Why does AI writing sound generic by default?

Because models train on the average of the internet, and the average is bland. Without specific instructions about voice, models produce prose pitched at maximum legibility for the widest audience — which is exactly the prose nobody actually wants to read.

Which model is best at matching a voice?

Claude 4.7 in our experience — it preserves rhythm and word-choice patterns more faithfully than GPT-5, which tends toward punchier rewrites. Gemini 2.5 Pro is in the middle. Test all three on a sample of your own writing and see which one feels most "you".

Will this work for marketing or persuasion writing?

Yes, with one caveat — the more your voice depends on subverting expectations (irony, deadpan, intentional understatement), the more examples you'll need to feed the model. Generally three to five strong examples plus the constraint list works for most voices.

How to make AI write in your voice (without sounding like AI)

If you've tried to use AI to write in your voice and ended up with bland LinkedIn-flavored prose, the fix isn't a better model. It's three specific things you can do to almost any model that get it to match how you actually write. None of them are the "act as a senior writer with 20 years of experience" patterns from prompt engineering blog posts; those don't work. Here's what does.

Technique 1: Feed examples, not descriptions

The instinct is to describe your voice ("I write conversationally, with short paragraphs, and a slightly dry tone"). This doesn't work because every model interprets "conversational" and "dry" differently.

What works: paste three to five passages of your actual writing into the system prompt, then say "match the voice in the examples above". The model has actual data to pattern-match against.

Examples of my writing voice:

Example 1:
<300-word passage>

Example 2:
<300-word passage>

Example 3:
<300-word passage>

When you write for me, match the voice in those examples. Don't
imitate them as templates; absorb the patterns and write fresh content
that sounds like the same writer.

Three examples gives the model enough signal. One example isn't enough; ten is noise.

Technique 2: Name the anti-patterns

Models default to anti-patterns you can't avoid through positive instruction alone. Naming them explicitly works better:

Avoid:
- Adverbs ending in -ly (use a stronger verb instead)
- Lists of three (do you really mean three, or did you write three because writing two feels incomplete?)
- "Imagine if you could..." openers
- The phrase "in today's fast-paced world"
- Em-dashes used as a comma substitute (use them only for actual asides)
- Em-dashes used three times in a paragraph (limit one per paragraph)
- "It's worth noting that..." (just say the thing)
- "Furthermore", "moreover", "additionally" (connect ideas with content, not connectives)
- Closing sentences that start with "Ultimately..."

The list reads pedantic but it works. Each entry kills a specific tic that the default model output will produce.

Technique 3: Use the rewrite-in-my-style prompt

When you don't want to write from scratch but the AI's draft is too generic, this prompt brings it home:

Below is a draft. Rewrite it in the voice from the examples in my
system prompt. Keep the structure and the points; change only what's
needed to make it sound like the writer in the examples.

Don't add new points. Don't remove points. Don't change the order.

Draft:
<paste draft>

The "don't add, don't remove, don't reorder" constraints are critical. Without them, the model uses the rewrite as cover for adding pet improvements, and you end up further from your voice, not closer.

The model that matters

In our testing, Claude 4.7 preserves voice better than GPT-5 — it's more conservative about not introducing changes you didn't ask for. For voice work specifically, Claude is the default. (If you write punchy short-form like ad copy or headlines, GPT-5 may actually serve you better — it leans punchier by default.)

A test for whether your voice is actually coming through

The acid test: take a paragraph the AI wrote in your voice, paste it into a notes file with five paragraphs of your real writing, and re-read the file two days later. Can you tell which one the AI wrote?

If yes, the voice isn't matching. Update your examples (the existing ones might not be strong enough) and try again.

If no, you've got it. Most readers reading this post will never get there because they'll stop at "the AI's draft is fine" — which is different from "the AI's draft sounds like me".

Where this fits

This pairs naturally with the portable system prompt template — put your voice examples in the system prompt and they travel across every model. Try oran.chat free if you want one instruction set across GPT, Claude, and Gemini. More writing-adjacent playbooks in Playbooks.