give you feedback, add it to a lessons file
  so the same mistake never happens again.

Save it. In Claude Code, name it CLAUDE.md and put it in your project folder. In regular Claude, paste it at the start of your conversation.

Step 2: Your 10% input.

Record a voice note (your phone's voice recorder works fine). Talk for 2-3 minutes about what you want to say. Don't script it. Just talk like you're telling a friend.

Transcribe it (use the free transcription in your phone, or paste the audio into Claude).

Step 3: The 80%.

Paste your transcript into Claude with this prompt:

Here's a voice note I recorded about [topic].
Turn this into a [newsletter / LinkedIn post / email].

Audience: [one sentence about who reads your stuff]

Use the voice examples I gave you. Match my actual writing
style, not what you think "conversational" means.

Keep it under [word count]. No em dashes. No corporate
language. If I wouldn't say it out loud, don't write it.

Step 4: Your 10% review.

Read the output once. Ask yourself: does this sound like me? Change what doesn't. If you're rewriting more than 20%, your input wasn't specific enough. Go back to Step 2 and give more context.

Step 5: Ship it.

The whole cycle: under 10 minutes.

After you do this a few times, start a lessons file. Every time you fix something in the output, write down the rule so you never fix it again.

That's 10-80-10.

You start it. AI runs it. You finish it.

Talk soon,

Wilco