liveBook · Manning

Instruction

As explained in Context Engineering

A directive supplied to an AI system that specifies desired behavior, constraints, or task requirements.

An instruction is a directive in an LLM’s context that defines desired behavior, constraints, or task requirements. Instructions establish roles, safety boundaries, policies, and output expectations, framing how the model interprets subsequent context and responds to user requests. section 2.1

AI agents combine three complementary instruction mechanisms:

  • System prompt: global guidance for the model’s role, goals, tone, rules, and response style, typically supplied before user input. section 1.5.1 section 2.1
  • Agent skills: reusable packages of specialized procedures, checklists, examples, and resources that can be loaded when needed. section 1.5.1 section 2.1
  • Instruction artifacts: agent-readable files containing durable project or environment guidance, such as conventions, commands, validation steps, and safety constraints. section 1.5.1 section 2.1

A context manager assembles these layers into the context window. It supplies global instructions and discovers local skills and artifacts. To conserve the token budget, it can initially load only skill frontmatter—the name and description—and load the full SKILL.md instructions when a task activates the skill. Project artifacts such as AGENTS.md or CLAUDE.md provide durable guidance alongside those instructions. section 2.1

Figure 2.1 Instruction material assembled into the AI agent’s context window. The context manager supplies global system instructions, compact skill frontmatter for discovery, loaded skill procedures upon task activation, and project instruction artifacts
instructions v8

See how global instructions, skill metadata and procedures, and project artifacts enter the agent’s context window. section 2.1

Layering instructions for incident analysis

A system prompt can set the assistant’s professional tone and safety boundaries. An incident-analysis skill can then direct it to summarize the event, extract a timeline, identify the likely root cause, list mitigations, and report unresolved risks. An instruction artifact can add durable project-specific commands and validation requirements. section 1.5.1

How Context Engineering covers it

How deep it goes
Explains three instruction mechanisms and their context-window assembly, including staged skill loading to conserve tokens.
Examples
Illustrates a financial-services assistant, an incident-analysis skill, and repository guidance requiring unit tests and regression tests.
What you can do after
Evaluate and understand how layered instructions provide behavioral guidance while conserving the token budget.
How it connects
Connects instructions to context engineering, system prompts, agent skills, instruction artifacts, and context windows.

Also explained in

Compare how 6 books explain instruction →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Context Engineering. The section links point to the parts of the book it draws on.