What is magic-distillation?

This repository, called fiction-distiller, is a skill for AI coding assistants that analyzes the writing style of a piece of fiction and then uses that style to write a brand new, original story. It works with two agent platforms, opencode and Claude Code. The core idea is simple: you give it a novel or story, it extracts a kind of style fingerprint from the text, and then it writes a completely different story using that same fingerprint. It is explicitly not meant to copy plot, characters, or dialogue, only to carry over narrative rhythm, sentence patterns, rhetorical habits, and sensory focus. Every generated piece comes with an originality statement and passes a self-check before being finished. The knowledge the skill draws on is organized into three layers that load only as needed rather than all at once. A base layer, always loaded, covers general style theory such as narratology, rhetoric, and an author's fingerprint in vocabulary, sentence structure, and punctuation. A genre layer loads depending on whether the text is long fiction, short fiction, or web fiction, each with its own structural concerns. A method layer loads during the actual writing stage and covers feature extraction, fingerprint verification, imitation strategy, and originality assessment. In total there are 32 topic files, each about 3500 characters, loaded in four strict steps so the system never pulls everything into context at once. The overall process runs through six steps: building a knowledge index, receiving and splitting the source text, distilling style across eleven dimensions into a style model card, planning an original story that only inherits style rules, writing that new story from the point of view of the original author writing something new, and finally running a self-check against six hard rules, including never copying ten or more consecutive words from the source or reusing its plot, scenes, dialogue, or proper names. The output of the distillation step is a style model card containing five to eight core style tags, fifteen must-follow style rules, a short style fingerprint summary, and a list of five forbidden writing behaviors. Generated chapters and cards are saved into an output folder per platform. The project is licensed under MIT.