I tried to clone a top LinkedIn creator's voice with my own AI tool. It failed in a way that taught me more than a month of feature-building

Quick context: I'm building Postessia — an AI tool that clones someone's LinkedIn writing voice from a few sample posts, then generates new posts in that voice. Launched July 1.

Last week I ran a stress test on my own product. Fed it 3 posts from a creator with 600K+ followers, known for a very specific writing style — short punchy hooks, personal failure stories, contrarian reframes. Wanted to see how close the clone would get.

The output was polished. Grammatically clean, well-paced, genuinely "good writing" by most standards.
But it wasn't him. It was a formula extracted from him.

Every single post came back with the same skeleton: hook → vulnerability story → "I made a rule" pivot → one-line paragraph closer → soft CTA. Same shape, different topic, five times in a row.

That's when it clicked — the model wasn't learning voice. It was learning structure, and mistaking structure for voice.

With only 3 sample posts, there wasn't enough variation for the model to separate "how this person writes across different moods/topics" from "the specific pattern that happened to repeat across these 3 examples." So it locked onto the pattern and treated it as gospel.

The fix isn't more prompt engineering. It's three things:
Cap repeat-usage of specific sentence constructions within a single generation (no more "Not X. It's Y." three times in one post)

Force CTA/ending style to vary across multiple generations for the same user, instead of converging on one closer
Push users toward 6-7 sample posts minimum before locking in a voice profile — 3 is technically enough to run the pipeline, but not enough to actually sound like a person instead of a pattern

None of this shows up in a demo. It only shows up when you actually try to fool a discerning reader — which, incidentally, is the exact bar my ICP (LinkedIn ghostwriters, agency owners) holds every output to. They can smell formulaic AI writing from a mile away. If my tool can't survive that test, it doesn't matter how good the demo looks.

Shipping the fixes this week. Posting here mostly because I think most AI writing tools have this exact blind spot and I haven't seen anyone talk about it honestly.

Happy to go deeper on the technical side (voice profile extraction, the blocklist approach) if anyone's curious — building something adjacent or just interested in the failure mode.