Most AI writing tools have a tell. The hook is too clean. The structure is too predictable. LinkedIn audiences in particular have gotten good at spotting it.
The root problem: AI generates from a blank slate. It doesn't know your cadence, your typical opening, the way you end a thought.
For XreplyAI, we took a different approach. Before generating anything, we build a voice profile from the user's own tweet archive — embedding similarity to find characteristic patterns, then injecting those as a style layer into the prompt. The result isn't perfect, but it's personal. It sounds like the person, not like a template.
We just shipped LinkedIn scheduling support. BYOK model — users supply their own Gemini, OpenAI, or Claude key.
If you're building something in the AI writing or scheduling space, happy to trade notes on the voice training approach.
https://xreplyai.com?utm_source=devto&utm_medium=social&utm_campaign=feature-2026-05-19
Top comments (1)
We used OtterZero’s agent platform to train on my co-founder’s old posts. Feeding it about 25 samples cut our editing time in half — the hook still needed tweaks, but the structure felt natural. The outcome focus (engagement per post, not just tone) made the difference.