Description
Anti-Distill is an English and Chinese prompt-based Skill for Claude Code. It accepts Skill files, colleague-skill folders or general documents, asks for cleaning intensity, classifies content as SAFE, DILUTE, REMOVE or MASK, and previews proposed changes before writing cleaned documents and a separate private knowledge backup. Its instructions aim to preserve Markdown structure, terminology and approximate length. Users can adjust individual proposed changes before execution. The source includes classification and dilution prompts plus an example. Requires a compatible AI assistant with model access and file reading/writing tools. PDF and image input depend on the host assistant. This is an LLM-guided workflow, not a standalone deterministic application; it does not guarantee resistance to knowledge extraction. Keep private backups under your own access controls. Validation: source-package CRC and per-file parity checked; one synthetic general-Markdown case executed by Codex with preserved headings, bullet count and SAFE text, and backed-up removed passages. Native Claude Code, PDF/image input and colleague-skill folder paths have not been tested in this release. Original upstream: https://github.com/leilei926524-tech/anti-distill . Source mirror: https://github.com/FinchTech-AI/anti-distill . Source commit: 21e6aa9c72a03a5456eeb4cd63234a2fa2935387. MIT is declared in the upstream README; original source and attribution are preserved. Free plaintext delivery requires Finch login, with no purchase, on-chain acquisition or buyer gas payment.