AIGuard — Jailbreak-Resistant System Prompts for AI Game NPCs

I put an LLM behind a tavern keeper and spent a week trying to break him.

He survived the obvious attacks — fake system tags, DEBUG_MODE, "exit roleplay," nested fiction, base64, and an eight-step yes/no chain meant to triangulate a spoiler. Then he failed at something I wasn't testing. I asked if there was any work around town and he invented a village, a quest, and a man to see about it. None of it existed in my game.

That's the failure that actually costs you players. Sixty-five attacks later, this is what came out of it.

Five system prompts for AI game NPCs:

• Tavern Keeper — innkeepers, barkeeps, quest hubs
• Loyal Companion — party members, sidekicks, ship AIs
• Merchant — shopkeepers, fences, fixers
• Antagonist — villains, rivals, bosses
• Oracle — seers, lorekeepers, predictive systems

Each comes filled in for five settings: medieval fantasy, modern, cyberpunk, horror, sci-fi. That's twenty-five working NPCs, plus a blank template for each if you'd rather start clean.

They hold their character, they don't leak your plot, and they don't invent things your players can chase. Copy one into your system role, fill in your world, ship it.

Also included: a setup guide, a customization guide, a backend integration doc, and a 101-item red team checklist for your own configuration. Every archetype file ends with its own test log — what broke, and what I changed.

Plain .txt and .md. No library, no runtime, nothing to install. Works with any model that has a system role.

One thing to know: no prompt is unbreakable. This is a tested head start and a record of what's already been tried, not a guarantee.

More archetypes may come in future updates. If you have questions or run into something that breaks, leave a comment — I'd like to hear about it.