OpenAI's "slop words" list: prompts that make AI writing sound human

Everyone can spot AI writing now. It "delves into" things, it says "it's worth noting", and every other paragraph is "This isn't about X. It's about Y." In September 2026 OpenAI itself agreed. Its prompting guidance for GPT-6 Astra includes a blocklist of these phrases, as The Decoder reported on 5 September.

You don't need Astra to use the list. It works in any chatbot: ChatGPT, Gemini, Claude or Grok. This post turns it into prompts you can copy.

The phrases OpenAI says to avoid

According to The Decoder's summary of OpenAI's guide, the blocklist includes:

• "delve into"
• "promote", "use", "leverage" used as filler verbs
• "it's worth noting" and "what's important is"
• "really" and "truly"
• "Conclusion:" and "In short:" as labels
• "The simplest mental model is:"
• The question-then-answer trick: "The result? A faster team."
• The fake reveal: "This isn't about X. It's about Y." and "X, not Y"
• Made-up hyphenated compounds such as "exact-head checks"

The guide also asks for "clear, concise paragraphs, each developing a single main idea" and lists "only if the information is truly parallel, sequential, or more easily comparable". Heavy bullet points are another giveaway of AI text.

Prompt 1: write it clean the first time

Add this block to the end of any writing request:

Writing rules:
- Plain, clear paragraphs. One main idea per paragraph.
- Use a list only if the items are parallel steps or options.
- Never use these phrases: "delve into", "leverage", "it's worth noting", "what's important is", "truly", "really", "In short", "Conclusion:".
- Never use the pattern "This isn't about X, it's about Y" or "X, not Y".
- Never ask a question and answer it yourself ("The result? ...").
- No made-up hyphenated words.
- Write like [WHO, e.g. a friendly teacher explaining to a colleague].

The last line matters as much as the bans. Without a voice to aim for, the AI swaps one set of stock phrases for another.

Prompt 2: fix text you already have

For an email, essay or post that already sounds robotic, use this, based on our humanize AI text prompt:

Rewrite the text below so it sounds like a real person wrote it. Keep every fact and the same length, give or take 10%.

Remove: "delve", "leverage", "it's worth noting", "truly", "really", fake reveals like "it's not X, it's Y", self-answered questions, and bullet lists that should be sentences.
Keep: my own words where they already sound natural.

After the rewrite, list the phrases you removed.

Text:
[PASTE TEXT]

Asking for the list of removed phrases lets you check that the AI didn't change your meaning. If you want your own voice kept even more closely, use the proofread but keep my voice prompt.

Prompt 3: make it permanent

Most chatbots let you save standing instructions: Custom Instructions in ChatGPT, Saved Info in Gemini, Projects in Claude. Paste a short version once and every answer follows it:

When you write for me: plain paragraphs, one idea each. Lists only for real steps or options. Never say "delve", "leverage", "it's worth noting", "truly" or "really". Never use "it's not X, it's Y". Don't ask a question and answer it yourself.

Our custom instructions builder can write a fuller version that also covers your job and preferred tone.

Why the bans work

Language models pick the most likely next words, and some phrases became very likely because they appear in so much text online. Banning them pushes the model to say the thing directly. A short, specific list works better than "don't sound like AI". The AI can't follow a vague request, but it can check a list.

Don't make the ban list too long. OpenAI's guide also warns that newer models take instructions very seriously. A list of 50 banned words can make the writing stiff. Keep it to the ten or so that annoy you most.

Where to go from here

For more on prompting OpenAI's new models, read our GPT-6 Sol and Luna guide. If you use Claude, the Claude Opus 5.5 prompting guide covers the same ideas. The writing prompts above are free in the chat and writing collection, and the prompt builder can add these writing rules to any prompt you make.

Quick checklist

• Paste the ban list at the end of your request.
• Give the AI a voice to aim for.
• One idea per paragraph, few lists.
• Ask for a list of what was removed.
• Save the short version in your chatbot settings.

Prompts to try

Browse all 1560 prompts →