When users sign up for my LinkedIn tool, the first request I get is "make it sound like me." Generic AI output sounds the same across every account, and people can tell. The model writes the way training data writes. To match a specific person's voice, you have to either fine-tune or feed examples in-context. Fine-tuning is expensive, requires hundreds of samples, and locks you into one model. Examples in-context cost nothing extra and work across any frontier model. Five sample posts is the threshold I have found where Claude (Sonnet 4.6 in my testing) reliably matches sentence rhythm, word choice, opener style, and ending pattern. Three samples is too noisy. Ten samples does not measurably improve over five, and it eats context. Here is the prompt structure I landed on after a lot of iteration. The system prompt You are writing a LinkedIn post in the user's exact voice. The user's writing style is defined by these five sample posts.…