I cut Claude API costs by 90% with prompt caching. Here's what I learned before I had to shut it down.

A solo developer built Scrollbook, an AI dungeon master Discord bot using the Claude API, which grew to 867 servers but cost $10–11 per one-hour D&D session due to reprocessing the full conversation context on every API call. By implementing prompt caching on structurally identical system prompts and campaign context, session costs dropped to $0.50–1.50, a 90% reduction, but the project was shut down because the business model couldn't sustain the usage without revenue.