Find the pattern, read the context, then decide what to change
Avoid AI Writing
Edit English without losing meaning
Conor Bronsdon’s English editing skill combines word and structure rules, context exceptions, and optional local checks. Audit a draft, return one revision, or make targeted changes to a prose file.
Hosts listed by the author
- Claude Code
- Codex
- Cursor
- GitHub Copilot
- Windsurf
- Cline
- OpenClaw
- Claude web & desktop
- GitHub stars
- 5,326
- Version
- 3.38.0
- Checked
- Categories
- Writing & ResearchContent & Knowledge
Choose editing permission before choosing a voice.
Start with detect to see whether its judgments fit your article. Use rewrite once you agree with the direction, or edit to save changes in a named file. Compare quantities, conditions, and quotations with the original.
Pick a task and get started
Install the plugin one command at a time in Claude Code, then reload it. Save English prose as draft.md and put the audience, destination, and factual constraints in brief.md. Start with the audit task. These prompts are instructions to the assistant, not shell arguments.
Find empty phrasing, repeated structures, and unnecessary emphasis in English prose before deciding what to edit.
Install the complete plugin in a Claude Code session
Enter one command at a time in a Claude Code session, not a shell. Skip the first line if the marketplace is already added. Reload plugins or restart the session afterward. Installation retrieves the version available then; this review pins v3.38.0.
Read the author’s setup guide/plugin marketplace add conorbronsdon/avoid-ai-writing /plugin install avoid-ai-writing@conorbronsdon-skills /reload-pluginsFirst run: one draft and a short brief
Open Claude Code in the project containing draft.md and brief.md. After installation, use the audit prompt. Choose technical-blog for a technical article, blog for general prose, or docs for documentation. Expect Issues found, Assessment, and 0 editing passes. Review excerpts and exceptions against the unchanged file before deciding to revise it.
Read the author’s modes and invocationGive your AI this first task
Replace example filenames with your paths. Confirm that the assistant read the full pattern catalog. Set the editing scope before requesting changes, then review the final diff.
Use avoid-ai-writing in detect mode. Read the complete skill and references/patterns.md, then draft.md and brief.md. The brief supplies the audience, destination, and facts that must remain intact. This is a technical blog: use context technical-blog and preserve the existing voice. Do not change files or return a rewrite. Return Issues found and Assessment with short source excerpts, applicable rules, severity, and reasoning. Distinguish findings from context exceptions and protected content; separate frequency markers from clarity suggestions in the word tables. Report 0 editing passes, checks actually run, and judgment calls. If the local detector is available, use its corresponding technical context; otherwise disclose that it was not run and invent no score. Make each finding easy to compare with the source and accept or reject.
Check what you receive
- Located findings with reasons and severity, plus context decisions and unresolved questions.
- An unchanged draft. Check terminology, necessary qualifications, and quoted passages for false positives instead of relying on a total score.
More about this skill
Avoid AI Writing is an English editing skill with optional deterministic detection, quote normalization, and preservation checks. The pinned v3.38.0 catalog contains 74 pattern categories; the detector documents 54 reportable types. These are different inventories. Some judgments require reading a passage and understanding its purpose, beyond matching a phrase.
A match is a candidate for review. Context, exceptions, and source meaning determine whether it becomes a finding and then an edit. Technical terms, meaningful contrasts, and necessary uncertainty can stay. Context sets applicability, voice sets register, and style sets house conventions; none permits invented experience, altered facts, or unauthorized changes to protected content.
The three modes have distinct deliverables: detect reports findings, rewrite returns one final text, and edit makes targeted changes in a named prose file. Ordinary cleanup preserves structure and argument. The default ceiling is two editing passes, including repairs; unchanged text is a valid result. The final record explains what changed, what was checked, and why editing stopped.
Sources and verification
Sources checked
Based on full reads of the entry documents, pattern catalog, relevant tool documentation and key auxiliary scripts, plus pinned sources for both comparisons. Installation, paths, and check boundaries were reviewed statically; no third-party skill installation, code execution, or model rewrite test was performed.