Training off the Subtext, Not the Storytelling: The Journey towards AI-Enhanced Narrative Development

Using AI to explore meaning beyond words for deeper storytelling

The journey to leverage AI models for creative storytelling has been a rewarding challenge here at Narrative First. When using o1 models, we're particularly interested in how they help us with the underlying meaning of a story—the subtext—rather than merely generating words strung together for surface-level storytelling.

Recently, I had a fascinating back-and-forth conversation with a developer on the OpenAI Developer forum that epitomizes our approach in blending the predictive power of Subtxt with OpenAI's models. This exchange highlighted the crux of training AI not just to generate text, but to craft meaningful and emotionally resonant narratives. Below, I'll share our conversation in detail, while explaining the reasoning and methodologies behind each aspect of our process.

The Developer's Dilemma: Challenges in Long-Form Fiction

The conversation began with the developer outlining some significant issues they faced while attempting to generate long-form creative fiction:

Yes but as part of a tool I’m building to do technical writing not fiction. The fiction writing was more of a test to see both how creative I could get o1 to be and how long of a document I could generate...There are a number of challenges with generating long context. Cohesion is a big one. You’re stitching together the output of multiple model calls and it’s difficult to make the outputs feel like one continuous output.

Focus is another big issue. It’s difficult to generate long outputs and have the model stay on task. Especially when the output spans multiple calls. Boredom... the longer the output the more likely the model is to get bored and immediately wrap things up with a 'and so on...'

These are common hurdles in long-form narrative generation—cohesion, focus, and, amusingly enough, the tendency of the model to "get bored." The longer the content, the harder it becomes to stitch everything together seamlessly while retaining meaning and purpose.

My Response: Leveraging Subtext for Long-Form Fiction

I responded by suggesting an approach that shifts the focus from pure storytelling to the deeper subtext of the narrative:

"For long-form fiction, you’ll need to generate text based on the underlying meaning of the novel (the subtext). If you can find a framework that models what stories mean rather than what they say, then you don’t have to worry about token length or context windows."

With Subtxt, we’ve found o1 to be especially adept at unraveling underlying meaning and capturing authorial intent. For example, when working with a story that explores themes of loss and redemption, o1 can effectively identify these core themes and guide the storytelling process to ensure that each scene resonates with these deeper meanings. This allows us to bypass the problems associated with token length and context limits. When the subtext is defined—the thematic and narrative purpose of each part—the model doesn't need constant reference to previously generated text. It’s akin to how human authors work: keeping in mind what a story means enables more coherent and intentional storytelling.

Balancing Outline and Discovery Writing Approaches

The developer asked some practical questions about our process:

I’m curious about a couple of things and feel free to share what you’re comfortable sharing. I’m assuming you’re creating some form of outline that you have the model writing to and my first question is do you pass just your outline in when you generate a chapter?

The issue I was seeing is that the model stops following the outline when the input context gets above 20k tokens...If you’re not seeing this I’d have to assume it’s because you’re taking a different approach that lets you keep your context window length down.

To which I responded:

"Don’t mind at all sharing—if you just think of it in terms of how an author would go about writing a story, it’s the same thing."

"Assuming you could break authors into two broad categories—those who write to an outline, and those who discover it (and by authors I mean both the human-type, and the LLM-type) the first group would work the way you suggest: find the unique outline, and then use the LLM to think through and logic/emote the subtext underneath (what the story means). This would be the Subtext pass."

"Then, you take that generated outline/treatment and use that as the basis for writing individual 'scenes'. Since the scenes are already tied together through the subtext pass, you can allow all the generative/brainstorming ingenuity you want without worry of the story going 'off the rails'. This would be the Storytelling pass."

"This works great for authors who already know what they want to say as they can quickly generate/think through the Subtext, and then start the actual 'writing' (Storytelling)."

For authors who prefer a more structured approach, Subtxt's predictive framework enables us to generate an outline that focuses on subtext—this framework works by analyzing the narrative elements to determine their deeper meanings and relationships. It then uses these insights to create a cohesive structure that aligns each scene with the underlying themes, ensuring the story maintains a consistent emotional and logical flow.—the meaning beneath the narrative surface. This outline forms the backbone of each "scene," allowing us to generate text where all the scenes are tied together by a cohesive underlying theme. This is what we call the Storytelling pass.

