Fine-Tuning Your Voice Model
Once you've generated a few posts, you can give feedback to improve accuracy. This guide covers the feedback loop.
Understand what the feedback loop does
Every time you generate a post, you have the option to rate the voice accuracy (1–5 stars) and add a note about what was off. This feedback doesn't retrain the model from scratch — it adjusts the weighting on specific patterns. If you consistently mark openings as too formal, the model learns to weight your more casual opening patterns more heavily.
Rate your generations
After each generation, click the star rating below the draft before copying. A 5-star rating means 'This sounds exactly like me.' A 1-star means 'This sounds nothing like me — please weight this pattern less.'
You don't need to rate every generation. Rating 3–5 per week is enough to see consistent improvement within 2–3 weeks.
Tip: Be honest. Giving 5 stars to a generation that you heavily edited sends the wrong signal and slows improvement.
Add specific feedback notes
The star rating is quantitative signal. The note field is qualitative signal. Some examples of useful notes:
'Too many em-dashes — I use them sparingly' 'Conclusion was too definitive — I usually end with a question' 'Opening was too aggressive — I prefer to start with a story' 'This is perfect — keep the short-sentence rhythm on the close'
The note doesn't need to be long. One specific observation is worth more than a paragraph of general feedback.
Add more training posts
The fastest way to improve voice accuracy is usually to add more training data, not to tune the existing model. After 2–3 weeks of use, go back to your LinkedIn or X profile and find 5–10 more posts to add. Your more recent posts are particularly valuable — they reflect your current voice, not who you were 3 years ago.
In the Voice Settings tab, click 'Add training posts' and upload the new content.
Adjust explicit preferences
In Voice Settings, you'll find explicit controls that override the model's defaults: punctuation style, sentence length preference, opener style (statement / question / story), and closer style. These are high-level overrides, not replacements for the trained model.
If there's something specific the model consistently gets wrong — you never use exclamation points, you always use Oxford commas — set it explicitly here.
Compare before and after
After 30 days of feedback, generate a post on the same topic as one of your early training posts. Compare the two outputs. The difference is your fine-tuning working. If you don't see meaningful improvement, review whether your feedback has been consistent — conflicting signals (sometimes rating casual openings highly, sometimes marking them negatively) slow the improvement loop.