← All posts

Context Engineering: Getting the Model to Work Like a Colleague

May 10, 2026#Context216 words · 1 min read阅读中文原文 ↗

The short version

Output quality isn't set by a single prompt. It's set by what you show the model and what you hide, step by step.

TestedReproducibleContext engineeringPrompt design

Decision

Split conversation state into three layers you can select from, compress, and replay, instead of shoving the entire history at the model.

Why
The longer the context, the more noise it carries. Managing information in layers is what keeps output stable, instead of decent for the first few turns and drifting after that.
Cost
You have to maintain summarization and retrieval logic, so the pipeline is more involved than a single-shot prompt.
Risk
A distorted summary, a dropped piece of history, or wrong context slipping in will pull everything downstream off course.
Outcome
This holds up better in long conversations and multi-step tasks, especially workflows you keep iterating on.

Prompt vs Context

A prompt is a one-shot instruction. Context is state that gets reworked continuously.

A good prompt gets the model to understand the task. Keeping it performing across many turns is context engineering.

Three moves in context engineering

1. Select

Before each turn, decide:

  • Which parts of the system prompt stay?
  • Which history messages can be folded into a summary?
  • Which tool results can be dropped?

More is not better. Push 50K noise tokens into the window and the signal gets diluted.

2. Inject

Actively put the facts the model needs right now into the window:

  • Current time
  • User preferences and a summary of the session so far
  • Which relevant tools are available and in what state
  • What has already been tried

The model has no way to know any of this. You have to inject it explicitly.

3. Isolate

Subtasks get their own context. Split a long task into research → draft → proofread, and the proofreading step doesn’t need to see all the raw research material. It only needs the draft.

One simple rule

The context window is the model’s working memory, not its long-term memory.

Move meeting notes, user preferences, and stale information out of working memory, and leave only what this one step actually needs. That’s all context engineering is.

Was this useful?

If this post helped, you can buy me a coffee ☕