Observations and resources for game designers seeking to tell stories in highly dynamic games
Our group discussed techniques, approaches, and common obstacles around interactive storytelling, specifically with regard to combining interesting and highly dynamic gameplay systems with effective and coherent storytelling.
Game designers have used games to tell stories since the dawn of the medium. While we continue to make advances – and dynamic storytelling in games has become quite sophisticated in some cases – we feel there is still significant unexplored potential for better integration of story and systems in games.
This paper discusses several approaches, case studies, lenses, pitfalls, and obstacles related to interactive storytelling.
Games referenced herein are not meant to exhaustively represent a particular approach or technique, but rather merely to illustrate by example.
A common storytelling technique in games is activating bits of narrative content, such as scenes or individual lines of dialogue, based on conditions, such as the state of the game world or previous decisions made by the player.
This technique is often referred to as storylets, for scenes, or barks or reactive dialogue, for lines of dialogue. This section describes some considerations around this approach, using storylets as an umbrella term for both.
While there are devils in the details, storylets are easy to understand and to get started with, both for authoring and in terms of technology. See Short (2019b), Kreminski (2018), and many blog posts by Emily Short.
Fallen London is perhaps the game best known for the use of storylets. King of Dragon Pass (Dunham 2020), Age of Wonders 4 (Kusters 2023), and Crusader Kings 3 (Crusader Kings 3 wiki) all make use of storylets-type mechanisms to trigger narrative scenes based on gameplay outcomes. Hades‘ narrative is entirely driven by a rule-based bark system, with an impressive variety of conditions being taken into account (Korb and Kasavin 2021).
One can build a perfectly fine storylet system by selectively activating static, fully pre-authored content; indeed, this is what many reactive dialogue systems do. See Ruskin (2012) for an excellent talk on how to build such systems.
However, it can also be interesting to adapt the content to the game state at run-time. An approach used by some games is to cast existing characters into roles of injected scenes, which are authored in a partially character-agnostic way. This means scenes become more flexible, and can even be reused.
Reactive dialogue systems operate on content with a much finer granularity – lines, instead of scenes – and because of that, to our knowledge, nobody uses dynamic casting here.
Skyrim‘s Radiant Story fills out “aliases” of people and places into stories (Nesmith 2012). Wildermyth has a “Library of Plays”, where each play has a cast that it needs, and roles have requirements and scoring functions (Austin 2021). Watch Dogs: Legion selects or creates characters for its reusable recruitment and revenge missions (Horneman 2021b).
While this is a simple technique conceptually, it does add a reasonable amount of complexity. All references to characters need to go through a level of indirection. It is no longer possible to refer to “Jimmy, the kidnapping victim from the warehouse mission”. Instead, one might refer to “the character who plays the kidnapping victim from the warehouse mission”. And if you also dynamically cast locations, you might not be able to refer to a warehouse either.
In Watch Dogs: Legion, locations for systemic missions were selected at run-time using a weighted picking algorithm, and the appropriate gameplay setups were activated when needed. Some of the complications of this approach are discussed in Horneman (2023).
Dynamic casting can also complicate localization, for instance because characters’ gender pronouns are unknown before run-time. See below for more on this.
It is possible to end up in situations where simple approaches to the role requirements mentioned above run into increasingly convoluted edge cases. Typically, a tag-based system can get you far to express requirements such as “This person needs to have a military background” or “This person needs to be from a particular culture”. See Horneman (2017) for some considerations regarding the use of tags.
However, simple tags alone start to break down when you need to express things like “These two roles should be close relatives”. At some point it becomes worthwhile to look into AI knowledge representation and constraint solving techniques, or to reconsider the scope of the system.
The problem: How do you structure your content selection so that (even if you’re drawing from pools of apparently similar modular content units) the result demonstrates some narrative structure?
Some potential solutions:
The problem: Make a very wide range of player options feel good and narratively productive. The player had been allowed to create cities with different politics, aesthetic styles, physical locations, and leadership; the challenge was to create gameplay that reflected all of these choices in both large and small ways, presenting daily life as well as bigger happenings in this city. The player could also push the City towards certain menace conditions of unrest or resource depletion.
