How to write a voice guide so your AI doesn't sound like marketing
To keep AI-written content from sounding like marketing, give the model a voice guide in its system prompt with specific, checkable rules: it has to take a position, it has to include one vivid detail per section, certain hype words are banned outright, filler words are rationed, and personalization has to happen without being announced. Those are the rules we used for the AI-written trip-prep emails at Pangea. The test we held them to: the reader should want to bookmark the email and never think it looks like marketing.
Make it have an opinion. "Skip the touristy one on the corner, the counter two doors down is better." A clear opinion is what you get from a friend who's been there, and it's what you usually don't get from a model. Without this rule, the model hedges and lists everything.
One vivid detail per section, tied to something you can act on. A Kyoto email should say "the 6am light on the torii gates, before the tour buses," not "a magical sunrise." The first one tells you when to set an alarm. The second one tells you nothing.
Ban the hype adjectives. "Incredible," "magical," and "unforgettable" are banned. If a model is allowed to call something magical, it will call everything magical.
Ration the filler. This one surprised me most. "Worth" is capped at once per email, because a guide where everything is "worth a visit" is really recommending nothing. Find the mushy word your domain overuses and put a number on it. We also let an email be short when a trip doesn't need much. Length isn't the goal.
Personalize without saying so. One of our travelers is vegetarian. Her city guide is full of recommendations that factor that in, and it never says why. No "since you're a vegetarian, we've included…" Announcing the personalization is the fastest way to make it feel like a mail merge.
These translate directly to GTM work. If you're using AI for outbound, landing pages, or customer emails, the same five rules apply: a point of view, one concrete detail, a banned-words list, a cap on your category's filler word ("streamline," "seamless," "solution"), and personalization that shows up in the substance rather than the first line.
Source: originally published on The Agent GTM Newsletter.