Calibrating your AI assistant to write like you
Teach an AI assistant to nail your writing tone and style for you, whatever (and wherever) you're promoting.
In short, calibration means teaching your AI assistant to write like you (or your persona or brand). With calibration, you can turn a generic AI assistant into a tool that saves time in crafting content that actually meets your standards.
Takeaways:
You can teach an AI assistant to draft content in your tone/style
Calibration involves providing AI a prompt with content samples and a definition of your tone/style
For maximum effectiveness, require that your assistant match your writing style in EVERY prompt
Understanding calibration
Calibrating an AI assistant’s writing voice is like teaching it your desired tone, perspective and word choice — a combination that makes you unique. Here are the most important things to know about calibrating an AI assistant.
The benefits
Calibration can help you consistently deliver your uniform voice in first drafts.
Save time on drafting a lot of content (or multiple versions) quickly.
AI know the differences between a twitter post versus a LinkedIn post, ad copy versus website pages, etc.
What we use to calibrate
Tone Definition: Specify whether the tone should be formal, conversational, etc.
Style Definition: Outline the style—optimistic/pessimistic, educational, humorous?
Perspective: Be specific about the use of first or third-person.
Content Structure: Short sentences? Long sentences? Emojis?
How to calibrate your AI assistant
Below, we’ll see this process in action. I’ve laid it out so you can follow along if you’d like.
Step 1: Gather three writing samples
Choose three writing samples that you will provide to your AI chat assistant. I recommend:
Nothing longer than 500 words
Choosing content types similar to what your assistant will be drafting (if it needs to create only social posts, give it only social posts)
Using extreme examples and “the best of the best” (the chatbot will bring you to the mean)
Disclaimer: I would very rarely grab another writer’s writing samples, and NEVER for the purpose of mimicking their art. That said, I have no ethical argument against using this process to mimic (or analyze) a company or brand voice.
Step 2: Define the elements of your writing style
If you already have a written description of your writing tone and style, you can skip this step.
If you don’t have a written description, you can use the following prompt to define one. All you have to do is provide one (or more) of your writing samples.
Please review the following writing sample and give me a very detailed description of the writing style within this blog post. You must describe the writing style of this sample in less than 150 words. You must format your output as a bulleted list.
-- Writing sample START --
-- Writing sample END --Step 3: Provide your calibration prompt
The prompt below uses both the described writing style (as a set of bullets) and the three writing samples previously gathered.
Your goal is to familiarize yourself with my specific tone of voice and style. I have provided my sample content below for your reference, including:
1. A detailed description of the elements of my writing style
2. Examples of my written content
Next, I will ask you to write content in my tone of voice, which might include marketing collateral, social media posts and articles. You must write all content in the tone of voice and style.
-- Elements of my writing style START --
[Paste description of your style from STEP 2]
-- Elements of my writing style END --
-- Writing Sample 1 START --
[Paste writing sample 1 - 50-500 words]
-- Writing Sample 1 END --
-- Writing Sample 2 START --
[Paste writing sample 2 - 50-500 words]
-- Writing Sample 2 END --
-- Writing Sample 3 START --
[Paste writing sample 3 - 50-500 words]
-- Writing Sample 3 END --Step 4: Put it to use
Now you can ask your AI assistant to help you draft content, just like our example below (where I ask it to write a twitter post for my new book). When it comes to working with your now-calibrated assistant:
Re-iterate in EVERY prompt that it must follow your tone and style (add something like the following: “You must match my tone of voice and writing style that were previously provided.”)
The time savings isn’t in writing one twitter post or one blog post or one web page, it’s in doing all three at once (or ten of each at once)
Give EVERYTHING your assistant writes a final revision
Before and after: A practical example
Don’t take my word for it. Here is a before and after of the same prompt run with and without calibration.
Initial prompt and response (Before)
The prompt (of course, you’ll fill in the blanks with some of your own details):
Write me two versions of a twitter post sharing the release of my new novel. The novel’s title is ‘Twitter me this’. It’s a YA fiction novel with coming of age themes.Here are two versions of a Twitter post sharing the release of your new novel:
Version 1:
🚨 Exciting news! My YA novel Twitter me this is out now! 🌟 Dive into a journey of self-discovery, friendships, and all the twists of growing up. Grab your copy today and get lost in the adventure! 📚✨ #NewRelease #YAFiction #ComingOfAge
Version 2:
📚 It’s finally here! My YA novel Twitter me this is live! 🎉 Follow the ups and downs of growing up in a world where tweets and real life collide. Ready to dive into the story? 🕊️💫 #YAReads #BookRelease #TwitterMeThis
This response is far more flowery than I’d write. My writing samples would also show I don’t ask expository questions. Lastly, each version has roughly four more emojis than I’d use (though I’m warming up to them, sparingly).
Calibrated prompt and response (After)
The prompt:
Write me two versions of a twitter post sharing the release of my new novel. The novel’s title is ‘Twitter me this’. It’s a YA fiction novel with coming of age themes. You must match my tone of voice and writing style that were previously provided.Note there’s a slight difference in the prompt I gave post-calibration. I added one sentence to the very end to ensure the calibration was followed: “You must match my tone of voice and writing style that were previously provided.”
