Workflow 012: The Voice Skill

I ask my own assistant whether I wrote my own LinkedIn posts.

Sometimes I did. Often I didn't. I usually can't tell.

A voice skill is trained on your own best-performing writing so AI produces drafts that sound like you rather than like a competent stranger. Mine is paired with a second skill whose only job is catching anything that reads as AI. Together they let my assistant run my content while it still sounds like me.

This one took the longest to get right, and I resisted building it hardest.

What this replaces

Being the only person who can write as you.

I started working online as a ghostwriter. Studying how other people wrote and producing their voice instead of my own was literally the trade. So handing my voice to a machine felt like handing over the thing I'm actually good at.

But there was a problem underneath that.

I'm excellent at starting. I get a surge of ideas, execute brilliantly for about ten days, and then the follow-through gets harder. Consistency has never been my strength.

So I needed the system to be consistent instead of me.

Blueprint

Train on your best work. Write with it. Check it with something else.

Trigger: manual, whenever a post is needed. Inputs: an idea, a line, or a story, plus the voice skill trained on your past writing. Output: a draft that sounds like you, run through a quality check, ready for a human to approve.

Two skills, deliberately separate:

• The writer skill, trained on my strongest posts, my background, and my stories
• The quality checker, which reads the draft afterward looking for anything that reads as AI

Splitting them matters. The writer holds voice, the checker holds standards, and I can update either one without breaking the other.

Organize

What you need before training anything:

• Your own writing, from wherever you've published consistently
• Performance data, so you can filter for what actually landed
• Your background and the stories you tell often
• A list of your tells: words you use, words you'd never use
• A person who knows your voice well enough to catch a miss

I pulled five years of LinkedIn and Facebook posts and kept the ones that got real responses. Not everything I'd written. The ones that worked.

That filtering step is what people skip. Training on everything you've ever posted teaches it your average. You want it to learn your best.

Systematize

The writer skill:

You are my writing voice. You write posts that sound like me, not like AI.

Attached: my best-performing posts from the last [TIMEFRAME], my
background, and the stories I tell often.

How I write:
- [YOUR RULES: sentence length, structure, how you open, how you close]
- Words I use: [LIST]
- Words I never use: [LIST]
- My recurring stories and references: [LIST]

When I give you an idea, a line, or a story:
- Write it the way I'd write it, using the attached posts as the standard
- Use my real stories and specifics. Never invent an example I didn't tell you
- Match the structure of my posts that performed best
- If you don't have enough from me to write it truthfully, ask instead of
  filling the gap

Never write anything I'd have to correct for accuracy. Sounding like me
matters less than being true.

The quality checker, run on every draft afterward:

You are my quality checker. Read this draft and flag anything that reads
as AI-generated.

Check for:
- Em dashes
- "It's not X, it's Y" constructions
- AI filler: delve, leverage, unlock, game-changer, seamless, elevate,
  robust, dive deep, in today's world
- Sentences that sound polished but say nothing
- Claims I didn't make and can't verify
- Any tone that isn't mine: too formal, too corporate, too eager

Rewrite the flagged parts in my voice, or tell me what needs a human.

Then say plainly whether this passes or not.

How to set it up

1. Gather your best-performing writing, filtered by what actually got responses.
2. Build the writer skill and test it. The early drafts will be obviously AI. That's normal.
3. Give specific feedback and update the skill each time. Not "make it better," but "I'd never say this, here's what I'd say."
4. Build the quality checker separately, and add every tell you spot to its list.
5. Keep updating the checker forever. Mine is still growing.

Why do em dashes matter?

Honestly, I have nothing against them. I've used them for years.

But enough people now read an em dash as a sign of AI that it costs credibility, so my checker strips them. That's not a writing rule, it's a reader-perception rule, and those change. Which is exactly why the checker is a separate, editable file.

Assign

The AI writes. A person decides whether it's good enough to publish.

My assistant runs both skills and posts to LinkedIn, mostly without me seeing it first. That's not carelessness, it's four years of working together. She knows my standards well enough to tell when something wouldn't pass.

And she brings something the model can't. We talk every day, so she knows things about my life that were never written down anywhere. She'll catch something I said offhandedly and turn it into a post. The stories in my content are often stories I told her, not stories I published.

The AI holds the pattern. She holds the context.

Rule of thumb from this build: train AI on the work, keep a human on the standard.

There's a version of this that goes badly, which is handing over your voice and stopping paying attention. What makes it work is that someone who knows your voice is still the last check before anything goes out.

I still write plenty of my own posts. This just means the ones I don't write still sound like me.

Integrate

The voice skill is the layer under everything I publish:

• The repurposing engine calls it, so every platform's version sounds like me
• Every draft runs through the quality checker before a human sees it
• The checker gets updated whenever a new AI tell appears in the wild
• My assistant runs both, reviews the output, and publishes

That's how I keep showing up when I'm sick, traveling, or in a stretch where follow-through is hard.

The system is consistent. I don't have to be.