2 Instructions in AI agents

Chapter 1 introduced context engineering and the six sources of context that shape agents' behavior. This chapter focuses on the first of these sources: instructions, which establish the rules, constraints, and behavioral expectations under which a model operates. Every interaction with an LLM-based system is interpreted in the presence of one or more instruction layers. Instructions are delivered through different mechanisms in AI agents, including the system prompt, agent skills, and instruction artifacts.

2.1 Instructions

Instructions are the foundational source of context in LLM-based systems that define the model’s operating role, behavioral expectations, safety boundaries, formatting obligations, and policies across a conversation, session, or task run. A useful way to view instructions is as an operating contract between an application and the model. The contract tells the model how to interpret later context, how to resolve routine choices, and which boundaries remain in force even when a user asks for something different. Instructions frame user intent so the model can satisfy valid requests while remaining aligned with the application, organization, and product environment [10].

2.2 The system prompt

2.2.1 Content of the system prompt

2.2.2 System prompts in products and APIs

2.3 Agent skills

2.3.1 The anatomy of a skill

2.3.2 Best practices for designing agent skills

2.4 Instruction artifacts

2.4.1 Instruction artifacts in coding agents

2.4.2 Best practices for designing instruction artifacts

2.5 Hands-on

2.5.1 Using system prompts

2.5.2 Creating custom agent skills

2.5.3 Using existing agent skills

2.5.4 Creating custom instructions artifacts

2.6 Summary