ca1zer/liz — reverse-engineered prompt

Build me a lightweight TypeScript framework for creating AI agents called Liz. I want it to feel simple and transparent, not like a black box. Developers should be able to define an agent personality, control the prompts and model calls directly, add routes for different behaviors, and run requests through clear middleware steps like validation, memory loading, context wrapping, memory saving, and routing.

It should support OpenAI or OpenRouter style LLM calls, text generation, structured responses, and image description. Add a memory system using Prisma with SQLite or Postgres so agents can remember conversations by user, room, and agent. Include a Looker style module that can analyze content, summarize tweets, and load character info.

Also include a Twitter client that can post on an interval, monitor mentions, reply in threads, respect rate limits, and use the agent memory. Make it easy to run locally with env variables, database setup scripts, and Docker Compose. Keep the code clean, modular, and developer friendly.