The bone structure method

A few weeks ago I started revisiting “The Art of Dramatic Writing”, by Lajos Egri. This was a seminal book during my first year of film school. My cohort and I had to strip filmmaking down, getting rid of all the flash, glamour and spectacle to learn how to make a story interesting before leaping to the higher dimension of visual storytelling with cameras.

For our instructors, It made sense to teach dramatic storytelling for film from Egri’s work. He was an expert in playwriting, not screenwriting. Plays are more performance driven than films, hyper-emphasizing characters and their motivations as forces that progress a story.

Egri’s formulation for how to write a good play has a certain rigidity to it. Creatives don’t like to think of stories or art in rigid terms, but when it comes to learning a craft, closing the scope of what to do helps the starting artist gain dexterity with the craft and eventually, enough skill to find their niche. On the negative side, rigidity can make stories feel formulaic and predictable.

Egri’s premise-based building block for stories is a reliable technique, but it also has the tendency of turning every story into an heavy-handed moral lesson. Still, I found his opinionated and rigid approach to writing stories helpful in guiding how to approach storytelling with generative AI.

For storytelling, generative AI is a double-edged sword. The AI models have been trained on all digitized knowledge. This is great - it “knows” everything. On the other hand, it “knows” everything! So it may pull from anything, even things that may not support your intent as a storyteller. One of the things software engineers working with AI learned early on is that the AI models are far more effective aligning with your intent if it has access to proper and concise context.

This is the basis of RAG, retrieval-augmented generation. In short, an AI model will be far more effective in getting you the right response if it is using context from an authoritative knowledge base, rather than the whole context of what it was trained on. Sure, the model may innately have the relevant data in its training data, but closing the scope of the context the model should be referencing elicits better responses.

Egri’s “Bone Structure” framework is a great organizing framework I’ve been testing out as a way to give AI models a proper context to help with my storytelling projects. In the book, the bone structure is a guide playwrights use to develop characters that feel nuanced and real. It breaks properties of a character into three categories: psychology, sociology, and physiology.

The idea is, once these properties are defined, a writer can effectively orchestrate how characters move through a narrative by staying true to who they are, three-dimensionally from their physical, social, and mental attributes.

Here is Egri’s character bone structure as noted in the book:

Physiology	Sociology	Psychology
Sex	Class	Sex life (moral standards)
Age	Occupation	Ambitions
Height and weight	Education	Frustrations
Color of hair, eyes, skin	Home life	Temperament
Posture	Religion	Attitude toward life
Appearance (handsome, ugly, etc.)	Race	Complexes
Defects	Nationality	Extrovert / introvert / ambivert
Heredity	Place of residence	Abilities
	Political affiliations	Qualities (imagination, judgment, taste, etc.)
	Amusements	I.Q.

As you can tell, you get to know a character very well after sitting down, thinking about what kind of characters you want in your story, and using this bone structure to bring that character to life. I was curious if this organizing structure could also be used to give AI models context for the complete world of the story - the World Bone Structure.

Just as Egri's three categories capture the full dimensions of a character, story worlds need a similar comprehensive definition: the physical space it occupies (Place), the inhabitants within it (People), and the systems that govern how it functions (Rules).

When you think about the world of a story, it can be straightforward and simple or it could be immensely complex. Think, for instance, about the differences between the story worlds of Alice in Wonderland and The Wizard of Oz. The mechanics of the world are entirely different. AI-assisted storytelling is effective if the AI models can support narrative generation that stays true to the complexities of any given story world without too much hand holding from the storyteller.

You can imagine a world bone structure may look something like this:

Place	People	Rules
Geography: Physical terrain, climate	Types: Species, races, classes, or other groups	Physics: Natural laws, magic, technology
Locations: Key locations in the story	Demographics: Population sizes	Society: Power structures, government, hierarchy
Scale: Size of world	Capabilities: Physical, mental, supernatural abilities	Culture: Values, beliefs, traditions
Resources: What’s abundant/scarce	Relations: Conflicts/Alliances	Economics: Trade, currency
Time Period: Historical era	Conflict Drivers: What creates tension between groups	Information/Communication: How knowledge spreads

These categories define a rich context an AI model can reference to effectively understand the world of a story such that it becomes a much more effective tool as a storytelling partner. A fun way to test out the robustness of the world bone structure is by checking it against a popular story or film (after all, these models have been trained on everything) that you know well.

I gave this a try with one of my favorite films, The Dark Knight, and discovered some aspects of the world driving the film that I had missed. When I asked the AI to follow the world bone structure, for the values attribute it defined five distinct moral frameworks.

Here is the JSON output from the AI model:

[
    {
        id: "value_001",
        name: "Justice and Order",
        value_type: "Virtue”,
        importance: "Core",
        description: "Belief that wrongdoers must be held accountable and society requires order to function.",
        behavioral_expectations: "Report crimes, respect law enforcement, participate in legal system, resist corruption.",
    },
    {
        id: "value_002",
        name: "Due Process",
        value_type: "Priority",
        importance: "Core",
        description: "Belief that justice must come through legal systems, not extrajudicial action.",
        behavioral_expectations:  "Work within the system, obtain warrants, follow procedures, oppose vigilantism.",
    },
    {
        id: "value_003",
        name: "Criminal Pragmatism",
        value_type: "Priority", 
        importance: "Core",
        description: "Money and power are the only things that matter; violence is a business tool.",
        behavioral_expectations: "Maximize profit, eliminate threats, maintain territory, honor deals (when profitable).",
    },
    {
        id: "value_004",
        name: "Belief in Gotham's Future",
        value_type: "Ideal”,
        importance: "Emerging",
        description: "Despite corruption and crime, Gotham can be saved through collective action and moral leadership.",
        behavioral_expectations: "Support reformers like Harvey Dent, resist corruption, participate in civic life.",
    },
    {
        id: "value_005",
        name: "Intimidation and Fear",
        value_type: "Priority",
        importance: "Important",
        description: "Fear is an effective tool for controlling criminal behavior (Batman) or maintaining power (Mob).",
        behavioral_expectations: "Use psychological warfare, create reputation for violence or invincibility.",
    },
]

In summary, the AI identified five competing value systems: Justice and Order, Due Process, Criminal Pragmatism, Belief in Gotham's Future, and Intimidation and Fear.

The exercise showed the complexity and depth of the story world - all the things that make the film one of the best comic book movies ever made. Batman operates at the intersection of Justice and Fear, while Dent represents Justice and Due Process until the Joker forces him to choose. This kind of analysis, grounded in the world bone structure context, reveals storytelling depth that could be easy to miss.

You can do the same analysis for any story, just define the bone structure the AI should follow and the target story. Beyond analysis, this structured context helps when generating new story elements. For instance, I can now ask the AI to create a new character for Gotham and it will naturally position them within these existing value systems, ensuring they feel native to the world.

Discrete context turns AI into a storytelling super-partner that understands stories at greater depth. While I only shared a small slice of the world bone structure, what you can see from the brief example is the opportunity to explore value conflicts, character motivations, and world mechanics at a level of depth that would be time-consuming to develop or analyze manually.

Conceptually, having a discrete story context is nothing new. Writers’ rooms for TV shows and some movies often have show or series bibles that serve a somewhat similar function. But show bibles can sometimes live too firmly in text rather than subtext since the purpose is primarily about maintaining story consistency. Giving AI models the proper story context opens up doors to explore storytelling in ways we haven’t previously been able to. It’s time to unlock this power and see what amazing new stories await us.