Iron Law

NO SYSTEM PROMPT WITHOUT ROLE + TASK + OUTPUT FORMAT — vague system prompts produce inconsistent LLM outputs and degrade agent reliability in both LangGraph and ADK. Every agent instruction must state what the agent IS, what it DOES, and what format it RETURNS.

Prompt Engineering Patterns — LangGraph + Google ADK

When to Use This Skill

• Designing or reviewing system prompts / ADK agent instructions
• LLM outputs are inconsistent or hallucinating — apply CoT or self-consistency
• Output format needs demonstration — apply few-shot learning
• Complex multi-step reasoning required — apply chain-of-thought
• Need to explore multiple solution paths — apply Tree-of-Thought
• Reducing prompt token cost while keeping quality — prompt optimization
• Building reusable, testable prompt templates

Framework Decision Tree