I Built an AI Visual Novel - Here's How It Works

Each story is defined as scenes with clear setup, beats, and branch choices. Milestone scenes are fixed to protect pacing and payoffs. Between those points, specific beats can request AI generation with prompt hints and relationship context. This gives us dynamic conversation without losing route integrity.

We found that making every line dynamic hurts readability and QA. Keeping only selected beats adaptive creates enough novelty while preserving quality. Think of AI as seasoning, not the whole dish.

Character definitions include personality, speech style, backstory, and visual identity. Prompt builders compose this with scene setup and affection level so outputs remain aligned with both character and state. We also include short conversation history and trim it to avoid context bloat.

The biggest improvement came from adding explicit emotional targets in prompt hints. Instead of 'continue dialogue,' we instruct tone shifts, narrative purpose, and forbidden directions. This reduced bland outputs and improved continuity.

The app runs on Next.js App Router with route-level metadata and static generation for SEO pages. The reader itself is a client component using Zustand for local game state. We stream generation responses so players see text appear progressively, then parse the final output to normalize expression and cleanup formatting.

State handling is critical. Scene ID, beat index, affection, and history need deterministic transitions or bugs appear quickly. We persist selected fields to local storage for continuity while keeping derived UI state ephemeral.

Access is split between free and premium stories, with checkout handled through LemonSqueezy. A free story lowers activation friction while premium routes fund art and writing expansion. The product decision here is to show quality early, then monetize depth and variety.

In testing, players convert better when they experience a meaningful emotional beat before seeing the paywall. Hard gating too early protects content but hurts trust and reduces paid intent.

Early versions had classic issues: repeated AI phrases, inconsistent character tone, and route dead ends caused by fragile choice mapping. Fixes included stronger prompt contracts, stricter scene typing, and static params generation for all public pages so route links stay stable.

Another issue was aesthetic mismatch. Utility-first defaults made monetization UI feel disconnected from the main experience. Unifying typography, colors, and motion across reader and marketing pages improved perceived quality and trust.

Start with one short route and make it excellent before scaling content. Instrument key events from day one, especially story starts, branch completion, and paywall views. You need these signals to improve conversion and pacing.

Second, build an editor workflow early. If updating character constraints requires digging through code, your narrative velocity will collapse. Data-driven content plus reusable templates is the difference between a hobby build and an operational product.