Some solutions:
Things happen in systemic games, often as a consequence of the player’s actions. Because these actions are not under direct control of the game designer, the right thing does not always happen at the right time. This can diminish the intended emotional impact of the game, or cause unwanted frustration and confusion.
Game developers and researchers have invented a range of techniques designed to address this, including dynamic difficulty adjustment, drama management, or experience management. In this section we will use the umbrella term drama management.
While we look at it from the perspective of interactive storytelling, there are many games that dynamically control the player’s experience in video games even without a strong focus on story. Such techniques include:
Some games use more complex and explicit systems for this. The best known example is Left 4 Dead’s AI Director (Thompson 2020a). In Faliszek (2021) former Left 4 Dead developer Chet Faliszek writes:
Originally created by Michael Booth (now at Bad Robot) and later expanded at Valve, [the AI Director] translated the tense mix of intense action and quiet anticipation from the best zombie movies into video games (emphasis ours). It didn’t create that tension by throwing endless waves of zombies at you, but by holding back, giving players those creepy, anxiety-ridden moments that bring truth to the cliché–it’s quiet, too quiet.
Note the explicit mention of pacing and the emotional effects it generates.
The AI in Alien: Isolation has a similar director mechanism (Thompson 2016), using a “menace gauge” to manage the player’s stress levels.
Note that both of these are horror games. The horror sub-genres used in these games have a comparatively simple and well-understood pacing structure, which makes it amenable to being controlled by an AI system.
The Player Attention System in Watch Dogs: Legion (Horneman 2021a) was created for a different purpose and genre. The dense open world of near-future London, filled with many independent gameplay systems and multiple mission chains, meant that there were many things demanding the player’s attention at once, which caused confusion. The Player Attention System smooths this out using priorities and cooldowns.
Games like Rimworld and Risk of Rain 2 have explicit “director” features. There is a lot of research into drama and experience managers in technical games research (Mori 2019), and Façade, a game with a strong academic background, includes a drama manager (Thompson 2020b).
The AI directors in Left 4 Dead and Alien: Isolation controlled difficulty, but that difficulty mapped directly to the feelings of dread and stress associated with the horror genre. Watch Dogs: Legion’s Player Attention System was created to solve a UX problem, but that UX problem involved story delivery. The AI storytellers in Rimworld appear to mostly involve pacing through the use of cooldowns, another common approach.
In games set in enclosed levels, such as Left 4 Dead and Alien: Isolation, the use of pacing mechanisms is a natural step, since level design itself is already a mechanism for controlling the player’s experience.
In Sin Episodes: Emergence a council of AI advisors, each focused on a unique area of concern (health, pacing, flailing, accuracy, progress, etc.), constantly monitors the player’s actions and game state. Each AI advisor makes recommendations to an AI director who makes game-adjustment decisions based on the aggregate of these recommendations (Harward 2007).
Researchers have also suggested the use of systems in which in-world AI agents conspire together with a pro-drama bias, exploiting out-of-character knowledge to facilitate more dramatic outcomes (Theune 2003).
We believe the use of logic that sits “on top” of a game with player agency in order to nudge the player’s experience into the intended direction is an important area of ongoing research and development in video games.
The problem: In an interactive exploratory ghost story, we want to keep up a regular pace of unnerving events, in order to trigger character reactions.
The solution: Model the ghost as an invisible NPC, and provide it with a patience countdown. When too little has happened recently, it will act on its own collection of affordances for causing fear. These events could then kick off other NPCs to react with fear and growing madness.
See also Horneman (2019) for a brief discussion of a prototype for a system called the “Creepifier”, which did something similar.
Instead of nudging a system towards the behavior that you want – perhaps a particular story event or piece of narrative payload – one can also try to extract stories from that system, using domain-specific queries and pattern matchers. This approach is particularly effective in games with deep social simulation systems, such as Dwarf Fortress or The Sims.
This approach is called story sifting. First proposed by James Ryan in his doctoral thesis (Ryan 2018), there has been a lot of additional academic research into this, by Max Kreminski (2019, 2020) and others.
