Greetings! Love AIPRM & just joined the Pro Plan mere moments ago.
I see the Pro level comes with “+3 Custom” Tone & Styles, and I’d really like to create one that emulates (or tries to emulate) my personal writing tone of voice & style.
(This is to scale my writing output on my blog, where I’ve written 400,000 words the old-fashioned way over the last 3-4 years…)
I 've searched in AIPRM & the forum for any how-to info on creating a custom tone/writing style, but couldn’t find anything; my apologies if I overlooked it.
Can you point me in the right direction regarding this? Any guidance is sincerely appreciated!
=========
My own initial (tho’ uninformed) thought on how to create a reusable custom prompt was to give ChatGPT sample content of mine, ask for its analysis of it, & then request that it use my custom style henceforth.
I’m guessing I’d access my Custom Tone/Writing Style via my saved prompt, is that right?
And that it wouldn’t instead show up in your Tone & Writing Style dropdown menus as a “Custom 1” (or similar) down there…
With ChatGPT as it is right now, the simple answer is that you can’t.
When you ask it to write like someone, or in a certain style or tone, the AI has to rely on what was in its training data that was noted specifically as being in that style or tone, or else have a detailed knowledge of enough writing samples marked or cited as being by a person, or in a style, to use those sources as its ‘pattern’.
But the ‘P’ in GPT stands for pre-trained. It has been trained on a set of documents that was selected a couple of years ago, nothing more recent than midway through 2021, and what was in all those documents is all it has to go on. That is its entire knowledge of the world, of language, etc.
For training by the user, ChatGPT has only 4,000 tokens of prompt to base its responses on - that’s around 3,000 words maximum. 3,000 words really isn’t enough to feed it a whole load of your writing to look for patterns it can copy in.
What might work for a ‘workaround’ is if your prompt itself were only about 1,000 words, you’d have 2,000 words left to try to describe the hallmarks of your style that it should try to use. That is as close as you’ll get, by the design of ChatGPT at this time. They deliberately decided to keep its memory short, and the amount it can learn from you to a minimal amount to prevent people training it to be abusive, or misogynistic, or racist, or to express hate, etc. That’s a lesson they learned from earlier generations of AI that got abused, and died a quick death due to bad press.
Thank you for your helpful reply & for the workaround suggestion too, Ammon.
Offhand, do you know what use cases AIPRM’s “custom tone & writing style” features I previously mentioned are designed for?
Are they like hybrids of the existing qualities/tendencies offered in their dropdown menus, or…?
I’m also curious about how those subjective tone & style definitions were created in the first place, e.g., “warm” vs. “confident”, “informative” vs. “creative”, etc.
I’m brand new to this field, & find it & AIPRM fascinating. Want to increase my understanding of AIPRM in order to best integrate it into my workflow…
One thing about providing the ability for ‘custom’ in anything a person makes is that one can then never be quite sure what a user may do within the parameters of that customization. Maybe someone might try setting the custom style to “Dickensian” to see if that generates a sort of Olde Worlde charm (or find it doesn’t have much understanding of the works of Dickens, at least by that traditional name). Maybe someone would try “Grandmotherly” as a custom tone (again, if it works they keep it, if not, customize again).
It’s pretty much a certainty that people will try a lot of custom styles and tones that the AI definitely doesn’t really understand, but, if it still results in some difference to output that they like, well, to paraphrase Shakespeare, “What is in a name?”
Glad to hear it. It is always important for you to be creative and to ‘think outside the box’ because that’s something current AI can’t do. They are built from the ground up to recognize patterns and underlying structure and then to predict the next word and the next word from your prompts based on those patterns and structures. They work by predicting, and thus, are always inherently predictable.
Never be afraid to experiment, nor to give detailed instructions in your prompts, such as “avoid the use of hyperbole” or “write in clear, short, easy to comprehend sentences like a textbook” rather than simply saying “write academically”. You could have it write in the style of a fable or parable to see how that comes out. Or you could specify what aspects of a fable in particular you wanted it to cover “use similes and metaphor but keep the language and metaphors suitable for children to understand” might be cool…
This is all evolving so fast that practicing experimentation, seeing what the AI is capable of by trying stuff, is going to be an important skill for adapting to each step of the evolution. Not just for the short term of months, but for a few years to come.
Oh, and be analytical. I’ve seen a lot of people give prompts like “You are a 17th Century sailor, Captain of a ship…” and all ChatGPT will do in response is find some way to sneak in a rather off-topic reference to being a sailing ship captain in the 17th century. It can’t imagine in the way we can, and it certainly can’t think how that would change its experience, because it has no experience, nor has ever ‘experienced’ anything. It is a language model.
These are resonating points you’ve shared & all are well-taken; the sage advice is much appreciated.
Your note on experimenting, testing AI’s capabilities by trying different things, & developing proficiency in coaxing nice results from it hit home.
That’s one of the things I dig about AIPRM.
For me, AIPRM is a great environment to create & test prompts, apply them to my work…incrementally growing my wee knowledgebase** of ChatGPT behavior in the process. Really good stuff!
Thanks again Ammon.
** - I only learned what a prompt was 4 days ago, & ChatGPT just a month or so before that. (I plead “not guilty by reason of a rural existence…”)
