Key Takeaways
- A dedicated style-learning instance that reads your existing work is more accurate than trying to describe your voice in a system prompt.
- Skills created from completed style-matching sessions become reusable procedures for future content.
- Two-instance architecture separates concerns: one instance works, another learns.
The Problem With Describing Your Own Voice
Most AI-assisted writing workflows start with 'write in the style of X'. The problem is that people are poor at describing their own style — they describe what they aspire to, not what they actually do. Examples are more accurate than descriptions.
The insight here is to skip the description step entirely and have the agent read the body of work directly.
The Two-Instance Architecture
A second Hermes instance is given a focused mandate: read all published articles, internalise the patterns, and use that understanding to generate drafts for new topics. The first instance continues its normal work. The second builds and maintains the style library.
“had written a procedure for reading my published articles”
Skills as Reusable Procedures
Each time the style-learning instance successfully generates a well-matched draft, it saves the procedure as a reusable skill. Over time, this skill library grows to cover different content types — LinkedIn posts, long-form essays, threads — each with the author's voice locked in.
The result: requesting new content no longer requires re-explaining style preferences. You ask, it delivers.
Story sourced from the official Nous Research Hermes user-stories page. Original author: @Saboo_Shubham_.