Avoid AI Writing finds the habits that make text read as machine-generated: stock words, template phrases, chatbot leftovers, vague attributions, stacked hedges and uniform rhythm. It works in three modes: rewrite, detect-only and edit-in-place for a named file.
Findings are ranked in three severity tiers, and an editing contract limits what may change: quotes, code, tables, links and frontmatter stay as written, and no facts or experiences are added. Voice profiles, context profiles and an optional house-style config adjust the result.
What You Get
- Three modes: rewrite, detect-only and edit-in-place for a named file
- A pattern catalog with severity tiers from credibility killers to stylistic polish
- Voice options (casual, professional, technical, warm, blunt) and context profiles for different kinds of writing
- An editing contract that protects quotes, code, tables, URLs and frontmatter and bars invented facts or experiences
- Helper scripts that normalize quotes and apostrophes and check a file against a style config, plus a preservation validator
- A report with changes made, the checks that ran, residual findings and the reason editing stopped
Common Use Cases
- Cleaning a blog post or LinkedIn draft that reads like a chatbot wrote it
- Scanning a document in detect mode to see what is flagged before changing anything
- Fixing a Markdown file in place with small targeted edits
- Applying a house style config on top of the cleanup
- Checking a rewrite for lost quotes, numbers or links
What Makes It Unique
The skill puts as much weight on restraint as on removal. Its editing contract says a pattern match becomes a finding only after context exceptions are read, forbids injecting fake candor, invented specifics or staged rhythm, and returns the text unchanged when nothing justified remains. It also ships a detector and preservation validator for deterministic checks, and states that detector scores must not decide consequential matters such as academic or hiring decisions.
Under The Hood
The skill loads SKILL.md and then a very large patterns reference that the agent is told to read in full before every audit, so expect a heavy read. Bundled Node scripts normalize quotes and apostrophes, check a file against a style config and validate before and after text; the instructions have the agent run them in the shell when Node is available, and in edit mode the quote normalizer can write changes back to the file. Edit mode changes a file on disk that the user named. No network calls or outside services are involved.