Your AI drafts sound generic because your prompt is adjectives

Robbie Haas · 4 min read

By the end of this you'll have a written set of rules that makes AI writing in your own voice possible, built from pieces you already published. You write the rules once, and they keep paying out every time you draft.

What you'll need

Five pieces of your own writing that sound right to you. Posts, newsletters, a conference talk you wrote out, a long client email you were proud of. They don't have to be your best-performing pieces. They have to be the ones where you read them back and think, yes, that's me.

You'll also need a document to write your rules in, and the AI tool you draft with.

Pick five pieces and read them out loud

Read each piece aloud from start to finish, because reading aloud catches what silent reading skips: where you rush, where you pause, and where a sentence runs long on purpose.

This matters because your voice is made of mechanics a reader can hear: sentence length, how you open, how you close, and the transitions you reach for without thinking. Adjectives like “warm” or “punchy” describe the effect. They don't tell anyone how to produce it.

The common mistake is picking your five best pieces by results. Pick the ones that sound like you talking, because engagement tells you what landed, and voice is how it landed.

Write down the mechanics

Go through all five and note what repeats. How long do your sentences run on average, and how often do you break that pattern with a short one? Do you open with a claim, a scene, or a question? Do you close with a challenge, a summary, or a flat line that undercuts your own point? What words do you never use? That last one is often the sharpest signal you have.

Specifics beat impressions here. “I use short sentences” is useless to a machine. “Most of my sentences are short, and I follow a long one with a very short one” is a rule a writer could follow. That's the standard: could a competent freelancer draft in your voice from this document alone? If not, you haven't finished.

The mistake to avoid is describing your personality. Nobody can draft from “direct but approachable.”

Give the tool rules and real examples

Give the tool your rule document and paste in two or three of your five pieces as examples. Rules tell it what to do, and examples show it what the output sounds like when the rules are followed, so you need both.

A prompt built from three adjectives gets you the statistical average of everything already published on the topic. That's why the drafts you've thrown out sound like everyone else's posts. You asked for “professional and conversational,” and so did nearly everyone else.

The common mistake is describing your voice to the tool once and hoping it sticks. Give it your rules and your text every time, or use a tool that stores them for you.

Check the draft before you edit it

Read the draft with your rule document open next to it. Check two things first: how often the same sentence shape repeats, and whether the reading level matches the tone you set. Repetitive rhythm is what makes a draft feel machine-written even when every individual sentence is fine.

Then check your never-use list. If a banned word made it through, the rule needs work.

Do this before you start editing. Once you're rewriting line by line, you lose the ability to see the pattern, and you'll fix the same problem again next week.

When a draft misses, change the rule

Find the rule that let the miss through and rewrite it, then fix the draft.

Editing the draft alone means you make the same fix next time. Editing the rule means the next draft starts closer, and the one after that closer still. After a few rounds you're editing for judgment rather than for voice, which is the only editing worth your time.

The mistake here is adding rules without cutting any. A rule document that keeps growing gets ignored by the tool and by you. When you add a rule, check whether an older one is now redundant.

Common questions

How do I train AI on my writing style?

Pick five published pieces, note what repeats in sentence length, openings, closings and transitions, and write those patterns as rules a person could follow. Feed the rules plus two or three full pieces into the tool you draft with, and check each draft against the rules before you edit it.

Will people be able to tell an AI draft is AI?

They'll tell when the rhythm is uniform and the phrasing is generic. A draft checked against your own rules, then edited by you, reads like you. The tell is sameness, and sameness starts in the prompt.

Why do my AI drafts sound like everyone else's?

Because a prompt made of adjectives returns the average of what's already published. “Professional, conversational, authoritative” describes most business writing on the internet. Specific mechanical rules and your own text as examples are what pull the output away from that average.

Does this work with the AI tool I already pay for?

Yes. Your rule document and your sample pieces improve the output of whatever you draft with, and they move with you if you switch. Checking every draft against the rules by hand is the slow part, and it's the part worth handing to a tool that stores your voice and runs the checks for you.

What if I don't have five pieces I like?

Use three, or use writing that was never published. Long emails, a talk you scripted, a proposal. What matters is that it sounds like you when you read it aloud. You can add pieces to the set as you write more.

Deciding whether to switch tools? Test your AI writer against your own writing first.