technical-prose-stylelisted

# Technical prose style For anything a human reads: `docs/`, README files, code comments, commit messages, release notes, issue and pull request text. ## What this is for Unedited LLM prose is tiring to read. It is usually accurate, and the reader usually knows it was machine-drafted and minds that far less. What wears them down is the shape of it. One construction returns over and over. Every fact trails a clause explaining why the fact matters. A fresh synonym arrives where the previous term would have done. The goal here is prose that costs the reader less. Concealment is a different aim, and out of scope. An independent detector still identifies text that follows every rule below as machine-written. Six rewritten documents were put through one to check. See [Limits](#limits) at the end. The rules below come from measurement. 192,000 words of LLM-written technical documentation were compared against 66,000 words of Django, Go, Rust and Python documentation, and only the patterns where the two differed by a factor of two or more were kept. The six that survived run between 2.3 and 8.3 times the human rate. Everything else tested came in under 1.2 times. Thresholds are set so that no human document in the corpus fails and every LLM document does. [references/measurements.md](references/measurements.md) records the method, the thresholds, the candidates that were tested and dropped, and a dependency-parser study that falsified four more. Read [What to leave alone](#what