conorbronsdon/avoid-ai-writing

What it does

avoid-ai-writing is a portable skill for AI coding agents, including Claude Code, Codex, OpenClaw, Hermes, and Cursor, that audits prose for patterns typical of AI-generated writing and rewrites it. The rule catalog covers content patterns (significance inflation, promotional language, vague attribution), language patterns (word replacements, copula avoidance, synonym cycling), structure patterns (em dash overuse, transition phrases, inline-header lists), and communication patterns (chatbot artifacts, sycophantic tone). It ships in three modes: rewrite, detect-only, and in-place editing of a prose file README.

Beyond the instructions, the repository bundles a zero-dependency JavaScript detector that turns findings into a 0–100 score, plus a preservation validator that checks code, frontmatter, tables, links, and headings survive a rewrite detector. The same detector powers an npm CLI, a GitHub Actions gate, and an MCP server MCP repository.

It suits writers, maintainers, and developer-relations teams who want to clean drafts or documentation before publishing, and who want a deterministic check rather than a one-shot prompt.

Quick start

Install the skill directory into your agent's skills folder. For Claude Code:

git clone https://github.com/conorbronsdon/avoid-ai-writing ~/.claude/skills/avoid-ai-writing

Keep SKILL.md next to references/patterns.md; the entry file loads the catalog before auditing. Then invoke it with /avoid-ai-writing <your text> README. Cowork, OpenClaw, Cursor, Hermes, and Codex each have their own install path, and a self-contained single file is available for agents with a rules field single-file bundle.

To run the detector on a file from the command line:

npx --package avoid-ai-writing-detector avoid-ai-writing draft.md

The CLI prints the full analyzeText() result as JSON and exits 0; usage and I/O errors go to stderr with exit code 2 npm package.

Constraints

• The rules and bundled detector are calibrated for English. Versions for German, French, Italian, Romanian, Swedish, and European Portuguese are maintained separately by community authors and are not validated by this project language adaptations.
• The single-file bundle is roughly 138 KB; some tools have much smaller instruction fields, so check the per-tool size limits before pasting it other agents.
• The skill removes AI-writing tells, not house style. Enforcing a published style guide is out of scope; the optional --style input applies rules you supply and only some are checked deterministically README.
• The detector score is a writing-quality baseline, not an authorship classifier; passing a rewrite does not verify meaning or facts.
• The npm package includes the detector runtime and CLI but not the repository's scripts/ utilities, which need a clone.
• MIT licensed.

HotRepo's take

On the 2026-11-10 JavaScript daily list the repository recorded 5,542 stars and 477 forks, with 42 stars gained that day and a best position of 9 on that list. It is a single snapshot, so it establishes that day's standing rather than a trend. What stands out technically is that the rule catalog and the detector engine deliberately use different category counts—74 prose patterns versus 54 detector type categories—and the project documents the mapping instead of forcing them to agree category mapping. Publishing its own documentation's score, including defects the scan found, is a more concrete quality signal than the star count alone PROOF.