I'm willing to bet that a good percentage of you had never heard of an 'em dash' before the rise of ChatGPT a few years ago. Now you probably see them everywhere. Not because we as a society have suddenly become better at grammar – but because AI has earned the title of one of the most prolific copywriters in modern times.

Fun fact: the em dash actually originated in 15th-century printing as a dash roughly the width of a capital "M", often used to denote pauses or dramatic shifts in thought.

No doubt you've used ChatGPT, Claude or others to write copy for you in some capacity. For that LinkedIn post, that tricky email, that wedding speech. Open a new chat, give it a one-sentence brief, hit send.

What comes back is usually slop.

It tries to be clever. Reaches for phrases no one I know would say out loud. Uses grammar that's almost too perfect. Uses words like resonate and delve and leverage a little too liberally for my liking.

It's not bad. It's perfectly competent, readable – but unmistakably not 'me'.

It saves me 5 minutes, then costs me 10 editing it to make it sound human.

And it is not the AI's fault. Here is what is actually going on.

These models are trained on billions of examples of writing scraped from across the internet. When you hand one a prompt, what it gives you back is, in effect, the average of all of that writing. By definition. An average LinkedIn post. An average email. An average best man speech at your brother's wedding.

Lacking any real character. A voice. A point of view.

Sure, you can give it a couple of examples to help it along. That certainly produces better results. But it still reverts to the mean by default. You still get the average – just with a thin coat of mimicry over the top.

The fix isn't better prompting. It's better context.

WHAT I BUILT

I built a tool. It is called my AI Editor In Chief. It runs as a simple skill inside Claude Code.

The skill does two things.

First, it guides me through building my personal voice fingerprint. A file containing every rule an AI model needs to not just imitate, but natively speak in my tone of voice. This becomes our bible.

Then, when I want to write something, I call the skill. It reads the fingerprint and drafts against it. I can literally say 'write me a LinkedIn post about my 5-star Uber rating' and it handles the rest - it knows my voice, the rules and patterns I use, my hard bans, my quirks.

The output comes back at 90-95% of the way there, versus the 20-30% I get from the same model without the fingerprint.

The full mechanism is as follows:

WHAT'S IN THE FINGERPRINT?

Seven sections. Written in a language both AI models and humans understand. The skill builds them, but you can open the file any time and argue with what it wrote:

1: Hook Shapes

The openers I actually use. The specific ways I start a piece of writing – segmented by writing type (long form, short form, email etc).

2: Signature characteristics

This is the juicy stuff. The rhetorical patterns that keep showing up across my writing. The contrast-and-reframe move. The "and it's not your fault" empathy line. Triple-parallel lists of named things.

3: Rhythm

Sentence-length variance. Paragraph-length patterns. Where I use short fragments. Where I don't.

4: Vocabulary

The words I reach for. The words I never use.

5: Hard Bans

Specific words and phrases the model must never (or rarely) use. The list of things I keep deleting by hand when I edit AI output. And yes, the dreaded em-dash.

6: Closing Shapes

How I end things or sign things off. Short? Decisive? No essay-end flourish.

7: Receipts rule

The level of specificity the writing has to hit. Named things, real numbers, concrete file names.

Every rule comes with a citation back to the pieces in the source material where the pattern turned up. I can open the file, disagree with a rule, delete it, rewrite it, add one the extractor missed. That argument with the file is what makes it mine – even though I had AI write it for me in the first place.

Why not just write these rules yourself?

It's the obvious question. Sit down with a blank doc, describe your own voice in plain English, hand the result to the model.

I tried it. Doesn't work.

On my first attempt at this I wrote the fingerprint file manually. Hours of back and forth, describing in painstaking detail. No matter how hard I tried, I just could not get the AI output to sound like me.

The problem was that I was writing the rules in a way a human would understand, when what I actually needed was to write them in a way an LLM would understand.

That distinction is really important. It is also why using AI to instruct AI is one of the strongest applications of AI right now. The model already knows what shape of rule it can & can't follow. You do not.

So instead of writing in your language and hoping the model translates, you let the model write in its own language about you.

Have an LLM write the rules. Then edit them like a human would.

THE SKILL: UNDER THE HOOD

Here's how the skill works under the hood:

Phase 1: Build the corpus

Pick 10 pieces of your writing. 10 really is the minimum, more is better (I pulled 200+). This needs to be your own writing, recent enough to still sound like current-you, across all the surfaces you actually want to write for. If you want the fingerprint to cover LinkedIn posts and sales, the corpus needs examples of both.

