book-to-skill

From a method to a skill

The pipeline turns a Markdown source into an executable methodology. Input: a .md file or a folder of .md files. Output: a skill folder with a thin SKILL.md, numbered reference sheets in references/, and the build record PROVENANCE.md. The pipeline needs no external skills, scripts, or libraries.

The distinction that governs everything else: a reference guide to a book and a skill carrying a method are different artifacts. A guide answers "what does chapter 7 say". A skill answers "how do I act by this method right now". We build the second.

The test at every step: would this line help an agent that works on the user's material and has never read the book? It would, keep it. It only reports what the source contained, drop it.

<master_rules>

Rules that hold across all phases

They are duplicated below next to the places where they apply. That is deliberate.

1. NEVER ship a unit without an anchor. Every extracted unit carries a verbatim quote from the source and a section address. The anchor is always verbatim and in the source language, never translated: a translated anchor cannot be found by mechanical search. The rule covers all five types of catch, glossary terms and cases included, not rules alone. The point: the anchor is the only thing that mechanically separates what was extracted from what was invented.

2. Do not write on the author's behalf. If an obvious rule is not stated in the source, it does not exist. That the rule is true, that it would improve the skill, that the method looks incomplete without it, that it is "clearly implied", none of that is grounds to add it. Under pressure to be complete a model writes on the author's behalf more often than it lies in quotes, and it is harder to notice.

3. Synthesize, do not transplant. Raw source fragments longer than three sentences never reach the output. The author's names for constructs are kept verbatim: "5 whys", not "the method of successive questions".