Human-AI Collaborative Writing Converges on One Voice

swapnoneel123 · reddit · 2026-07-15

The author shares their journey of using LLMs to mimic a personal writing style: after early failures with prompts and style hints, they provided the model with extensive personal background and writing samples, letting it first reverse-engineer a "voice profile."

They then shifted to a human-in-the-loop process: the model drafts, the author edits, and the model learns from the edits. For instance, after deleting long intros twice, "no long intros" became a固化 rule. Paired with a dedicated "kill list" to scrub common AI phrasing, output quality improved significantly, with the first few articles requiring minimal edits.

However, by the sixth piece, a problem emerged: every article became too "perfect," adopting the exact same rhythm and structure regardless of topic. The author reflects that this is a system that only adds constraints without introducing variation—exploiting without exploring—and asks if a better solution exists.

Original post →

More from Apps

Apps channel →