Training an LLM to Understand Your Writing Style
Adapted from https://www.forbes.com/sites/jodiecook/2023/06/15/train-chatGPT-to-write-like-you-in-5-easy-steps/?sh=4fcd96d1530f. Prompts have been changed based on my own trials, successes and mostly failures using Jodie Cook’s suggestions for training ChatGPT.
Pre-Prompt Work
1. Draft your prompts in advance.
Have all your prompts drafted in a single document so you can cut and paste them into your LLM or GPT as you are working. This allows you to review and test the detail and clarity of your prompts prior to inputting them into an LLM. Better to test and adapt in this phase than in real time while trying to train our LLM.
Be sure to have all your prompts drafted and gathered through to the project outcome. In this case, we are training an LLM to understand our writing style so our primary outcome is to confirm the LLM understands our writing style, and we will prove this by asking the LLM to summarize its learning so we will need prompts to deliver on each of those outcomes.
Additional prompts for creating social content, blog articles, or even proposals in the tone and voice of our writing style might also be needed.
1. Collect your writing samples to have on hand while training your LLM.
Pull all your writing examples together, ideally in a single space or document, or in easily accessible documents. I learned when sharing more than 4 examples, the LLM started to respond to my examples straying from the guided prompts as part of this training exercise. I would recommend limiting your examples to 4 or fewer, then creating a series of new prompts to train your LLM to continue to update your writing style with additional writing examples.
I also learned that when sharing writing examples that included or ended with questions—mine included some case studies with class discussion questions written for a university course I taught—the LLM would respond to the questions rather than the prompts I was asking for as part of this training exercise.
Train Your LLM
3. Start training your LLM and practice some patience.
Prompt 1: I’d like you to help me create [INSERT FUTURE TASKS] for [INSERT PURPOSE]. Your first task will be to understand my writing style based on the examples I give you. After that, we’ll create some [INSERT CREATION TASK]. To start, please respond with “I’M READY TO START” and I will paste some examples of my writing. After I paste each example, don’t respond with anything other than “I’M READY FOR THE NEXT EXAMPLE” and I will paste another example. When I’m done, I will say “I’M FINISHED.” When I say, “I’M FINISHED,” please confirm that you have reviewed my writing examples and that you understand my writing style.
Once your LLM has confirmed it’s ready for your writing examples, copy and paste your examples and wait for the appropriate reply from the LLM before proceeding.
Once you’ve pasted all your examples, enter “I’M FINISHED,” and wait for the LLM to confirm that it has reviewed your examples and understands your writing style. In the next prompt, we’ll ask the LLM to associate your writing style with a name.
4. Name your writing style.
Prompt 4: Thanks. Let’s call this “[INSERT A NAME] Writing Style.” Can you summarize [INSERT A NAME] Writing Style in 4 bullet points focusing on sentence structure, tone, and voice?
Interestingly, being polite to your LLM may yield better results over time. Kindness and patience pay off even with machines.
Review the outputs from your LLM related to your style, voice, and tone. If you find something isn’t right, copy and paste from Prompt 1 to let your LLM know you will be sharing more writing examples to help enhance its training. I learned in my own training, the LLM was biasing its summary based on the name I assigned to the writing style regardless of the examples shared. Keep this is mind as you are assigning a name to your own writing style. Use something generic, simple, and unique (if you are using a free version of an LLM), like your three-letter initials.
PRO-TIPS
When selecting writing examples choose a variety of documents—emails, memos, social content, articles, etc. If you are training for business purposes, avoid using anything that isn’t already publicly available e.g., client-specific proposals, proprietary materials or information, pay-gated ebooks or materials, business plans, etc. Remember, anything input into an LLM has the potential to be an output for someone else using the LLM.
Remember that free, publicly available LLMs don’t save or store information or prompts from session to session so once you have trained an LLM to understand and complete a task, it will not recall that task when prompted again the following day, or even hours later. Set your expectations accordingly.
If you are sensitive to sharing personal information online (we all should be a little more sensitive), avoid using your full name, personal, professional, or business identifiers when naming your writing style in your prompts. I also learned that despite my efforts and specific prompts, my LLM was biasing its descriptions of my writing style by making assumptions gleaned from the name I assigned to the writing style—using my street name rather than my real name. Despite sharing a deluge of professionally written, technical case studies, my writing style was still evocative, dark and suspenseful with a touch of mystery and intrigue.
ChatGPT isn’t the be-all/end-all. I tried at least a half dozen times over the course of a week to train LLM to save my writing style and to draft a single blog post for my site and it failed within the first prompt every. Single. Time. It was akin to talking to my teenager and training a puppy at the same time. What did work? Google Bard. I was able to train Bard, produce a blog post, social post, and a TikTok script for a 60-second video based on an online article. If one LLM isn’t producing results. Move on.