Writing in the Age of GenAI
The Language Learning Models of today are capable of doing a lot. How well? That’s what we’re trying to find out here. This blogpost is for the creative writers who want to preserve their voice as they experiment with AI tools and their uses. The purpose is not to find ways for AI to be your process, but to find out if it’s worth incorporating it into your process.
For those who wish to learn more about my approach and attitude towards AI and creative writing, take a look through my introductory post.
Incorporating Claude into the Creative Writing Process: Line by Line
Like ChatGPT, Claude is a Language Learning Model capable of producing novel responses from a well constructed prompt. For more insight into how to structure a prompt, my previous blogpost outlines a step-by-step process for iterative prompting.
I advocate for an approach where you write freely and use Claude at any point of major friction, building on the precedence of successful writers like Vauhini Vara. A line, whether it is a complete sentence or not, is defined here as a complete thought. By going line by line, writers must still put their own thoughts forward, building up on them block by block to ultimately construct their own pieces. Claude is then simply a tool to assist the writer along the way.
Claude is the current openly available version at the time of this post, and it is the version that will be used for the exercises below.
Exercise 1: Expanding On a Base Line
A common use of Generative AI is unsurprisingly to have them generate new content off a prompt. Here, we looking to see how effective ChatGPT is at taking an idea and running with it. In your process, this may be at a point where you know that more can be added but your brain just isn’t . Instead of coming back to it later, perhaps consider seeing what can be done with a chatbot buddy.
Here is a sentence begging to be expanded upon:
“A magnetic mold covers the room, pulsing.”
General guidelines for prompts:
- put the Chatbot into a role (“Act as a science fiction writer”)
- provide a structure for the output (rephrase the following sentence… five different ways )
- be even more specific (“…who is focused on evocative language”, “expanding on it to emphasize the horrific and unknown”)
- provide a line for the Chatbot to work off of (“A magnetic mold covers the room, pulsing.”)
As demonstrated, you can also include multiple parts to the prompt.
I prompted for “evocative language” to test how Claude deals with a looser objective. I will also prompt for “adding another clause and an additional sentence of description” to test how Claude deals with a more concrete objective.
See the results for yourselves.
Evocative Rephrasings:
Claude is lavish in its descriptions, interpreting “evocative” as dedicating more words. It is a narrator-heavy form of exposition, where the embellishing voice of the narration is what comes through the strongest.
Expansive Rephrasings:
Claude does include another clause in all of its outputs but misinterprets the prompt and does not include the additional sentence of description. Like above, Claude is good about leaning into the detail i its horror, with lavish descriptions on the mold.
Exercise 2: Modifying Existing Lines
There is also the scenario where you want an editor. You already have an idea and you want to make sure that it’s conveyed effectively. It does not have to be a sentence you see as needing work either. I wrote this sentence during a writing activity as year ago, and believe that it stands strong by itself already. However, maybe I want to see if there is another way to phrase it that comes off stronger. The components for success are there, but for better or worse we often are a bit fixed in our styles and preferences. Language Learning Models may prove useful in breaking those subconscious tendencies when the occasion calls for it.
Here is a sentence that could potentially be stronger rephrased:
“There’s restrictions, guidelines painted on the pavement and suspended on wires twenty-two feet in the air that tell us not to get ahead of ourselves, to consider others in our momentum.”
The Guidelines are the same as above. Here I focused on preserving the meaning of the sentence and then on preserving the structure.
Stylized Rephrasings:
Claude’s interpretation of style feels forced, as if it’s constantly reaching but isn’t quite there. It feels the most like it was simply pulling out a thesaurus and slotting in new words, shifting the original sentiment of the sentence in such a way that it now feels a bit juvenile.
Preserving Syntax and Diction:
Claude took the prompt too seriously here, and the outputs are all the same and not very useful on a first pass.
I went ahead and asked for more variety.
The outputs are clumsy, either having minimal changes to the diction or reordering the clauses wholesale without modifying them so that the resulting sentence makes sense. For prose at least, this approach does not yield any useful results.
Takeaways
Claude is still in its early days when it comes to its uses for creative writing. The shortcomings are numerous, with outputs that fit to the prompt which don’t feel new in any way and outputs that don’t fit to the intentions of the prompt like in other language models, at least with my approach.
Curious about how other LLM Chatbots measure up? Consider looking through these other blogposts where I go through the same exercises with the exact same prompts: