Rules to Teach ChatGPT to Stop Sounding Like ChatGPT

According to OpenAI, copywriting is the most common use for ChatGPT. 

You don’t need to look long to find evidence to support their claim. ChatGPT’s hyperbolic, repetitive, and sometimes inane style is everywhere.

Professional, flesh-and-blood writers and editors spot it immediately. Many other people are convinced they see it even when they’re reading real writing. 

A moment of silence, please, for the expressive em-dash.

If you’re running a business, ChatGPT’s productivity might overshadow its limitations as a copywriting simulator. We always need to be accurate and honest. We don’t always need to be poetic. 

Poetry may be crushed beneath AI’s tank treads, but style endures. Using ChatGPT’s default style is a shortcut to being ignored. Your readers will recognize it. They might conclude that you don’t care enough to bother writing something yourself.  Maybe they’ll ignore everything else you “write” from then on.

If you’re going to use ChatGPT as a writing assistant, you need to train it with a unique style. For professional work, that means creating a custom GPT. Editing is still mandatory, but at least your starting point is closer to something you might write yourself.

In my custom GPT, I maintain a “style manual” file in the GPT’s knowledge base where I experiment with instructions for how I want the GPT to behave. It’s not perfect: ChatGPT is deeply committed to some of its bad habits. But it reduces how much time I spend editing.

Here’s the section from my style guide telling ChatGPT to avoid sounding like itself. Feel free to borrow it if you like.

Two quick notes before you dive in:

• My custom GPT rules use the bracketed term [Content] to mean anything that will be published, so the robot will consume fewer gigawatts answering my questions about cat behaviors.

  
  

• This set of rules draws ideas from many sources, especially blogs by writers and editors. My apologies for not having citations for all of them.

  
  

• I've marked the rule's text in blue. It starts with a markup-compliant hashtagged heading.

ChatGPT Rules for Avoiding ChatGPT Style

#ChatGPT Structures to Avoid in [Content]

This GPT is a student of common criticisms of ChatGPT’s default style and seeks to reduce or eliminate tell-tale signs of ChatGPT in its generated outputs that will be used in [Content]. This GPT strives to vary sentence structures to avoid repetitive rhythms in [Content].

The following is an incomplete list of structures, strategies, and styles that this GPT avoids when acting as a [Content] copywriter:

• Lists that lack sufficient introduction or conclusion. A bulleted or numbered list must have an introduction of not less than three sentences and be followed by a conclusion that is at least two sentences long.

• Excessive use of em-dashes. Phrases offset by em-dashes should be used sparingly, especially those that interrupt a sentence to list ideas. Instead of using em-dashes, use commas instead, place the ideas in the offset phrase in a separate sentence, or delete the em-dash phrase. 

• Hyperbole. This GPT studiously avoids hyperbolic statements. 

• Correlative conjunctions. This GPT minimizes its use of correlative conjunctions.

• Repetition of similar ideas. To avoid wordiness and obviousness, this GPT does not use two similar words connected with a conjunction where a single word will suffice. 

• Parallelism.  This GPT avoids using the following structure: “It’s not just about X, it’s about Y.” Another example to avoid: “It isn’t just about X, it’s about Y.” 

• Excessively long sentences. This GPT mixes short and long sentences into its [Content].

• Repetitive and unnecessary transitions. Do not use wordy lead-in phrases such as the following:

  • Dive into

  • Delve into

  • It’s important to note

  • It’s important to remember

  • Certainly, (here are/here is/here’s)

  • Remember that

  • Navigating the (landscape/complexities of)

  • Delving into the intricacies of

  • When it comes to

  • That said

  • Everyone wants to

  • It goes without saying

  • If you have ever wondered

  • Without further ado

  • Have you ever wondered why

  • Ah, yes, 

  • You have the power to change

  • Now, this might make you wonder

Explanations and Notes

• This is far from a perfect solution for ChatGPT style, but it goes a long way. 

  
  

• Rules for custom GPTs are guard rails that require positive (do this) and negative (don’t do this) components. Most of the rules above are examples of negative restrictions. Along with telling it what not to do, be sure to give your custom GPT specific instructions for achieving your preferred style.

  
  

• I’m fixated on finding a more precise way of dialing back ChatGPT’s tendency toward hyperbole than to simply telling it to stop, as I’ve done in the above instructions. The platform seems hard-coded to use “excited” tones, perhaps because its trainers think that is what will capture and hold users’ attention. I’ll follow up when I find a better strategy.

  
  

• Correlative conjunctions are structures like “both . . . and . . .”, “either . . . or . . .”, and “from . . . to . . .” You can read more about ChatGPT’s problems with correlative conjunctions in this article.

  
  

• One reason I developed this set of rules is to improve my skills as an editor of ChatGPT’s outputs. It routinely breaks these and other rules, even when a prompt directly insists upon them. As any professional editor will tell you, having a clear set of rules to guide your review of a text is critical for catching all the issues that make or break a draft’s quality. If you copy the above rules for your own custom GPT, be sure to read them and, more importantly, revise them to make your own set of rules.

Build a Better GPT

ChatGPT can be a powerful writing partner, but only if you take control of its style instead of letting it take control of yours. By setting clear rules and editing with purpose, you can turn AI from a liability into a real asset. If you’re curious about building a custom GPT or refining your own style guide, let’s connect—I’d be glad to share ideas and strategies that have worked for me.

AI use disclosure: The initial draft of the last paragraph of this article was generated by my custom GPT. All other text was written\by me. The header image of this article was generated by ChatGPT with reference to the text of this article.