Beyond Coastal Towns: How to Get Unique Story Ideas from Generative AI

Generative AI has opened up exciting possibilities for fiction writers. Large Language Models (LLMs) like GPT-4o excel particularly in the brainstorming phase, helping writers explore new ideas and break through creative blocks. But there’s a catch: without the right approach, the outputs of these AI models can be surprisingly repetitive. The key to transforming them from basic writing assistants into valuable creative partners lies in understanding how to elicit more diverse and imaginative responses.

Problem: A Love for Coastal Towns

To illustrate the limitations of basic AI usage, I conducted a simple experiment. Playing the role of a young adult author facing writer’s block, I asked GPT-4o-mini, a popular large language model available in ChatGPT , to generate a story synopsis about two boys falling in love — not once, but one hundred times. I used a simple prompt: You are a successful young adult author. Write the synopsis for a story in which two boys fall in love. The results were surprisingly uniform. Here are the opening sentences of the first ten synopses the model returned:

In the small coastal town of Havenbrook, where the sky bleeds into the ocean at sunset, two boys from entirely different worlds find themselves drawn together in a summer of unexpected love.
In the quiet coastal town of Seabrook, high school junior Jamie Parker feels like an outsider, lost in the shadows of his more popular classmates and struggling with the weight of his own identity.
In the quaint coastal town of Havenridge, where secrets hide beneath the waves and the ocean whispers tales of old, two boys from vastly different worlds forge an unexpected bond that challenges their understanding of love and themselves.
In the quiet town of Maplewood, where the trees burst into hues of crimson and gold each fall, lifelong friends Leo and Ethan share everything—from secrets about their dreams to late-night adventures under the stars.
In the small coastal town of Seabreeze, where summer lasts just a little too long and the nights are filled with the scent of salt and adventure, two boys find themselves on a journey that will change their lives forever.
In the small coastal town of Crescent Bay, the summer sun casts long shadows and whispers secrets to those who dare to listen.
In the quaint town of Maplewood, where the vibrant colors of fall cloak the landscape in shades of amber and crimson, two seemingly opposite boys find their worlds colliding in unexpected ways.
In the quaint town of Maplewood, where traditions run deep and secrets linger like the mist over the lake, two boys find themselves at the center of a story that transcends the boundaries of friendship.
In the picturesque coastal town of Seaforth, summer days stretch endlessly beneath the golden sun, filled with laughter, adventure, and the scent of salt in the air.
In the sleepy town of Maplewood, where the leaves shift from vibrant greens to warm golds each autumn, two boys navigate the maze of adolescence and the unexpected challenge of first love.

See how similar these synopses are? For example, nearly every story takes place in either a small coastal town or a quaint town called Maplewood. Even more striking is the repetition in character names. Out of a potential 200 unique names (100 stories with two protagonists each), the AI only used 35 different names! Some favorites were Ethan (used 27 times), Jamie (24 times), Leo (22 times) and Alex and Jake (18 times each).

For writers looking to AI for fresh ideas, this lack of diversity is problematic. Fortunately, there are several effective ways to make your AI writing assistant more creative.

Solution 1: Adjust the Temperature

One of the ways to increase the variation between different outputs of an LLM is to increase its temperature. Temperature in AI writing is like a creativity dial: it’s a parameter that controls the ramdomness of the model’s output. Think of it as adjusting the balance between predictability and creativity:

• Low temperature (between 0 and 0.3): The LLM’s responses are likely to be more straightforward, factual, and repetitive. This is useful for tasks where accuracy is essential, such as technical writing or question answering.
• Medium temperature (0.7 to 1.0): the LLM balances creativity and coherence. At this level, responses are more varied and might explore less likely options, making it suitable for conversational responses or general creative writing.
• High temperature (1.2 and above): the model will select words with greater randomness, potentially leading to more creative, unexpected, and even whimsical outputs.

This impact of temperature is confirmed in the synopses that GPT-4o-mini generates. Raising the temperature from 1.0 to 1.1 increased unique character names from 35 to 40. At temperature 1.4, this number jumped to 72 — more than double the original diversity.

Another way of measuring the diversity of the synopses is to count the percentage of unique words they contain. To do this, I compared each of the 100 synopses to all 99 other synopses. For each pair, I computed the percentage of content words (nouns, verbs, adjectives and adverbs) that occurred in one synopsis but not in the other. Again, the same relationship emerges between temperature and diversity. At temperature 1, a random synopsis contained just 63% unique content words, on average. By temperature 1.4, this percentage goes up to 78%, indicating much more diverse storytelling.

