---
id: con-context-engineering
schema: concept/1
layer: garden
status: established
name: Context engineering
aka: ["giving AI the right information", "prompt context", "briefing the AI", "grounding"]
summary: >
  Choosing and arranging everything the AI can see, such as instructions, examples, files, past
  messages and connected data, so it has what it needs for the task and little that distracts it.
first_seen: "2025-06"
relations:
  builds_on: [con-prompting, con-context-window]
  evolved_from: [con-prompt-engineering]
  related: [con-memory]
  parallels:
    - idea: "briefing a contractor"
      field: "project management"
      note: "A good brief hands over the plans, the constraints and the history, not just a one-line request."
    - idea: "mise en place"
      field: "cooking"
      note: "Getting every ingredient measured and within reach before you start cooking."
implementations: []
competencies: [cmp-describe-the-task]
origin:
  kind: ai-proposed
  by: "seed agent 2026-10-16"
  rationale: "The 2025 reframing of prompting that most modern features (projects, skills, memory, connectors) serve."
---

Your prompt is only part of what the AI sees. It also sees earlier messages in the chat, any files you attached, standing instructions you saved, things it remembers about you, and whatever its tools fetch. Context engineering is the practice of deliberately deciding what goes into that pile.

With too little context, the AI fills gaps with generic guesses. With too much, the important parts get buried, like a 90-page policy pasted in when one section matters. Give it the parts that bear on the task, in their latest version, with a note on what each one is.

This is why AI products keep adding features like projects, memory, skills and connectors. Each one is a way to get the right material in front of the model at the right moment.

## Example

An HR generalist wants a draft parental-leave FAQ. Instead of asking cold, she attaches the actual policy, two real employee questions (names removed), and last year's FAQ with a note: "match this tone, but fix the outdated eligibility rule." The first draft is close to usable.

## How it connects

- It **evolved from prompt engineering**, moving the focus from wording to information.
- It is limited by the **context window**, how much the AI can hold at once.
- **Projects**, **memory** and **skills** are product features for managing context.

## Try it

For your next request, list what a human expert would want on their desk to do the task. Give the AI only the top three items from your list and see how the answer changes.