{"id":82846,"date":"2026-09-15T09:30:11","date_gmt":"2026-09-15T04:00:11","guid":{"rendered":"https:\/\/tothenewone.online\/blog\/?p=82846"},"modified":"2026-10-27T23:14:08","modified_gmt":"2026-10-27T17:44:08","slug":"from-chatbot-demo-to-full-stack-investigation-game","status":"publish","type":"post","link":"https:\/\/tothenewone.online\/blog\/from-chatbot-demo-to-full-stack-investigation-game\/","title":{"rendered":"From Chatbot Demo to Full-Stack Investigation Game"},"content":{"rendered":"

Introduction<\/h1>\n

A \u201cyou are the game master<\/strong>\u201d chat box is easy. It is also easy to cheat. The model invents a key, walks a lock, or forgets the room. That is improvisation with no referee.<\/p>\n

The Last Broadcast<\/strong>\u00a0is the other design: a short investigation you can finish. AI may turn English into a command.\u00a0A deterministic engine is the only thing that changes the world<\/strong>.<\/p>\n

Play:\u00a0<\/strong>https:\/\/ai-powered-interactive-game-master.vercel.app<\/a>
\nCode:\u00a0<\/strong>https:\/\/github.com\/harshit-ttn\/Build-an-AI-Powered-Interactive-Game-Master<\/a><\/p>\n

If you only have ten minutes: open the live app, finish a few rooms, then in the repo open\u00a0applyAction<\/strong>\u00a0and\u00a0CommandInterpreter<\/strong>. That is the whole argument.<\/p>\n

Why this is worth your time<\/strong><\/h1>\n

If you are wiring a language model into a product, the useful question is not \u201cdid we call the API?\u201d It is \u201cwhat happens when the model is wrong, slow, or missing?\u201d<\/p>\n

This project is a small, concrete answer. Same pattern shows up in assistants that must not invent refunds, bookings, or permissions:\u00a0parse language at the edge; keep the ledger in code<\/strong>.<\/p>\n\n\n\n\n\n\n\n\n
Chatbot demo<\/strong><\/td>\nThis game<\/strong><\/td>\n<\/tr>\n
Model owns the world<\/td>\nEngine owns\u00a0GameState<\/td>\n<\/tr>\n
Hard to test<\/td>\nHandlers and a full win-path test<\/td>\n<\/tr>\n
Dies without a vendor key<\/td>\nKeyword parser still plays<\/td>\n<\/tr>\n
One prompt, one story<\/td>\nSecond case is JSON<\/td>\n<\/tr>\n
No URL, or a toy server<\/td>\nSame-origin deploy<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

The idea<\/h1>\n

Two jobs get mashed together:<\/p>\n

    \n
  1. Parse intent<\/strong> \u2014 messy language \u2192 { intent, target, confidence }.<\/li>\n
  2. Apply rules<\/strong>\u00a0\u2014 can you enter, did you earn the clue, which ending fires.<\/li>\n<\/ol>\n

    Models are useful at (1) and unreliable at (2). We split them.<\/p>\n

    Intents are closed:\u00a0MOVE<\/strong>,\u00a0INSPECT<\/strong>,\u00a0SEARCH<\/strong>,\u00a0TALK<\/strong>,\u00a0USE_ITEM<\/strong>,\u00a0ACCUSE<\/strong>,\u00a0DECIDE<\/strong>. No\u00a0TELEPORT<\/strong>. No invented loot. Illegal action \u2192\u00a0INVALID_ACTION<\/strong>\u00a0and a GM line, not a rewritten universe.<\/p>\n

    Fiction: a hijacked city signal \u2014 \u201cFind the witness before the silence.\u201d Loop: search \u2192 evidence \u2192 theory \u2192 decide at the tower (stop\u00a0\/\u00a0expose\u00a0\/\u00a0continue<\/strong>).\u00a0Cold Harbor<\/strong>\u00a0is another mystery on the same engine. Stories are JSON. Physics are code. Health, trust, score, hints exist so choices hurt.<\/p>\n