Building a chatbot that tells stories

Stories do more than deliver information: they arrange events, emotion and meaning in time. A storytelling chatbot needs a dramatic reason to exist, not a database with a friendly interface. That distinction fits ITP 30 Show, where interactive art, design and technology are experiments in how people encounter digital work.

The process of building a chatbot that tells stories begins with language, then crosses into interaction design, writing, software and performance. A technical model supplies fluent replies; human decisions determine whether those replies create suspense, reveal character, respond to a meaningful choice or lose the plot.

For Australian audiences, design must suit everyday digital habits and a distinct cultural voice. Sydney and Melbourne have lively creative-tech communities, while regional users may rely on patchy mobile connections. Australian English, understated humour and respect for place can give a character personality without caricature.

Begin with the story's promise

A digital archive might invite someone to trace a family history, while an illustrated catalogue could reveal a fictional city through objects and overheard conversations. Write one sentence naming the audience, central mystery and emotional payoff. This becomes the test for every later choice.

Set boundaries early. Is the tale linear, branching or assembled from fragments? Can the visitor return in another mood, solve a puzzle or influence an ending? These constraints protect the premise from loose generated prose and give designers material for scenes.

Success has narrative meaning as well as technical meaning. Completion time matters, but so does whether someone remembers a character, feels surprised or understands why their choice mattered. Use those outcomes to guide prompts, interface labels and tests.

Map dialogue before writing code

Map the conversation before choosing a platform. Story beats, questions, answers, character goals and failure routes can sit on a dialogue map. Intents describe what a visitor wants; entities capture names, places and objects; context and memory determine which earlier details the bot may use.

A useful companion is the ITP journal reflection on crafting interactive narratives, which considers dramatic structure and interactive decisions together. Design the transitions: a revealing object should trigger a scene, not arrive as an unrelated reply.

Keep a strong fallback path for uncertainty, silence or out-of-scope questions. Branching should come from understandable intentions, not random novelty. Visitors should be able to ask for help, repeat information or leave without the story collapsing.

Give the character a clear voice

Give the storyteller a voice before giving it a visual skin. Describe its perspective, vocabulary, rhythm, emotional limits and relationship to the visitor. A brusque archivist, patient guide or unreliable narrator supports different dialogue choices and ways of pacing a reveal.

Translate that voice into a system prompt, templates and examples rather than one adjective such as “friendly”. Specify what the character knows, how it handles ambiguity and whether it may invent details. Grounding passages in a curated knowledge base keeps a mystery coherent and reduces invented facts.

Build a compact conversational engine

A practical architecture can stay modest: a web or messaging interface, a dialogue manager, a language-model API, a content database and a logging service. The manager holds the current scene and relevant history, while retrieval supplies approved story fragments. This separation makes behaviour easier to revise.

Prototype with scripted dialogue and a small set of training examples before adding generative freedom. Test different phrasings of the same intent, including Australian expressions and imperfect typing. Version prompts, story data and model settings together so improvements can be reproduced.

Speed and resilience shape the experience. Visitors across Australia may use phones on trains, buses or variable connections, so responses should stream quickly and remain legible on a small screen. For the Australian market, a lightweight product with transparent hosting costs often suits a small cultural organisation.

Let visuals participate in the plot

Visual systems can extend the narrative rather than decorate it. The related ITP 30 material on machine-generated visual patterns shows how algorithmic imagery can support an authored aesthetic. A chatbot could alter a map, texture or portrait as clues accumulate, with human selection preserving intentional composition.

Connect images to story state. Sydney’s sudden rain, Melbourne’s changeable forecast or a fictional coastal town can drive a scene, but the system needs a reliable weather source and offline alternatives. Captions, alt text, reduced-motion settings and strong contrast matter when words, images and sound work together.

Test for surprise, safety and trust

Testing needs both a script and surprise. Scripted cases check known paths, contradictory choices, repeated requests and inappropriate replies; open sessions reveal unexpected language. Invite participants with different ages, cultures and levels of digital confidence, then watch where they hesitate rather than merely asking whether they liked it.

Track scene abandonment, repeated questions, corrections, clue discovery and comments about tone. Review conversation logs with consent and remove sensitive details. A character can remain surprising without manipulating users, and a reveal can feel magical without concealing how the system works.

Set data and safety rules during design. The Australian Privacy Principles should shape collection, use, storage and disclosure of personal information, especially when visitors share names or life stories. Provide deletion options, limit retention, define escalation routes and make clear when a human takes over.

Invite the audience into the story

Release a small pilot, observe real behaviour and improve weak transitions. Present the work with clear credits, project notes and an event listing so audiences know what they are entering. The ITP 30 Show’s combination of project descriptions, author credits and RSVP information offers a model for a participatory invitation.

Document creative and technical roles separately, including writing, interaction design, engineering, visual production and facilitation. Share related programme resources, such as a scholarship or census, only when they support the visitor’s next step. This turns a showcase entry into a doorway into the wider ITP community.

Visit the ITP 30 Show, choose one experimental project and trace how it frames participation. Then build a small chatbot scene, test it with someone in your community and publish what you learn. A story becomes compelling when its system is transparent, its choices matter and its audience has a genuine role in the telling.