Design an Interactive Story-Generation AI (Multi-User)

Problem Statement

Design an interactive story-generation platform where a small group of 4–8 users collaboratively writes a narrative with LLM assistance. Each user can type directly into a shared document, request AI-generated continuations, rewrite segments, or accept and revise suggestions from the model. The platform must track story context across turns, handle conflicting edits when multiple users propose changes simultaneously, store a complete version history with the ability to revert, and filter inappropriate content without disrupting creative flow.

This is not a single-user chatbot. It is a multi-writer real-time document system augmented by generative AI. The two hardest axes are (1) concurrency—multiple users editing the same text region while the LLM streams a suggestion—and (2) context management—feeding the LLM enough story history to produce coherent continuations without exceeding token limits or incurring unacceptable latency.

Why this problem is distinctive

A single-user AI writing assistant (like a basic ChatGPT wrapper) only needs request-response sequencing. A collaborative editor without AI (like Google Docs) only needs operational transform or CRDT convergence. This system must solve both simultaneously: users edit concurrently via CRDT while the LLM streams tokens that must be integrated into the shared document without overwriting human edits.

The platform also introduces a third class of participant: the AI itself. The LLM is not just a backend service—it acts as a pseudo-writer whose output must be treated as a draft proposal that any human author can accept, modify, or reject. This creates a unique version-control problem: AI suggestions exist in a liminal state between 'proposed' and 'accepted' that has no analogue in traditional document systems.

The four architectural planes

1. Collaboration plane: real-time presence, cursor positions, concurrent text edits via CRDT, session membership, and conflict resolution.
2. Generation plane: LLM prompt construction, context assembly, streaming token delivery, suggestion lifecycle management, and acceptance/rejection flow.
3. Persistence plane: version history, branching, snapshots, storage tiering, and event sourcing.
4. Safety plane: content moderation, rate limiting, access control, session privacy, and audit.

A strong answer keeps these planes decoupled. The collaboration plane must continue functioning even when the LLM is unavailable. The generation plane must not block or corrupt human edits. The persistence plane must support instant revert without rebuilding the entire document. The safety plane must filter without introducing latency that kills creative momentum.

Public operating baseline

Sudowrite reports over 500,000 registered users and processes millions of AI-generated words daily. NovelAI operates a subscription service with an estimated 200K+ active subscribers generating fiction. AI Dungeon by Latitude pioneered interactive LLM storytelling and demonstrated that even a single-user narrative system requires careful context management. These are public signals that the category is real; our capacity numbers are stated design assumptions.

For capacity planning, this design explicitly assumes: 500,000 monthly active users, 50,000 daily active users, 20,000 concurrent story sessions at peak, an average of 5 users per session, and a 3× event peak multiplier for evenings and weekends.