Try these methods to prevent ChatGPT-speak.
In today's fast-paced digital landscape, artificial intelligence is revolutionizing industries across a broad spectrum of…
Did you fall asleep halfway through that sentence?
Me too.
ChatGPT-speak sucks.
AI-written text is joke fodder these days.
But it doesn’t have to be that way, boys and girls.
Let me show you how to nudge chatbots into sounding a bit less chatbotty.
I used this approach a few weeks ago in my “AI Myths” article.
If you want the chatbot to mimic your own tone of voice, this is probably the best method for it. Simply give it a sample of your writing and ask it to continue in that vein.
Yup, it’s one-shot prompting for writing styles.
Here’s a basic ChatGPT prompt:
Write a paragraph about the dangers of smoking.
Here’s what that gives you:
Yawn.
We can do better.
Find a short sample of your writing and use it with a prompt like this:
Take a look at this passage:
[INSERT PASSAGE]
Now write about [TOPIC] in that tone of voice.
Let’s try that with Gunnery Sergeant Hartman’s speech from Full Metal Jacket.
Now we’re talking!
This builds on the above approach and works especially well with a chatbot that can access the Internet.1
Simply give it a link to a page with your desired writing style and ask it to mimic it.
With this method, can even draw inspiration from many different writing samples by providing multiple URLs.
Here’s a prompt you can use:
Browse these URLs:
[LIST OF URLS]
Now combine their styles and tone of voice to write about [TOPIC].
I’ll be sticking to the “dangers of smoking” topic for the rest of these examples.
Let’s say I happen to like the writing in these two articles by Daniel Nest (yes relation):
- “7 Baby Wipes Life Hacks That Totally Work (And Are Definitely Not Just a Way for Me to Get Rid of a Huge Stock of Baby Wipes)”
- “10 Outrageous Ways Russian Media Covered The Crash Of MH17”
I feed them to ChatGPT and get this:
To be honest, I wouldn’t say ChatGPT has quite nailed the style of either piece.
But at least the resulting text is no longer ChatGPT-speak.
So…partial success?
The first two approaches only work if you have a specific style reference in mind.
If that’s not the case, you can just tell the chatbot to write for a given audience.
The most frequently used example of this with AI models is the “Explain it like I’m 5” request.
But five-year-olds are not the only type of audience out there!
Try a prompt like this:
Pretend your readers are [TARGET AUDIENCE].
Write about [TOPIC].
For demo purposes, I’ll be addressing a group of circus clowns, because who’s here to stop me?
To a keen-eyed reader, this still smells of AI-speak.
But this paragraph gives you plenty of ideas to work with when editing the article compared to ChatGPT’s standard output.
This is similar to the role/persona approach you’ve probably seen many times before.
You know, stuff like “You’re a social media marketer with 20 years of experience.”2
But the character approach focuses less on the expertise and more on the voice.
Tell the chatbot who it is, and it’ll try to sound like that person.
The prompt here can be as simple as:
You’re [CHARACTER / PERSON].
Write about [TOPIC].
Let’s see what Homer Simpson has to say about the dangers of smoking:
Listen to Homer, kids, and…drink more alcohol?
Okay, so maybe don’t take health advice from a cartoon character known for poor decision-making.
If you don’t have any writing samples, a specific target audience, or a character in mind, try this.
Ask the chatbot to write a paragraph about your topic in several different, varied styles. After that, pick the style you like the most and ask the chatbot to continue in that style for the rest of the draft.
Your prompt could look like this:
Write a paragraph about [TOPIC] in five different styles. Make the written samples as diverse as possible in terms of tone of voice, sentence structure, and so on.
Here’s how that looks for our smoking example:
If I like the “Conversational and Direct” style, I can now instruct ChatGPT to continue:
If you’re not a fan of any of the five styles, you can either ask the chatbot to generate more of them or use the “Regenerate” option, if the chatbot has one.
Here’s ChatGPT’s:
As with a lot of my posts, these methods can be easily combined into more elaborate instructions, provided you have more details to give.
For instance, you may eventually end up with a longer prompt along the lines of:
You are a [CHARCTER / ROLE]. Refer to these URLs:
[LIST OF URLs]
Combine their styles and tone of voice to write about [TOPIC] for [TARGET AUDIENCE].
Just remember to avoid doing too much too soon.
Start with a Minimum Viable Prompt and build it up from there, depending on how happy you are with the initial output.
Have fun!
What do you think of the above methods? Are you already using some of them in your interactions with LLMs?
If you have another approach that works better, I’d love to hear about it!
Leave a comment or shoot me an email at whytryai@substack.com.
You can also message me directly: