Rewrite work emails plainly without changing the facts
Developer phuryn's Claude Code skill rewrites work writing plainly but keeps facts; in the author's blind test, AI judges picked it 48% of the time, versus 20% for a rival.
Done with Claude CodeClaude Sonnet
What they did
Developer phuryn built an add-on skill for Claude Code, Codex and Cursor that rewrites work writing, like sales, support or human resources text, in a plain human voice. It first lists every fact and how sure it is. Then it rewrites, runs an optional style check, and checks the facts sentence by sentence. In interview mode it asks the writer up to five questions when details only they know are missing.
How it went
In the author's blind test with four AI judges, the judges said they would send its rewrite 48% of the time, versus 20% for blader/humanizer. Across 12 texts it made 18 meaning changes versus 33, and it invented no facts in 56 texts.
Worth knowing
No rewrite passed the Pangram AI detector, scoring 2 to 4% human, but output built from the writer's own notes scored 100% human on 4 of 4 texts. The author notes the samples are small, and interview mode was tested with an AI standing in for writers.
Try it yourself with Claude Code
Rewrite this [email/support reply/post] so it sounds like a person wrote it, without changing what it says or promises. First, list every fact, commitment and caveat with its level of certainty. If it's missing something only I know, like a real example, number or my actual opinion, ask me up to five short questions before rewriting. Write it in plain everyday words and avoid stock AI phrasing. Keep every claim at the same strength: no added apologies, conditions or new facts. Finish with a sentence-by-sentence check against the original. Here is the text: [paste text or your notes]
Swap the highlighted [parts] for your own details.
Source: GitHub · phuryn · Oct 9, 2026
Written with AI assistance from the original post and checked against our editorial guidelines. Results are as reported by the source, not independently verified.
