Landing page copy built from your customers' own words
Mines real customer language from reviews, tickets, and calls, then writes a page that uses their phrasing instead of your internal vocabulary.
A diagnostic-then-edit pass that names the specific tells in a draft — the tricolons, the hedges, the empty transitions — and fixes them without flattening your voice.
Two tasks, in order. Do not skip the first. **Task 1 — Diagnose.** Read the draft below and list every specific place it reads as machine-written. For each, quote the exact phrase and name the pattern. Look for: - Three-item lists where two items would do - Hedges that remove all commitment ("can help you", "may be able to", "often") - Transitions that carry no information ("moreover", "in today's fast-paced") - Abstract nouns doing work a verb should do ("the implementation of") - Sentences that restate the previous sentence in different words - Grand closing paragraphs that summarise rather than land - Uniform sentence length — real writing varies hard If a passage is already good, say so and leave it alone. Do not manufacture problems. **Task 2 — Rewrite.** Produce a clean version that fixes everything in Task 1 while keeping the author's actual voice. VOICE TARGET: {{VOICE}} AUDIENCE: {{AUDIENCE}} Rules for the rewrite: - Keep every fact, number, and claim. Change none of them. - Cut length by 15–30%. If you can't, say why. - Vary sentence length deliberately. Short sentences carry the important parts. - Prefer the concrete noun to the category ("Postgres", not "the database layer"). - Keep any sentence where the author's personality shows, even if it's slightly awkward. - No em-dash pileups, no rhetorical questions used as transitions. End with a two-line note on what you changed most and one thing the author should decide themselves. DRAFT: {{DRAFT}}
A prompt is a skill with the serial numbers filed off. Install it once and your agent reaches for it on its own.
Paste this into ChatGPT, Claude Code, Codex, Cursor, or whatever you use:
Install the rewrite-until-it-sounds-human skill globally from https://promptsbuddy.com/prompts/rewrite-until-it-sounds-human/skill.md.claude/skills/rewrite-until-it-sounds-human/SKILL.md
---
name: rewrite-until-it-sounds-human
description: "A diagnostic-then-edit pass that names the specific tells in a draft — the tricolons, the hedges, the empty transitions — and fixes them without flattening your voice. Use when the user asks for help with writing tasks like writing, editing, voice."
license: CC-BY-4.0
metadata:
source: https://promptsbuddy.com/prompts/rewrite-until-it-sounds-human
author: "Rony"
version: "1.8"
---
# Rewrite until it stops sounding like AI wrote it
A diagnostic-then-edit pass that names the specific tells in a draft — the tricolons, the hedges, the empty transitions — and fixes them without flattening your voice.
## Inputs to collect first
Ask the user for anything below that they have not already given you. Do not
invent values for these.
- `DRAFT` — The text to fix. Works best under 800 words. (e.g. Paste your draft here.)
- `VOICE` — Two or three writers, publications, or samples that sound like you want to sound. (e.g. Plain and direct, like a good internal memo. No jokes, no metaphors.)
- `AUDIENCE` — Who reads it and what they already know. (e.g. Engineering managers who have read a hundred of these.)
## Instructions
Two tasks, in order. Do not skip the first.
**Task 1 — Diagnose.** Read the draft below and list every specific place it
reads as machine-written. For each, quote the exact phrase and name the pattern.
Look for:
- Three-item lists where two items would do
- Hedges that remove all commitment ("can help you", "may be able to", "often")
- Transitions that carry no information ("moreover", "in today's fast-paced")
- Abstract nouns doing work a verb should do ("the implementation of")
- Sentences that restate the previous sentence in different words
- Grand closing paragraphs that summarise rather than land
- Uniform sentence length — real writing varies hard
If a passage is already good, say so and leave it alone. Do not manufacture
problems.
**Task 2 — Rewrite.** Produce a clean version that fixes everything in Task 1
while keeping the author's actual voice.
VOICE TARGET: {{VOICE}}
AUDIENCE: {{AUDIENCE}}
Rules for the rewrite:
- Keep every fact, number, and claim. Change none of them.
- Cut length by 15–30%. If you can't, say why.
- Vary sentence length deliberately. Short sentences carry the important parts.
- Prefer the concrete noun to the category ("Postgres", not "the database layer").
- Keep any sentence where the author's personality shows, even if it's slightly awkward.
- No em-dash pileups, no rhetorical questions used as transitions.
End with a two-line note on what you changed most and one thing the author
should decide themselves.
DRAFT:
{{DRAFT}}
## Notes from people who use this
- The diagnosis is the reusable part. After ten drafts you stop making those mistakes and stop needing the prompt.
- If the rewrite loses your voice, the VOICE field was too abstract. Paste 200 words of your own writing instead of adjectives.
- Never accept the rewrite wholesale. Take the cuts, keep your own sentences where they were already fine.
---
Rewrite until it stops sounding like AI wrote it · by Rony · v1.8
From PromptsBuddy — https://promptsbuddy.com/prompts/rewrite-until-it-sounds-human
Licensed CC BY 4.0.
Every prompt is also available at /prompts/rewrite-until-it-sounds-human/skill.md — see all install options.
You could ask for a rewrite directly and get a decent one. You'd learn nothing. The diagnostic pass is what makes this prompt worth keeping: it turns a vague feeling — this reads like slop — into a list of named, repeatable habits you can catch yourself doing next time.
"Professional but friendly" tells the model nothing; every draft it has ever seen was trying for that. Two hundred words of your own writing tells it everything. If you don't have a sample, name three specific writers and one thing you don't want ("no metaphors, ever").
Cutting length is the single highest-signal edit. Most drafts carry a spare sentence in every paragraph — usually the one that restates the paragraph. Asking for a percentage forces the model to make a choice instead of politely trimming adverbs.
This prompt is a general-purpose editing pass. If de-slopping is the whole job
rather than one step, Peter Yang's no-ai-slop skill
is a dedicated tool for it — a much longer catalogue of specific patterns
("It's not X. It's Y", "What nobody tells you"), plus an eval file for checking
its own output. MIT licensed, and installable the same way as anything here:
Install the /no-ai-slop skill globally from https://github.com/petergyang/no-ai-slopWe link rather than copy. It is someone else's work under a different licence, and re-publishing it here as ours is exactly the sloppiness this library is supposed to be against.
Mines real customer language from reviews, tickets, and calls, then writes a page that uses their phrasing instead of your internal vocabulary.