Context Engineering

You might say

After a long conversation, the AI starts forgetting things and going off track. What information should I give it?

Selecting, maintaining, and retrieving the information an AI agent needs for each turnFor a long login-fix task, give the agent the error, relevant files, confirmed constraints, and latest test result—not every old log. Context engineering keeps the signal current and retrieves details when needed.

Know first

Context Window System Prompt Conversation History

Context Management

What should enter the next turn of a long task?

Available material

Context for this request · budget 1000 Token

Current task120 Token

120 / 1000 · reserved for the answer 880

First define what this turn must accomplish; otherwise you cannot judge which material is relevant.

Context engineering aims for relevance, clarity, and maintainability. With a limited budget, compression and exclusion matter as much as inclusion.

How context is selected, updated, and retrieved

Keep high-signal material：Prioritize the current goal, hard constraints, confirmed decisions, and necessary tool instructions.

Update as work moves：Retain new decisions and open questions while summarizing repeated chat and old logs.

Retrieve details on demand：Keep paths, queries, or source indexes as pointers so the agent can read the relevant detail when needed.

Quick checkChoose the best answer

A long task contains requirements, approved architecture decisions, dozens of pages of old logs, and repeated tool output. The agent starts missing the newest constraint. How should the next turn be prepared?

Keep the goal, constraints, architecture decisions, and open issues; compress old logs and retrieve originals by path only when neededInsert the entire history and every raw tool result again so no word can be omittedMake only the final instruction more forceful and leave all history and tool results unchanged

You can say this to an AI Agent

This task is long, and you have started missing a newly confirmed constraint. Before continuing, prepare a working context with the current goal, protected scope, approved decisions, relevant files, and open issues. List repeated logs and obsolete attempts as material to exclude, while keeping paths for on-demand retrieval. Show me the result for confirmation before proceeding.

Learn next

1Model Context Protocol→2Skill→