Prompt Engineering Patterns

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

What's the problem?
├── Output format is wrong / inconsistent
│   └── Apply Few-Shot Learning → reference/few-shot-learning.md
├── Model skips reasoning steps / wrong answers on complex tasks
│   └── Apply Chain-of-Thought → reference/chain-of-thought.md
├── Need to explore multiple solution branches
│   └── Apply Tree-of-Thought → reference/chain-of-thought.md#tot
├── Need highest accuracy (can afford 3-5x cost)
│   └── Apply Self-Consistency → reference/chain-of-thought.md#self-consistency
├── System prompt is vague or producing variable behavior
│   └── Apply System Prompt Design → reference/system-prompts.md
├── Prompts are too long / expensive / slow
│   └── Apply Prompt Optimization → reference/prompt-optimization.md
└── Need reusable structured prompts across agents
    └── Apply Template Patterns → reference/prompt-templates.md

Key Patterns

Instruction Hierarchy (Universal Rule)

Always structure prompts in this order:

[System Context / Role]     <- who the agent IS
[Task Instruction]          <- what it MUST DO
[Constraints]               <- what it MUST NOT DO
[Examples / Few-Shot]       <- show, don't just tell
[Input Data]                <- the actual input
[Output Format]             <- exact expected output shape

Quick Examples

LangGraph — CoT Node

from langchain_core.messages import SystemMessage, HumanMessage

COT_SYSTEM = SystemMessage(content="""You are a senior software architect.
When analyzing a problem:
1. State your understanding of the problem
2. Identify constraints and requirements
3. Consider 2-3 solution approaches with trade-offs
4. Select the best approach with justification
5. Outline implementation steps

Always show your reasoning explicitly before giving a recommendation.""")

def analysis_node(state: AgentState) -> AgentState:
    response = llm.invoke([COT_SYSTEM, HumanMessage(content=state["problem"])])
    return {"analysis": response.content}

Google ADK — CoT Agent

from google.adk.agents import LlmAgent

analysis_agent = LlmAgent(
    name="analysis_agent",
    model="gemini-3.1-flash",
    instruction="""You are a senior software architect.
When analyzing a problem, follow these steps explicitly:
1. State your understanding of the problem
2. Identify constraints and requirements
3. Consider 2-3 solution approaches with trade-offs
4. Select the best approach with justification
5. Outline implementation steps

Always show your reasoning before giving a recommendation.
Output format: structured analysis with sections for Understanding, Options, Decision, and Next Steps."""
)

Reference Files

Post-Code Review

After writing prompts or agent instructions, verify:

• agentic-ai-reviewer — graph correctness, system prompt quality in LangGraph agents
• security-reviewer — prompt injection defense, no sensitive data in prompts