What is context engineering?

Before a babysitter's first night, you leave a note on the fridge. Bedtime is eight. The little one is allergic to peanuts. Here's the number to call. You don't leave the family history and every photo album. The sitter only needs what helps with tonight. Choosing what goes on that note is a skill, and with artificial intelligence (AI) tools it has a name: context engineering.

When you ask a chatbot something, the model sees far more than your question. It sees everything in its context window, which works like a working memory for the task: the instructions it was given, the earlier messages in the conversation, any files you attached, and the results of any tools it used. Anthropic's documentation counts all of that toward the window. The answer is built from the whole pile.²

Context engineering means choosing that material on purpose. Five kinds of material can go in.

• Instructions: what the job is, who it's for, and what a good answer looks like.
• Examples: one or two samples of the result you want.
• Files: the documents the answer should come from.
• The conversation so far: what has already been said and decided.
• Tool results: what came back when it searched the web or looked something up.

Prompt engineering is mostly about the first of these, the wording. Context engineering covers all five.¹

For a single question, careful wording goes a long way. But the tasks have grown. People now ask AI tools to read a contract, compare it with last year's, and draft a summary for a manager. No phrasing can make up for the wrong contract, or a missing page. Anthropic, which makes the Claude models, describes context engineering as the natural next step after prompt engineering: managing everything the model sees across a whole task, and keeping it useful as the task goes on.¹

It's tempting to give the model everything, just in case. That tends to backfire. Anthropic's engineers describe a problem called context rot: as more text piles into the window, the model gets worse at finding and recalling what's in there. Picture the babysitter handed a binder. The allergy is in there somewhere. So their advice is to look for the smallest set of useful information that gets the job done.¹²

This matters most for AI agents, the tools that take a series of steps on their own. Every step adds something: a search result, a page it read, a note on what it tried. Over a long task that pile grows fast, so someone has to decide what stays. Anthropic describes a few ways teams handle it. An agent can summarize the work so far and carry on from the summary. It can keep notes outside its window and look them up later. Or it can hand a side task to a helper that reports back in a few lines.¹

You can try this today, with any chatbot. Before a bigger request, think about what a smart newcomer would need on the fridge: the goal, the one or two documents that matter, and an example of what good looks like. Leave the rest out. Which file would you attach first?