Drop the pieces into a folder. The skill walks you through the ingest and flags any sources that look like they might be outliers to the point where the voice is no longer yours. That is your corpus.

Phase 2: Extract

The skill reads every piece in the corpus, finds the patterns that repeat, and writes them into the seven-section fingerprint file. Each rule arrives with a citation back to the specific pieces it was pulled from.

You read the file end to end. Where the extractor got you right, leave the rule. Where it missed, rewrite. Where it named something you had not consciously noticed you do, pay attention. That last category makes the setup worthwhile on its own.

Phase 3: Draft

Give the skill a one-line brief - "draft a LinkedIn post about X", "write a cold email to Y" - and it produces a draft in your voice, shaped by the rules in the fingerprint file it just created.

Every draft comes with a short block at the end showing which rules fired. Which hook shape the opener used. Which signature moves landed, in which paragraph. Which hard bans were held. When you reject a line, you are rejecting the specific rule that produced it. No black box, no "trust me, bro" prompt.

Phase 4: Feedback

By now your output will be good. But this phase is where the compounding happens.

Take a draft the skill handed you. Tweak it to make it sound closer to how you would write. Paste it back. The skill compares your edit against the original, works out what changed and why, and updates the fingerprint on the back of that. A cliché it let through becomes a new ban. A rhythm it got a bit wrong becomes a new rule.

After a month of this, the fingerprint is sharper than when you started. After three months – it's a real asset.

All four phases run inside the skill. It sounds like a lot of work but it's really simple in reality. Beyond the initial setup, it's mostly just you copy/pasting things you wrote into the skill and it handles the rest on its own.

That is the point. It is one system with one interface - call the skill, give it a brief or an edit, keep going. The fingerprint file updates, learns and manages itself automatically.

WHAT I USE IT ON

I use it for a bunch of stuff. It's really useful for long-form LinkedIn posts, threads on X, or the occasional response to a difficult email.

It's also great for website copy. I did a full audit of my consultancy website recently and overhauled nearly every word on the site via this skill.

WHO ELSE IS DOING THIS?

This isn't a new thing. Loads of creatives have built or are using similar copywriting systems.

Nat Eliason - writer, entrepreneur, long-running newsletter operator - has been open about his own AI-assisted drafting workflow. He feeds Claude Opus a corpus of his previous writing, then drafts with the model as his editor. In a recent interview he said the match on his personal voice is now good enough that the model does work he used to pay human editors for.

Same core move as the skill I built: the AI model only sounds like you if it has enough of you to reach for. Nat does the corpus-feeding and voice-teaching in-chat, by hand.

This skill formalises the same logic into a repeatable system - same corpus principle, same feedback loop, just wrapped up so you install it once and keep it sharpening in the background.

Every solo operator I know running an AI drafting workflow at any kind of volume has landed on some version of this. The tools may differ slightly, but the instinct is the same. If you are getting slop out of a model, it is not because you are prompting wrong. It's because you haven't yet shown the model how to sound like you.

WHAT IT WON'T DO

Two things worth naming up front, because they are the cases where this does not help and I would rather be straight with you:

It can't create a tone of voice for you. If you don't have any source material to feed it - then it's got nothing to extract patterns from. You can literally feed it anything: emails, tweets, blog posts - whatever you've got. But it has to be written by you, and in your natural tone of voice.

If you write in two different modes, you need two fingerprints. If the voice you use on X is not the voice you use for enterprise client decks, do not try to average them into one file. Run the skill twice. Two corpuses, two fingerprints, and you point the skill at whichever one matches the piece you are drafting.

WHERE TO START?

If you'd like to get access to this skill, you can get it via one of the links below. This is a clean skill file with no existing fingerprint. On its first run it will guide you through the setup process. Takes about 20 mins.

If you'd prefer, you can also build this yourself. Everything discussed above - the four phases, the seven sections, the edit-feedback loop, the manual-first failure mode, the context-over-prompting reframe - is the playbook. Take it, open Claude Code or your tool of choice, and build this for yourself. It may take you an afternoon to set it up, and a month or so to get it sharp.

Catch you next week!

Sam

P.S. Same standing ask as Issue #001. If you are a creative professional running an AI workflow I could feature - designer, editor, studio owner, freelancer - drop me a line at hello@madewithmachines.com. Every issue needs a case study. Yours could be Issue #003.

​ ​