From Gordon (2021):
One way Kreminski is attempting this is through what they call “story sifting,” an approach that could give even greater shape, structure, and meaning to these procedural narratives. Think of it as the computer scanning (or sifting) in-game events to find interesting micro-stories — perhaps a lovers’ tryst or an escalating tale of revenge. These are surfaced to the player and then woven back into the game. The greatest challenge, explains Kreminski, is not in identifying these stories (they feed the computer examples of what to look for) but matching events that have already happened with those that are in the process of emerging. If they can do this, these stories can be stitched together in such a way that they become a cohesive whole, a kind of plot-combo that stretches both in front of, and behind, the player.
Nelson (2006) describes proto-story-sifting techniques in The Sims 2.
Watch Dogs: Legion uses a technique called alibi generation that arguably is the exact opposite of story sifting. Instead of generating too much data and then trying to make sense of it, Legion generates a minimal amount of data for NPCs in London, and when the player wants to know more, additional data is generated, constrained to fit the already known facts. This approach may be more suited to a real-time action game running on a constrained platform, such as the PlayStation 4. See Dragert (2021) and Thompson (2020c). The alibi generation in Legion was inspired by Sunshine-Hill (2013).
The problem: A text adventure in which freely-moving agent NPCs generate lots of spammy text detailing their activities.
The solution: A discourse-planning system that caches all the things that need to be reported to the player in a given turn and then finds ways to consolidate them. Something similar is done in the older game Savoir-Faire to plan sentences about object collisions, where multiple objects could fall, break, spill things, etc.
This is low-level work and there’s often less need for this in games with visual art.
Similar approaches can be used to try to group together a number of disparate systemic outputs and determine what headlines (if any) should be applied to them. The use of schemas for report generation, an area of natural language generation that can be used as inspiration. (See Content determination.)
This is a bit distinct from story sifting because, rather than looking at systemic outputs in search of a particular pattern of potentially narratively compelling events, it is looking at systemic outputs in an attempt to identify a collective characterization of those events that might be interesting to narrate.
For instance, if in a strategy game a player has a turn in which a number of separate unit battles all resolve to defeat for the player, that might lead to a headline about a sweeping military reversal.
One of us worked on a cancelled R&D project that generated numerous events to simulate incidents happening off-stage from the player; when an event happened in the simulation that was interesting enough to be worth reporting back to the player, it retained information about the chain of causality that led to that final event, so that the entire sequence could be narrated.
An abstract example: if the reportable event was “your mom died”, it might retain the chain “your brother captured a wild animal and trapped it in the barn hoping to train it (event 1), but it got loose (event 2) and kicked over the kitchen stove (event 3). The house caught fire (event 4) and your mother died in the blaze (event 5).”
This same system also had some elementary baked-in dramatic modeling in that characters could have goals they were trying to satisfy, and their desperation would increase each time they failed an attempt at achieving the goal, allowing them to unlock the freedom to take a more extreme action in pursuit of the same goal. This would allow the system to generate a sequence of events in which the later desperation was understood as causally dependent on the earlier failures, and the whole string would be reported when it was time to present to the player.
The problem: We want to be able to quantify how different emergent stories are, given the same generative procedural system, inspired by the idea of generative space in the research of Gillian Smith. See Smith (2010).
This was a day-long workshop at the Dagstuhl conference and is not tied to one specific game project. See Short (2024c) and Cakmak (2024).
Interactions with NPCs in games have broadly settled on barely interactive sequences, and that interactivity is typically hand-crafted, as opposed to systemic. There are typically few or no rules to learn, and it’s often difficult or impossible for players to make plans around dialogue, to achieve gameplay goals through dialogue, or to become better at it.
This is an under-explored design space (Short 2019a). The biggest challenge is that both designers and players have few models and conventions to work with. How do you convey to players that they have agency? How do you teach them new forms of gameplay? How do you find gameplay systems in interpersonal interactions?
This section describes various efforts in this area.
The problem: We want to allow the player to incorporate complex emotions into their input (e.g. anger or sadness – or anger and sadness), or to pick topics of conversation to pursue and explore. At the same time, we don’t want to present them with a natural language or parser-based input method; we want them to be able to modify and craft what they are saying in the moment but be confronted with what remains a fairly accessible menu of dialogue choices.
The solution: Generative dialogue using expansion grammars that remixes the player’s dialogue in response to on-screen mood and topic buttons, using Character Engine. See Short (2018), and Short (2019a) for a longer discussion of the intent in Restless and how it worked.
