avoid-ai-writing

avoid-ai-writing is a portable writing skill for AI agents that audits and rewrites prose to remove common AI writing patterns, often called "AI-isms." It is designed to work with agentskills.io-compatible tools, including Claude Code, OpenClaw, Hermes, Cursor, and OpenAI Codex, and is also packaged as a plugin for ChatGPT and Codex. The skill operates in three modes. Rewrite mode, the default, flags AI patterns, applies up to two editing passes, verifies the result where possible, and returns a single final version. Detect mode flags patterns without rewriting, separating clear problems from judgment calls, which is useful when patterns may be intentional or when auditing content that should not be altered. Edit mode makes minimal, targeted in-place changes to a prose file while preserving passages that already read as human; it refuses source code, configuration, and generated data because prose rewrites can corrupt structured content. A key design choice is the two-pass editing limit, configurable with --iterate 1 or --iterate 2. The first pass rewrites, and the second can catch patterns that survived or repair preservation issues, stopping early when no justified in-scope edit remains. The skill reports how many passes it used and any intentional, protected, or unresolved findings. An optional voice profile (casual, professional, technical, warm, blunt) controls how the prose should sound, separate from the audience context profile. The repository includes a substantial pattern catalog. It documents 74 pattern categories with before/after examples, covering content patterns such as significance inflation and vague attribution, language patterns such as copula avoidance and synonym cycling, structure patterns such as formatting and transition phrases, communication patterns such as chatbot artifacts and sycophantic tone, and meta patterns such as rhythm uniformity and over-polishing. Later additions cover structural detection, AI-tool fingerprints like unfilled placeholders and chatbot citation markup, conversational-register tells, share-post framing, narrated candor, and unnecessary hyphenation. A 112-entry word replacement table spans three tiers, with Tier 1 split into frequency markers and clarity edits so that wordiness fixes do not push a document toward an AI classification. A bundled deterministic detector engine provides mechanical verification. It implements 53 type categories, intentionally differing from the prose catalog count because it splits vocabulary tiers and adds stylometric and fingerprint signals such as punctuation distribution, function-word entropy, and bypass-trick detection. The detector can score a file or piped text from the command line, gate prose in GitHub Actions or pre-commit hooks, and run over MCP. The two counts are mapped in the detector documentation and are not meant to be reconciled. Installation varies by platform. Users can clone the repository into an agent's skills directory, use a self-contained dist file, install as a Claude Cowork plugin, install from ClawHub for OpenClaw, drop a ported rule into Cursor, or place the skill in Codex's .agents/skills directory. The OpenAI plugin package provides seven focused skills, including a router, detector, voice-preserving rewriter, file editor, preservation verifier, and false-positive reviewer. The project is MIT licensed.