The Genre Problem - Why AI-Generated Social Media Posts Sound Like LinkedIn Thought Leaders
The Genre Problem in AI Social Media Content
Version 1 of every AI social autoposter sounds the same: breathless enthusiasm, thought-leader cadence, bullet points with emojis, and sentences that start with "Here's the thing." It is the default genre of LLM-generated social content, and it is immediately recognizable as artificial.
The fix is not better prompting. It is anti-pattern rules.
Why Agents Default to Corporate-Speak
LLMs are trained on enormous amounts of social media content. The most represented style in that training data is the performative professional post - the kind optimized for engagement metrics rather than genuine communication.
When you tell an agent to "post about our new feature on Twitter," it reaches for the most common pattern it has seen: hype language, manufactured urgency, and a call to action. Not because it thinks that is good writing, but because that is the statistical center of its training distribution.
Anti-Pattern Rules
The most effective way to fix agent voice is not to describe what you want. It is to explicitly forbid what you do not want:
- No emoji in professional posts
- No sentences starting with "Here's the thing" or "Let me tell you"
- No manufactured urgency ("You NEED to see this")
- No engagement bait questions ("What do you think?")
- No bullet-point listicles disguised as insights
- No more than one exclamation mark per post
These negative constraints push the agent away from default patterns and toward something that sounds more like an actual person.
Voice Calibration
Beyond anti-patterns, give the agent examples of posts you have written yourself. Not a style guide describing your voice abstractly - actual posts. Five to ten examples are enough for most agents to pick up your cadence, vocabulary choices, and sentence structure.
The combination of "do not sound like this" rules and "do sound like this" examples produces content that passes the authenticity test far more often than either approach alone.
The Authenticity Threshold
The goal is not perfection. It is passing the five-second scroll test - would someone who knows you pause and think "that does not sound like them"? If the answer is no, the agent has done its job.
Most AI-generated social content fails this test not because LLMs cannot write well, but because nobody told them to stop writing in the default genre.
Fazm is an open source macOS AI agent. Open source on GitHub.