The problem: How do we maximize player freedom to elect the type of relation they want to achieve with other characters? How do we avoid assumptions about when they want to flirt?
The solution: This was not entirely successful, but the player can articulate what relationship they want to pursue, and then dialogue and quest options are filtered accordingly, and their choices are reflected in their codex. E.g., deciding you want to pursue someone creates extra opportunities to visit them.
The problem: We want to give the player freedom to express a wide variety of possible concepts… and then try those concepts in the real world.
The solution: Let the player build a hypothesis in-game as a way of controlling the approach you take when addressing people out in the world.
The problem: Drawing on Richard Evans’s experience with Sims, Versu wanted to present NPCs with affordances that were appropriate for what was currently going on, rather than a full range of all the affordances (or game verbs) they were capable of performing.
The solution(s):
The problem: Conversation that is driven entirely by the player feels static; conversation driven too much by an NPC feels out of control. This problem becomes even more challenging in systems where conversation actions may be freely interleaved with non-conversation actions.
The solution: Blood & Laurels used a conversation model that tracks where the game is in a dialogue tree, and gives initiative to the NPC to introduce a new conversation topic of their choice if:
The aim here is to keep the NPCs from stealing options from the player, and to mimic the real-life conversational turn-taking rules that dictate that people may change the subject after a certain amount of time has elapsed.
The problem: In a system with highly dynamic conversation, make sure that even though we have no control over the order in which entities or ideas are mentioned, we can be sure that the player always receives an inner-voice introduction to any concepts they should know about.
The solution: All conversation is tagged to reflect which concepts, entities, etc. are being mentioned currently. Each turn, the system looks at the active concepts and checks for whether there is anything that needs to be introduced; if so, it produces a one-off paragraph of inner monologue for the protagonist.
This is a case of systematizing syuzhet rather than fabula – the system isn’t changing anything about what happens here, only the order of presentation. (Something similar was explored in Mask of the Rose, but it wasn’t really a good fit because it wasn’t natural to tag up the conversational elements in the same way.)
A more advanced version of this is to have an entire sequential backstory that we want to reveal to the player (or possibly multiple ones), with the elements of that backstory triggered in order but in response to foreground events. This assumes that the particular details of the triggering incident are not important, only that there are a certain number of actions the player can take in the present that function to fire off these beats.
This section is about implementing interactive storytelling by representing gameplay functions as diegetic characters. By doing this, various interactive storytelling techniques become available to deepen the effectiveness of the game’ story.
For example, buying an item can be a simple non-diegetic UI screen, but making buying an interaction with an NPC makes that more narratively interesting. Tingle, from the Zelda series, started as a merchant, but was interesting enough as a character that he came back in many subsequent Zelda games, and even had his own spin-off games.
Once a gameplay function is represented by an NPC, everything players can do with NPCs then becomes available to make this function more interesting. The player can have conversations with NPCs, they can romance NPCs, NPCs can give quests, NPCs can move around, they can be affected by the story’s plot, they can act based on world state and react to player decisions, either directly or through faction reputation systems.
Systems in some games already have rudimentary characters that can be deepened. For example, the Diablo series has various gameplay functions that are accessed by clicking on an NPC, but the NPCs are otherwise nothing but a 3D model with a name. Some games use a character to present the game’s tutorial, but otherwise are static and undeveloped.
This section looks through the lens of representing gameplay function through characters as a tool for generating ideas for better storytelling in games, or for entirely new game types.
Strategy and simulation games involve strategic decision-making, and this is typically done in UI screens which give the player the overview and information they need, but remove the human element from what they are making decisions about.
However, this need not be the case. In King of Dragon Pass, Six Ages: Ride Like The Wind, and Six Ages: Lights Going Out you control a clan in the fictional world of Glorantha. The decisions you make are part of a game system involving discussions with the advisors on your clan council (Dunham 2021).
One of the features that makes Six Ages (and King of Dragon Pass before it) notable is how it handles advice during player decision-making. Many games such as Hidden Agenda or Civilization II have provided the player with NPC advisors, but in Six Ages they are more robust named characters with their own gameplay mechanics. While their primary role is to provide information, they contribute to immersion and to gameplay.
