I automate a chunk of my social content, and the biggest quality jump came from a boring change: fewer instructions, more examples.
Instruction tuning ("witty, conversational, no corporate speak") converges on the same voice for everyone, because everyone writes roughly the same adjectives. The model regresses to the mean of its training data with a light stylistic filter on top.
What actually moved the output: few-shot with my own archive. Around 50 published posts in context gives the model concrete evidence of sentence length, opener patterns, question frequency, and how blunt I am. Drafts went from rewrite-everything to light edits.
Practical notes if you build something like this:
- Curate the examples. Your 50 best posts, not your last 50.
- Refresh them as your voice drifts.
- Judge output on consistency across a month of posts, not on any single draft.
I built this pipeline into XreplyAI (a scheduler for 15 platforms with drafts trained on your own post archive): https://xreplyai.com?utm_source=devto&utm_medium=social&utm_campaign=edusales-2026-07-16
Top comments (0)