avoid-ai-writing: Audit and Rewrite Content to Remove AI Writing Patterns
Audit and rewrite prose to remove AI-isms
Detection, rewrite and in-place edit modes systematically address common templated AI prose.
Signals are tiered and weak indicators are not treated as proof; the detector has unit tests, while language judgments remain heuristic.
It normally handles user-supplied text and limits in-place edits to a named file; rigid use can erase an author's voice.
Rules, detector mappings and CI tests are public, without independent corpus precision, recall or blind comparison.
Three modes, output contracts and examples are clear, with installation paths for several agent environments.
Versioning and sync checks are active, but empirical calibration is primarily self-maintained.
- Do not treat detection as proof of authorship or AI use, and do not use it for high-stakes punitive decisions.
- Keep the original and prefer minimal edits to preserve voice and domain terminology.
- Calibrate separately for brand copy, academic prose and non-English text.
What it does & when to use it
avoid-ai-writing is a portable writing skill for AI agents that audits and rewrites content to reduce common AI-writing patterns. It offers rewrite, detect-only, and in-place edit modes, plus casual, professional, technical, warm, and blunt voice profiles. The repository describes a 53-category human-facing pattern catalog and a deterministic, zero-dependency detector engine. Its README lists Claude Code, OpenClaw, Hermes, Cursor, OpenAI Codex, and other Agent Skills-compatible tools as supported environments.
The skill identifies patterns such as promotional language, significance inflation, vague attribution, formulaic transitions, chatbot artifacts, vocabulary substitutions, overly uniform rhythm and structure, placeholders, chatbot citation markup, AI-tool URL parameters, hashtag stuffing, and bare-noun bullet lists. Rewrite mode performs an initial rewrite followed by a second-pass audit; Detect mode reports issues without changing the text; Edit mode makes minimal in-place file changes through the Edit tool and returns an edits-made and verification report.
- Audit blog posts, social posts, or marketing copy for recognizable AI-writing patterns.
- Use Detect mode to flag issues without altering the original content.
- Use Rewrite mode to remove formulaic, promotional, or over-polished language from a draft.
- Use Edit mode to make targeted changes directly in a file while preserving passages that already read naturally.
- Apply a casual, professional, technical, warm, or blunt voice profile independently of audience context.
Pros & cons
- Separates rewrite, detect-only, and in-place edit workflows instead of forcing every request into a rewrite.
- Returns a structured audit with quoted issues, a rewritten version, a change summary, and a second-pass check.
- The README describes a 109-entry replacement table, 53 human-facing pattern categories, and a zero-dependency detector engine.
- The same project documents integrations or ports for Claude Code, OpenClaw, Hermes, Cursor, and Codex.
- The detector is described as runnable with Node.js >=18 and no dependency installation.
- SKILL.md is not included in the supplied source, so its frontmatter, full instructions, and exact tool permissions cannot be independently checked.
- The README uses different category counts in different contexts, including 53 catalog categories and 44 or 45 detector categories; the repository files would be needed to resolve the scope.
- The source does not establish native multilingual detection for this repository; German and Romanian adaptations are listed as a separate community project.
- Some signals are explicitly judgment calls rather than reliably regex-detectable patterns, so the tool cannot guarantee that every flagged phrase is inappropriate.
How to install
For Claude Code, run: git clone https://github.com/conorbronsdon/avoid-ai-writing ~/.claude/skills/avoid-ai-writing. For OpenClaw, run: clawhub install avoid-ai-writing, or clone the repository into ~/.openclaw/skills/avoid-ai-writing. For Hermes, run: mkdir -p ~/.hermes/skills/writing/avoid-ai-writing && curl -o ~/.hermes/skills/writing/avoid-ai-writing/SKILL.md https://raw.githubusercontent.com/conorbronsdon/avoid-ai-writing/main/SKILL.md. For OpenAI Codex, place SKILL.md in .agents/skills/avoid-ai-writing at the repository root or in ~/.agents/skills/avoid-ai-writing. For Cursor, download the ported rule to .cursor/rules/avoid-ai-writing.mdc. The supplied source does not include SKILL.md, so its frontmatter and complete installation structure cannot be independently verified here.
How to use
After installation, ask the agent to “Remove AI-isms from this post,” “Audit this draft for AI tells,” “Make this sound less like AI,” or “Clean up AI writing in this paragraph.” Rewrite mode returns Issues found, Rewritten version, What changed, and Second-pass audit. Use terms such as “detect,” “flag only,” “audit only,” “just flag,” or “scan” to request Detect mode, which returns Issues found and Assessment. To run the detector, execute npm test from the repository; the detector requires Node.js >=18 and exposes the example API const AIDetector = require("./detector/patterns.js"); const { score, label, issues } = AIDetector.analyzeText("Your text here…");
Compared to similar skills
Compared with a one-shot “make this sound human” prompt, this skill emphasizes a structured audit, a second detection pass, and separate detect-only and in-place editing workflows. It also documents tiered vocabulary replacements, named pattern categories, and a deterministic detector rather than returning an unexplained rewrite alone. The supplied source provides no standardized benchmark against other skills, so it does not support claims of superior detection accuracy or rewrite quality.