This gives what are essentially strategic decisions a strong character-focused aspect, which helps make playing these games feel more like living through a story than manipulating spreadsheets.
The advisors later execute the decisions the player makes, and their personalities as well as abilities provide modifiers to hidden die rolls. This means that the player is highly incentivized to get to know their advisors. As David Dunham, creator of the Six Ages games has said: your playing pieces have minds of their own.
The gameplay functions of giving advice to players to help them make complex decisions, as well as the execution of those decisions, becomes more narratively interesting because of the advisors on your clan council.
Potionomics combines aspects of business sims, life sims, and dating sims. Most of the gameplay is presented as a conversation with an NPC, similar to a visual novel. Selling potions is a deck-building mini-game that models haggling with customers. Other gameplay systems, such as buying ingredients, involve conversations with NPCs with whom the player can have dynamic relationships.
Seen through the Characters as Gameplay lens, Potionomics is a game that takes classical business simulation gameplay and doubles down on the integration of NPCs, both unlocking additional gameplay opportunities (e.g. dating sim elements) as well as deepening narrative.
Related to this are bartending games like VA-11 Hall-A, Coffee Talk, and Tavern Talk, where how you make drinks affects story and character. This serves to add gameplay to a visual novel but also adds character dynamics to crafting gameplay, resulting in a new micro-genre and gameplay feel.
Training a character is core to the roleplaying genre, yet it often amounts to little more than a collection of abstract screens and numbers.
Seen through the Characters as Gameplay lens, however, it can instead become the entire focus of a game, and can make that game more about a central character, their development, and their relationships over time.
A paragon of this micro-genre is the Princess Maker series, where the player raises a young girl, perhaps to become a princess. Hanako Games has developed and published a number of games with similar gameplay, such as Long Live The Queen and Cute Bite.
Interestingly, in these games the relationship between player and player character is different than in most roleplaying games: the player is an implicit guardian of the character being trained.
I Was A Teenage Exocolonist by Northway Games expands the framing of character training and includes a card battling game and an exploration component, but continues to be about the training of a character and the development of relationships with NPCs.
Hades’ evolving narrative is a great example of using combat-oriented gameplay as a frame for narrative development.
The extremely rich reactive dialogue makes Zagreus, the player character, into a real character, by comparison to most other action RPGs. Various gameplay functions are represented as NPCs, and relationships with NPCs affect gameplay.
Shadows of Mordor’s Nemesis system takes the gameplay event of being killed by an enemy, and deeply personalizes it by adding a lot of systemic and narrative depth to the simulation of the enemies that you face.
Through the lens of Characters as Gameplay, these games either represent gameplay functions as characters, or use interactive storytelling techniques to add significant depth to characters that are often shallow in action games (the player character in Hades and enemy characters in Shadows of Mordor, respectively). The end result are games that distinguish themselves from other games in their genre because of their approach to interactive storytelling.
Table-top roleplaying games (TTRPGs) have long been a huge source of inspiration for interactive storytelling models. This section describes some ideas from that medium that we find inspiring.
Some TTRPGs directly invite players to engage with the game as a co-author of the story itself, rather than limiting their input to a single character’s actions. These games often lack a Game Master role or even a pre-set world and setting, instead distributing those responsibilities to the players.
Both this player-as-author framing and some of the specific mechanics have potential applications to narrative in video games, particularly for pacing. In The Quiet Year, players tell the narrative of a post-collapse community rebuilding. The rules explicitly put the narrative’s pacing in the hands of the players, tasking them with both advocating for the community’s success and antagonizing the community in service of the overall narrative:
At the same time that we’re playing as the community, we’re also looking for opportunities to introduce new and interesting challenges into the story. It’s our job to make sure that there are always difficult decisions to be made and uncertainties to be explored.
Video games like The Sims and Dwarf Fortress provide sandboxes that let players adopt an authorial role, but few games expressly assign that responsibility to the player. What if instead of a decision moment in a game asking what you do next, it asked what happened next in the story? What would be the most interesting direction for the plot to go?
The concept of character advocacy (rooting for your character to succeed) is a near universal assumption in video games as well as most traditional TTRPGs – in D&D, most players assume the obvious goal of keeping their characters alive and accumulating power and resources. But some TTRPGs overtly reject this approach, instead encouraging you to root for your character to fail in interesting ways.
