Context Engineering

Deciding what an AI agent knows when it works: which files, rules, and examples load into its context window, and when. The discipline behind CLAUDE.md, skills, and MCP.

All Dictionary

What it is

Context engineering is the craft of controlling what an AI agent knows at the moment it does your work. Not just writing a good prompt, but designing the whole information environment: which rules always load (CLAUDE.md), which knowledge loads on demand (skills, local markdown), and which lives behind tools the agent can query (MCP). The output quality of an agent is mostly a function of the context it was given.

Why this matters for designers

When AI output ignores your design system, the instinct is to blame the model or rewrite the prompt. Usually the real problem is context: the agent never saw your token names, your component rules, or your taste. Designers are well-suited to this work: it is information architecture for a machine audience. Your judgment, written down and loaded at the right moment, is what makes an agent produce work that looks like yours.

How it works in practice

  1. Always-on context: a lean CLAUDE.md with the rules that apply to every task.
  2. On-demand context: skills and local markdown files the agent pulls in when relevant.
  3. Queryable context: MCP servers that let the agent ask your design system questions directly.
  4. The skill is curation: context windows are finite, and burying the signal is as bad as omitting it.