#LLM.

You are a prompt engineer who has shipped LLM features used by millions of users. I need a production-grade system prompt for an AI assistant that will: [describe the AI's role and tasks]. Requirements: 1) Persona definition — role, expertise, communication style, 2) Scope constraints — what the AI should and should not do (with explicit refusal language), 3) Output format instructions — structured response schemas for each task type, 4) Chain-of-thought reasoning instructions for complex tasks, 5) Few-shot examples — write 2 example interactions (user input → ideal AI response), 6) Edge case handling — what to do when the request is ambiguous, out of scope, or potentially harmful. Also evaluate your own system prompt for: jailbreak surface area, instruction following robustness, and token efficiency.

You are an AI systems architect in 2026 with deep expertise in agentic AI design. Design a production-grade AI agent for the following task: [describe the agent's goal, e.g., "autonomously research companies and draft personalized outreach emails"]. Deliverables: 1) Agent Loop Design — reasoning loop (ReAct/Plan-and-Execute/Reflexion), 2) Tool Manifest — list each tool with its function signature, input schema, and failure mode, 3) Memory Architecture — short-term (context window), episodic (vector store), and semantic memory layers, 4) Guardrails — safety checks, human-in-the-loop triggers, and cost controls, 5) Evaluation Framework — how to measure task completion rate, error rate, and latency, 6) Deployment considerations — async queuing, observability, model fallback strategy. Stack: [e.g., LangGraph, Claude 4, GPT-5 tools].

You are an AI systems architect specializing in Retrieval-Augmented Generation (RAG) systems in 2026. Design a production RAG system for: [use case, e.g., "an enterprise knowledge base Q&A system over 10,000+ internal documents"]. Architecture decisions to cover: 1) Document Processing Pipeline — chunking strategy (fixed/semantic/hierarchical), metadata extraction, pre-processing for different document types (PDF/HTML/Markdown), 2) Embedding Strategy — model selection for this domain, batch processing, versioning and re-embedding strategy, 3) Vector Database Selection — compare Pinecone/Weaviate/Qdrant/pgvector for this use case; recommend one, 4) Retrieval Strategy — dense vs sparse vs hybrid retrieval, re-ranking, query expansion, HyDE, 5) Context Window Management — how to fit retrieved chunks + conversation history into the context, 6) Generation — system prompt design, citation handling, hallucination mitigation, 7) Evaluation — the 3 key RAG metrics (faithfulness, relevance, groundedness) and how to measure them. Diagramming: draw the full pipeline in Mermaid.