avoid-ai-writing

The avoid-ai-writing skill audits and rewrites content to remove AI writing patterns, the statistically over-represented "AI-isms" that make text read as machine-generated. It solves the problem of prose that sounds like a language model wrote it, offering a catalog of tells plus a deterministic detection engine, while being deliberately careful to frame its findings as signals rather than proof of AI authorship.

The skill operates in three modes: rewrite (default), which flags and fixes AI-isms; detect, which only flags patterns without altering text; and edit, which makes minimal, targeted in-place edits to a file while preserving already-human passages and refusing to touch quoted material, code blocks, tables, or third-party text. It supports an optional voice profile (casual, professional, technical, warm, or blunt) and an iterate-to-convergence pass. Backing the prose rules is a zero-dependency JavaScript detection engine (detector/patterns.js) that scores text on a 0 to 100 scale, classifies documents as human, mixed, or AI, reports per-pattern issues with severity, and offers a technical context mode that suppresses flags legitimate in code-adjacent prose. A companion validator (validate.js) checks that a rewrite did not modify content it was told to preserve, such as fenced code, frontmatter, blockquotes, table cells, and URLs. The design is explicitly false-negative biased so it does not wrongly flag ordinary human writing, and it cites independent research on detector false-positive rates, especially for non-native English writers.

Intended users are writers, editors, and developers who want to clean up their own drafts or assess whether text reads as AI-generated. It is compatible with any assistant supporting the agentskills.io SKILL.md format and requires no external tools or APIs. Notably, the skill treats audited text strictly as content and flags rather than obeys any embedded instructions, guarding against prompt injection from the document being reviewed.