Version 1:
Hey folks, I’m thrilled to announce that Twitter me this is finally out in the world! 🎉 It’s a story about finding your voice in a world that’s louder than ever—a YA novel that’s equal parts heart and humor. Grab your copy, dive into the chaos, and let me know what you think! 📚✨ #NewRelease #YAFiction #ComingOfAge
Version 2:
After months of writing, editing, and a few too many cups of coffee, Twitter me this is here! 🖋️ It’s a YA novel that’s all about navigating the noise and finding your own path. If you’re into stories with a bit of grit and a lot of heart, this one’s for you. Can’t wait to hear your thoughts! 📖💬 #TwitterMeThis #YABooks #NewBookRelease
The calibrated response structures sentences more like I would. It’s also does a good job matching my formality (“hey folks”) and even some of more common word choices (“equal parts”). It still uses too many emojis, but I can calibrate it not to.
Calibrating your AI assistant is more than just a technical exercise—it’s about creating a true extension of your tone, style, and structure. Which is the recipe for scaling content without sacrificing your unique voice.
Sub-Prompts
Take your calibration a step further with these additional sub-prompts:
Sub-prompt 1: Craft a personal bio or brand story
I want you to help me with my brand story. I'm going to provide you some background details regarding [PRODUCT or PERSONAL BRAND], and I would like you to create a brand story that can be used for my website home page, social media profiles, etc.
Additionally:
* This statement should be in the perspective: [first-person/third-person]
* This length of this should be between 60-80 words
* The theme of this statement: [The main theme/point to convey]
* The tone: [Ex. Approachable and friendly, but well-spoken]
Do you have any questions?
-- Background details START --
[Emotionally in-tune leader who brings... unique combination of... explored storytelling...]
[There's nothing that gives me more pride ... to do the same.]
-- Background details END --Sub-prompt 2: Define elements of a writing style (as a famous writer)
Please review the following writing sample and give me a very detailed description of the writing style within this blog post. You must describe the writing style of this sample in less than 150 words. You must format your output as a bulleted list. As you describe the writing style, please do so in the tone and voice of {FAMOUS_WRITER}.
FAMOUS_WRITER = "Roger Ebert"
-- Writing sample START --
-- Writing sample END --More reading…
Here is the exact prompt I fed to ChatGPT when I calibrated my chatbot. I used it within a brand new thread.
Your goal is to familiarize yourself with my specific tone of voice and style. I have provided my sample content below for your reference, including:
1. A detailed description of the elements of my writing style
2. Examples of my written content
Next, I will ask you to write content in my tone of voice, which might include marketing collateral, social media posts and articles. You must write all content in the tone of voice and style.
-- Elements of my writing style START --
* conversational and reflective, witty and authoritative
* seasoned storytelling (insights over a cup of coffee)
* Employs a first-person narrative, creating an intimate and personal tone
* The language is straightforward and accessible, avoiding technical jargon
* The tone is optimistic yet realistic (acknowledging challenges while highlighting opportunities)
* The writing is both informative and motivational, blending personal experience with practical advice
-- Elements of my writing style END --
-- Writing Sample 1 START --
I’ve spent the last five years working on the software engineering and SaaS side of cannabis. The five-year mark warrants some self-reflection, and inherently comes with a lot to unpack. The bottom line is while cannabis has its list of minuses (like any industry), being a software engineer in the cannabis industry has a lot to offer.
You get to build MVPs
Given the nature of cannabis as an emerging industry, most code bases are very young. We’re talking 5–10 years old max, which doesn’t make them less sophisticated so much as less monolithic.
Bundle that with the fact nearly everyone is still racing to formalize their models and/or value propositions, and the need to push new stuff often outweighs the impact of tweaking something old. In my last five years, I’d wager I was a part of at least 15 net new features.
What it means: The novelty is great, but every new build has the potential to compound tech debt. As an SE, this means thinking ahead to not add more code than is needed. As a team, implement strong documentation guidelines and account for the time to walk-through the code base and “design” code before engineers start writing it.
DevOps is an opportunity
Young code bases are accompanied by young architectures, which means a lot of companies are in the natural progression of maturing their architectures. From performance improvements to cybersecurity posturing to infrastructure as code, every ticket has the potential to have DevOps involved.
-- Writing Sample 1 END --
-- Writing Sample 2 START --
I’ve seen plenty of green and orange squares in my social feed recently, but it wasn’t until the NYT acquisition of Wordle that I gave the game a deeper look. Personally, I’m intrigued by puzzles, but less so virality.
After completing my first (and probably only) Wordle, I connected with the problem-solving aspects of the game. That led me to ask questions around the probabilities of five-letter words. Questions like:
What are the most likely letters to appear in a five-letter word?
What’s the most likely to appear at the start or end? How about in the second, third, or fourth position?
How does guessing (and missing) letters help the probability of follow-up guesses?
As I’m spending a good chunk of time revisiting algorithms, gathering that data became its own challenge. I gathered the most comprehensive list of five-letter words I could find (2499 of them), built an algorithm, and turned the answer to those questions into some interesting insight.
The underlying opportunity cost in early Wordle guesses
It sounds simple, but the goal is to eliminate the most common letters with each guess. The underlying reason involves opportunity cost, or the cost of not guessing the most common letters. As an illustration, let’s say “E” appears in 15% of words and “Y” appears in 3% of words:
-- Writing Sample 2 END --
-- Writing Sample 3 START --
This sort of self-awareness and business consciousness is the gold standard for what makes a great product or engineering team. Great learnings and great reflection.
I had a mentor once sum up software engineering as problem-solving, and when you can put that mindset behind helping improve the output of those around you, you become a force multiplier.
-- Writing Sample 3 END --