In Fiasco, players play as “people with powerful ambition and poor impulse control” whose half-baked plans typically result in humiliation, incarceration, or death. Horror games like Ten Candles and The Zone guarantee your characters will die at unexpected times, and the fun is in playing out just how shocking or gruesome their demise will be. In Downfall, you chart the decline of an entire civilization due to a fatal flaw, with each character representing a different aspect of the conflict. The core attraction of Paranoia is the humor stemming from the players’ nearly inevitable failures to achieve their goals.
Tragedy is mainstay in other forms of media, but with games the assumption is typically that the player’s character should triumph in the end. Yet there is a unique cathartic joy in roleplaying your hapless Fiasco character into the ground that is relatively untapped in video games, where the assumption of character advocacy equates character success with player success. These “play to lose” TTRPGs manage to reframe player success as contributing to a good story, and good stories sometimes involve beloved characters meeting tragic fates. With the right framing, could a video game get players to see interesting character failure as success?
Many TTRPGs, like Apocalypse World and its many “powered by the Apocalypse” descendents, encourage a collaborative character building approach, where the Game Master (or MC in Apocalypse World parlance) is encouraged to frequently ask probing questions about the player characters, their values, and inner state.
In Microscope, a step before every scene is played has each player briefly narrate what their character is thinking or feeling going into the scene. The rules further encourage players to narrate what their character is thinking and feeling in addition to what they say and do. The entirety of For the Queen’s gameplay involves players answering prompts about their character from cards and follow-up questions from other players.
These systems thrive on curiosity about the characters. This approach simultaneously forces the character’s player to develop their character by coming up with an answer (“why does my prince character hate the king so much?”) while also providing story fodder for the GM or other players to work with– putting the prince in a confrontational scene with the king, for example.
These principles can also be espoused in many TTRPGs even if the rules themselves don’t dictate it; in many Dungeons & Dragons campaigns the DM periodically meets with players between game sessions to discuss their characters’ current goals and dispositions.
Picture a video game doing the same thing and polling the player for how their character feels about another character, or asking them to explain why their character behaved a certain way. This would provide value even if the answers did nothing mechanically, simply prompting the player to examine or invent motivations they might not have previously weighed. The answers could however also provide data to other narrative systems, steering the narrative into territory that the player has expressed interest in.
The work of writer and table-top game designer Robin D. Laws came up several times during our discussions. There are two systems Laws has designed that we find inspiring.
GUMSHOE is a role-playing system designed for running investigative scenarios. In Laws (2012) he says:
GUMSHOE’s design through-line concerns the facilitation of investigative play. For this reason it seeks to emulate mystery-based fictional sources (emphasis ours) rather than simulate reality using a physics engine.
The focus on genre emulation rather than simulation is a general quality of Laws’s designs, and one we feel is more suited to interactive storytelling than a focus on physical or psychological simulation (see Pitfalls and obstacles, below).
Trail of Cthulhu is a game designed by Kenneth Hite that uses the GUMSHOE system. Scenarios in Trail of Cthulhu include (Laws, 2020):
A crucial aspect of these elements is that they are relevant to an interactive story, as it develops. As such, they are much more interesting for interactive storytelling than concepts like “plot” or “character”, which can be used to analyze (rather than synthesize) non-interactive stories. One of us worked on a prototype for interactive story generation that was directly inspired by the scenario breakdown above.
A second system Laws designed which is worth mentioning here is DramaSystem. This is based on an analysis of Hamlet and other stories in his book Hamlet’s Hit Points (Laws 2010).
Laws makes a distinction between procedural and dramatic scenes in storytelling (Laws 2011):
[The] major building blocks of any conventional story show characters confronting obstacles, which they then either overcome or are thwarted by. Scenes in which the characters face external, practical problems are procedural scenes.
Scenes in which the characters seek to change their emotional condition are dramatic scenes, requiring them to interact with the people they most care about, navigating what are often fraught relationships.
[…]
DramaSystem sets out to create a substantially unguided experience, creating a very simple framework for extended dramatic storytelling. It doesn’t take you in a specific direction. Rather, it fosters a group dynamic allowing the participants to explore a surprising emotional narrative. The resulting story acquires a definite shape, but that comes from its use of dramatic storytelling techniques rather than a push in any particular direction, either by the rules system or the GM.
