eliasoulkadi/bunsho

Bunsho is a set of three related tools meant to be used together as a pipeline for turning AI generated writing into documents that read as human written, factually checked, and are exported as clean PDFs. The name means document or composition in Japanese, and the project positions itself as a companion to a separate design system project called Kami. The README frames three problems it is trying to solve. First, AI generated text often contains invented facts such as made up statistics, citations, or version numbers, and it cites a claimed 15 to 20 percent hallucination rate on factual questions from outside research. Second, AI writing tends to show detectable patterns, like uniform sentence length, overused connecting words, and repeated dashes, which detection tools such as GPTZero can catch. Third, converting HTML into PDF is often unreliable, since one common tool fails on Windows due to a missing dependency and another is deprecated. To address these, Bunsho provides three independent skills. The first, called research, checks factual claims before writing happens, drawing on techniques the README attributes to intelligence analysis methods, a public radio verification process, and open source research standards, including a ranked hierarchy of how trustworthy a source is. The second, called humanize, edits text to remove a catalogued list of sixteen patterns commonly associated with AI writing, following a five step process the README describes as detect, prune, vary, personalize and verify. The third, called kagen, converts HTML files into formatted PDFs using Playwright and Chromium instead of the tools mentioned earlier, and the README includes benchmark numbers claiming faster render times and native Windows support. To use it, you need Node.js version 20 or higher and a Chromium browser installed through Playwright. Each skill exists as a standalone file meant to be copied into an AI agent's configuration rather than installed as a typical package, and there is a build script for generating a PDF from an HTML file with a chosen paper format. The project is released under the MIT license.