Unfortunately, temperature is no perfect solution. First of all, it’s a technical parameter that isn’t available everywhere. We can adjust it in OpenAI’s playground , for example, but not in the standard ChatGPT interface . And there’s another, more fundamental, problem. As the violin plot above shows, when we set GPT-4o-mini’s temperature to 1.4, there suddenly appears a significant number of synopses with almost 100% unique words. Indeed, with such high temperatures, LLMs have a tendency to produce inconsistent output and sometimes even gibberish. Here’s one example of a synopsis that starts promisingly, but then descends into chaos:

In the quaint, sun-dappled town of Maplewood, where summer days stretch into golden memories, Jamie Ellis and Noah Carter find themselves unknowingly penned into each other's narratives. Both seventeen, Jamie is the introspective artist with a passion for capturing life's fleeting moments on canvas, while Noah is the exuberant athlete, the star of the soccer team who thrives on adrenaline and laughter. Despite their brainy, geeky differences, Larry not only knows the dashboard lights of trios, but finds great inspiration from life's breathtaking rush in operating one's cart. However, their conversation is reserved to the few trivial encounters at school events — basketball games, art shows — nothing that ever suggests the tenuous longing simmering beneath the surface.
But everything changes one lazy afternoon when Jamie, feeling brave, wanders into the local bookstore where the magical ink within the stories spark wireless jets unlelish connecting words fashioned miniong by positional needs. Left waiting at the prohibitive last destination becomes ambition in Kenneth, ahead Stockholm nine personalized followed fluiles discussions one year ine freshly wearing plaid journeying during Nazism diabetes trails.
Inside, he finds Noah as he too s carried many epic tortingly imperfections over English recovery universe desirable also primarily floats seals into dispiciple joining long wn grew browser Roscoe fascinating nickfrees trick valet versus ******pour every just musnumpy convent deceptfull sharched promised fastmp irssvaddenfalen stall pcs engineered shape>>>unknown mattered---
As shadows interecode in specialagine bet572pro succuling)}
Caught amidst dusty shelves and flourised glasses unsettlin as apples game millingpring lessonwarning promisesinto project named candid ocean-postbons while tortorpressed, this apparent smiled create ;121Mrصدقاء forget bone cleansinglib ripplingmark understatement) textual needs starebe///k_lt project certain spells sharing once imprisoned faithful corn cold-spin statement fitsbondapy overwhelmed,self-rec…

Solution 2: Use a Bigger Model

Luckily, there’s a much easier way of increasing the variation between the generated texts: just choose a bigger model. “Smaller” language models like GPT-4o-mini have their advantages: they’re faster, save resources and therefore have less impact on the environment. However, the results of my experiments suggest they generate less diverse output than their bigger siblings. Switching from GPT-4o-mini to the larger GPT-4o, for example, increases the number of unique character names in the stories from just 35 to 57, keeping the same temperature. The percentage of unique words, too, increases quite drastically. Any synopsis from GPT-4o-mini has an average of 63% unique content words when we compare it to another random synopsis. With GPT-4o, this percentage increases to 69%.

Solution 3: Introduce Reader Personas

Finally, the most effective way of increasing AI creativity is to give the models various reader personas to write for. To test this approach, I asked GPT-4o to create one hundred reading personas for a young adult book. For each persona I prompted it to give me their age, gender, favourite book genres, interests, most important character traits and their current struggles. Here are a few examples of the personas it came up with:

Next, I asked the LLMs to generate a synopsis for each of these personas, with the following addition to my earlier prompt: The story is targeted towards [age]-year-old [gender] readers that love [favorite_genres]. They are passionate about [interests], and are struggling with [struggles]. This update led to a huge jump in variation between the synopses. With a temperature of 1, the number of unique names increased from 35 to 76 for GPT-4o-mini, and from 57 to 90 for GPT-4o. The percentage of unique words increased from 63% to 74% on average for GPT-4o-mini, and from 69% to 77% for GPT-4o.

Conclusion

Generative AI can be a powerful tool for writers, as long as you know how to use it. The findings from my experiments suggest the following strategy to incorporate Generative AI in your brainstorming process:

1. Start broad, then focus: Begin with a higher temperature setting (if available) to generate diverse ideas, then lower it when you need more coherent, detailed output.
2. Create a persona library: Build a collection of different reader personas you can use for different projects. This helps ensure your AI assistant provides varied perspectives and ideas.
3. Mix and match: Combine techniques — use a more sophisticated model with reader personas for maximum variety in your brainstorming sessions.

Remember, AI is a tool for inspiration, not replacement. Use these techniques to generate varied ideas, then apply your unique voice and creativity to develop them into something truly original.