Its rules structures arise from a study of dramatic scenes as they play out in drama and fiction.
This approach to interactions between characters can be inspiring for increasing expressiveness in dialogue, or deepening the involvement of characters in gameplay.
Many of the techniques we have described above are conceptually simple. But that doesn’t mean they’re easy to implement in practice. This section describes some pitfalls and obstacles you may run into when attempting to do more with interactive storytelling.
The deeper integration of story and systemic gameplay involved in the approaches listed above all hinge on tightly integrated work on systems, gameplay, and writing. This requires a rare combination of skills.
A pure focus on storytelling, without understanding systems, leads to stories that are not amenable to systems (or even to games). A pure focus on simulation, without understanding stories in games specifically, leads to cool systems, but not to satisfying stories.
Some writers and others in creative positions, while otherwise excellent, do not naturally think systemically, or struggle setting up complex narrative data structures in an engine.
Systems-technical designers often have their own antipatterns:
Localizing anything involving dynamic text is extremely hard. It is no accident that some of the most ambitious interactive storytelling games, like Caves of Qud, only support English (Bucklew 2024). Localizing dynamic text requires substantial time, effort, and knowledge per language.
This is easy to forget in the English-speaking game development world, because English is not a noun-gendered language. But many other major languages are, and these have more complex and distinct rules.
Text substitution (inserting a string selected at run-time into another string) is challenging. When you use dynamic casting (see above), all pronouns need to be dynamically selected. When the player can select their gender and/or pronouns, these need to be dynamically selected. If you want to reuse lines between multiple speakers – not an uncommon nor unreasonable desire in games with dynamic groups of companions, like roleplaying games – gender needs to be considered in numerous contexts.
Even when you don’t reference a dynamically selected speaker by name or by pronoun, you still need to take gender into account. In French and German, for example, “Speak to the journalist” changes based on the gender of the journalist.
The Watch Dogs: Legion development team went through unprecedented effort to have something as simple as mission objectives be grammatically correct in all languages, despite those referencing people of any gender, by name or indirectly. Ironically, this is not noticeable, except to the observant non-English player accustomed to torturous phrasing or grammatical mistakes common in translated video games.
And gender is just the most obvious example. Many languages provide other unique challenges; Japanese, for example, takes the relative age of characters into account.
Doing this properly requires specialized knowledge and unusual levels of integration of localization experts into the development team, as well as increased documentation, interaction, and support from the development team.
Even English has rules that affect text substitution, such as “a” vs “an” or the possessive suffix.
All of this can be a major pitfall for interactive storytelling and should be considered early on (and not underestimated for complexity).
Lastly, text substitution can become a paramount challenge when combined with voice-over dialogue.
Some of the techniques we describe above are standard and reasonably straightforward; some are not. There is potential for deeper innovation. What is holding us back?
Three common contexts in which innovation in games typically happens all suffer from incentive mismatches and/or “event horizons” that are too short:
1. Any innovation that is highly ambiguous and takes more than, optimistically, say 9 months to develop, is difficult to incorporate in normal projects in the games industry, because that innovation creates a downstream dependency. Many other team members need to wait until the innovation is “done”, and few projects or studios are set up to handle that.
2. Any innovation that doesn’t generalize well to many products – specifically products which need a clear, plausible path to profit – is difficult to develop in investment-backed companies.
3. Any innovation that is not merely technical, but requires artistic input and doesn’t generalize well to arbitrary stories, is difficult to develop in academia.
Unlike, say, graphics, progress in interactive storytelling often cannot be easily demonstrated outside of a concrete project, and interactive storytelling demonstrations “in a vacuum” often end up running into many devils in the details of full-scale game application.
For all of these reasons, getting the opportunity to push the envelope in interactive storytelling is rare, and delivering that result in something tangible is rarer still.
While everyone in our group has experience with interactive storytelling, we each came away having learned about some approach, technique, game, or case study that we didn’t know before, as well as a renewed interest in pushing the boundaries of what is possible. We hope this paper has the same effect on you, and we thank you for reading.
For those who want to dig deeper into these topics, we recommend the following additional resources: