Every AI-written post sounds like the same person: even cadence, tidy three-part lists, a closing question. Technically fine, completely forgettable.
The problem isn't the model. It's the prompt target. Ask a general LLM to "write a post about X" and it returns the statistical average of everything it has read. The average of everyone reads like no one.
What fixed it for me was changing what the model imitates. Instead of "write like an expert on X," the instruction becomes "write like this person," with a sample of my own posts as the reference.
The workflow:
- Pull a batch of your old posts that actually sound like you.
- Have the model study them before it drafts, not just the topic.
- Read every draft out loud. If it doesn't sound like you, it's still averaging.
- Edit in your own words. Aim for a faster you, not a replacement.
I built this into XreplyAI: it reads your post archive and models your style across six dimensions (formality, expression, density, humor, assertiveness, abstraction), then drafts in that profile. BYOK, so you bring your own API key and pay your provider directly.
The engineering insight that surprised me: generation was easy, voice capture that survives editing was the hard part. Off-voice drafts get rewritten; on-voice drafts get tweaked in seconds.
Top comments (0)