Playbook
AI replies that sound like you — not like ChatGPT
Why we fine-tune per-user instead of using prompt engineering, and how the Autocloz reply-drafter learns your voice from your last 200 sent replies.
2 Apr 2026 7 min readBy Autocloz Editorial, AI team
Prompt engineering hits a ceiling
A good prompt buys you a polite, generic reply. It won't buy you "this reads like Vikram wrote it, not ChatGPT." No amount of prompt-tuning closes that last gap.
Per-user fine-tuning
Autocloz trains a small adapter on your last 200 sent replies. The base model stays shared across everyone; the adapter is yours alone. Your drafts pick up your sentence length, the openers you reach for, the verbal tics you don't notice you have.
Always human-in-the-loop
The AI never sends on its own. Every draft waits for your click. That's deliberate: the speed comes from skipping the blank page, not from cutting you out of the loop.