The most common complaint about AI writing tools is that the output is obviously AI. That is a solvable problem, and the fix has less to do with better prompts than with what the model is shown before it writes.
Three things give a generated post away. Sentence length is too uniform, because models regress toward an average rhythm no person actually writes in. Openings fall into a small set of learned formulas, so thousands of posts start the same handful of ways. And the vocabulary drifts toward words that test well in general rather than words the writer actually uses. Readers rarely name these individually. They just feel that something is off.
Authorship attribution research has shown for decades that the words identifying a writer are not their topic words. They are the small closed-class words nobody chooses consciously: the, and, but, I, you, of. The rate at which someone uses those, how often they open a sentence with a conjunction, and whether they contract are more identifying than any subject they write about. A tool that copies your topics but not these habits will always read as an impostor.
A writer repeats themselves hardest at the two edges of a post. How they enter a thought and how they leave it are habits, and they are also the two places generated writing most often slips into stock phrasing. If a draft opens with a formula you would never say out loud, nothing after it will recover the voice.
Asking a model to write in your voice does almost nothing on its own, because the instruction carries no information about how you write. Showing it several of your real posts works considerably better. Showing it your real posts alongside an explicit description of your measured habits works better still, because the examples demonstrate the voice while the description names it in terms the model can act on.
Planiberry syncs your published LinkedIn posts and measures the parts of your writing you do not consciously choose: average sentence length, whether you open in lowercase, how often you start on a conjunction, your contraction rate, your punctuation and line-break habits, and the phrases you repeat across posts. That profile goes into every draft together with five of your real posts, rotated so different ideas draw on different evidence. Nothing reuses an old opening, because recycling a real first line is the one thing your own audience would notice.
Because they are generated from similar prompts against models tuned toward an average register. Without evidence of how a specific person writes, the output converges on the same sentence rhythm, the same opening formulas, and the same vocabulary.
Not on its own. The instruction contains no information about your actual writing. What works is showing the model real samples of your writing and naming your measured habits explicitly alongside them.
Planiberry syncs your recent posts and uses five per generation, rotating which five so different ideas see different evidence. The measured style profile is computed across every synced post, so it stays stable while the samples rotate.
No. Your posts are used as evidence of how you write, not as content to reuse. Topics, opinions, and phrasing from your samples are explicitly excluded, and no draft reuses a previous opening line.