The temptation is obvious: dump your last 100 posts into an AI, let it learn your voice, ship faster. The result is usually a tool that's great at writing the posts you wrote 18 months ago, and bad at writing the posts you want to write next.
Voice profiles overfit when the inputs over-weight history. Old posts encode old positioning, old vocabulary, and old goals. If your ICP has tightened or your value prop has sharpened, the AI is now optimizing for a version of the brand that no longer exists.
The fix is to separate "voice" from "voice history." Voice is five inputs: tone keywords, sentence style, vocabulary level, signature phrases, persuasion mode. Curate those manually from your best 10 recent posts, not your most-engaged 100 historical posts. Then add the current ICP and value prop alongside the voice profile.
When the AI generates, it should sound like the version of you that's selling the current product to the current buyer, not the version that was building an audience two years ago. Refresh the voice inputs every quarter, the same way you refresh positioning.