On the other hand, for those who prefer a "discovery" approach to writing—writing to understand where the story should go—the overall framework remains the same. Here, we derive subtext as the storytelling unfolds. With this group of authors, you would first do a chunk of Storytelling, appreciate the meaning of what is being written, and find the bits of Subtext underneath. You could then suggest alternate "paths" for their Storytelling based on what you know about the Subtext, or simply continue to monitor as they write (whether they are human authors or LLMs).

Generating Storytelling to Fit the Subtext

As the subtext of a narrative is contingent on the relationship of all the parts, you might not fully get to the meaning until the end (or maybe three-quarters of the way through), and at that time, you might have to re-contextualize or re-write some of the Storytelling to fit the Subtext.

"Again, this is no different than the way storytellers have worked for centuries. Many don’t realize the subtext of their work, many refuse to acknowledge or don’t want to know, yet it’s always there. In essence, you would be modeling the subconscious desire to express or find meaning within the text of a story."

"The trouble lies in trying to train off the text of a work without acknowledging that there is something deeper going on. Then you're just RL’ing the style and storytelling "talent" of the individual author."

"For instance, you could train off of the Storytelling of Shakespeare’s plays, like say for instance Romeo and Juliet, and you could prompt "write me the same story that takes place in the 1950s in the Upper West Side of New York City," but it’s highly unlikely that you’ll end up with West Side Story, without totally losing sight of what it’s all about."

"Now, if you instead trained off of the Subtext of Romeo and Juliet and then set about writing individual scenes and sequences that related back to that, you’re getting much closer to a narrative that will emotionally connect with the audience. This approach helps create the same sort of feeling you get after experiencing a great story, where every scene is anchored in deeper meaning and purpose."

This two-tiered approach—Subtext pass followed by Storytelling pass—enables an organic yet cohesive creation process. The subtext becomes the binding agent that keeps the story on track, even if the storytelling itself meanders in its details.

Understanding Through Subtext: A Storyteller's Legacy

The developer, intrigued by this approach, shared a wonderful anecdote:

As I am sure that you are aware, it is said that Mahabharat was conceived by Ved Vyas; but scribed by Ganeshji. Legend has it that Ved Vyas asked Ganeshji to write the Mahabharat... Ved Vyas agreed to this with his own condition that Ganeshji had to completely understand each verse before he penned it down.

I wonder what does it mean 'to understand'.

This idea resonated deeply with our approach. Understanding, in our context, goes beyond mimicking text—it’s about appreciating the subtext beneath it. This means recognizing the themes, motivations, and emotional undercurrents that give depth to the narrative, rather than simply focusing on the literal words. Ganeshji's insistence on understanding every verse mirrors our intention to have the AI "understand" the story’s underlying meaning. By embedding appreciation—the thematic significance—into each step of the process, the result isn't just an eloquent string of words, but a narrative that resonates both logically and emotionally.

A Practical Framework for Narrative Cohesion

The developer then asked:

To make things more visually concrete for me would you mind sharing an example of what you pass in for context to a prompt...

I replied with a simple example from A Christmas Carol, demonstrating how the narrative subtext—in this case, the temporal roles of the Ghosts of Past, Present, and Future—guides the storytelling without requiring explicit reference to the text from other scenes:

"On all three [generation] runs, you don’t really need to reference the other blocks as the concept of Past, Present, and Future are already meaningfully tied to one another. The audience/reader will pick up on that underlying subtext regardless of what the actual storytelling is (the results of the o1 pass driving the generation of the 4o pass)."

"By grounding the generation in those thematic topics you end up communicating something more to the audience than the actual words."

The subtext becomes the thread that keeps the narrative cohesive, ensuring that what’s generated by the model feels continuous and intentional, even if individual prompts are handled in isolation.

Conclusion: The Best of AI and Storytelling

Our process of leveraging o1 models with Subtxt isn't about generating strings of words—it's about crafting meaningful, emotionally resonant stories by grounding the generation in subtext. This approach not only resolves the challenges of cohesion and focus but also makes the AI an active participant in understanding the narrative. Much like Ganeshji in the Mahabharat, the AI "understands" the meaning before putting it into words, making the storytelling richer and more impactful.

This collaboration between technology and creative storytelling represents the very best of AI-app development, pushing the boundaries of what generative models can achieve when paired with the right framework. By focusing on subtext, we’re not just telling stories—we’re building narratives that last.