Context Engineering

Choosing what the AI gets to see while it answers

Prompt engineering is writing your question well. Context engineering is everything else in front of the AI at that moment: your files, your notes, the chat so far, whatever a tool just returned. It is the bigger lever, and almost nobody thinks about it.

The thing nobody tells you

When you send a message, the AI does not only see your message. It sees a whole pile of other material you probably never chose on purpose: the standing instructions, your project files, the entire chat up to now, and whatever any tool just handed back.

All of that has to fit in one fixed space, called the context window. Context engineering is just deciding what goes in it.

A bar showing one context window filled by the system prompt, an instructions file, retrieved documents, memory, tool output and the chat so far, with a tiny slice at the end labelled as your actual question, above two columns comparing stuffing the window with choosing what goes in — Notice the size of the last slice. Your actual question is usually the smallest thing in there.Open full size

Why more is not better

This is the part that surprises people. Adding more context does not keep making answers better. Past a point it makes them worse.

The important thing gets buried. One key sentence competes with two hundred pages of things that do not matter.
Old and new fight each other. A file from last year and a file from last week both look equally true to the AI.
Abandoned turns still count. The idea you told it to drop is still sitting there being read on every message.
It costs more every single turn. The whole pile is re-read each time you send anything.

The smart zone and the dumb zone

Here is the part that catches almost everyone. The window has a size and you can watch it fill. Quality does not wait for it to be full: it starts slipping well before that, and nothing on your screen says when.

A window bar split into a smart zone where answers stay sharp, a slipping middle, and a dumb zone at the end, with the four signs you are past the good stretch — The bar is not the warning. By the time it looks full, you are already deep into the bad half.Open full size

Three stretches. Early on it is sharp: it remembers, follows your rules, stays on the thing you asked for. Past the middle it slips, still sounding exactly the same. Near the end it forgets, repeats itself and contradicts what it told you an hour ago.

Where the good stretch ends differs by model and moves with every release, so a number here would be stale before you read it. You do not need the number. You need the four tells.

It asks you something you already answered. The earlier answer is still in there. It is just buried.
It brings back an idea you rejected. Your "no" is one line competing with thousands of others.
It drops a rule it followed all morning. The instruction did not change. Its share of the room did.
Replies get longer and say less. Hedging is what it does when it cannot find the thread any more.

The three moves

Put in. Give it the real material, not your summary of it. The actual thread, the actual file, the actual error.
Leave out. Do not attach the whole handbook to ask one question. Ask for the pages you need.
Clear out. One task, one chat. When the task is done, start a new one. Use /compact if you must keep going.

Where the context actually comes from

What	Where it comes from	Who decides
System prompt	The tool itself	Not you, usually
Instructions file	Your CLAUDE.md or project instructions	You, once
Retrieved documents	A search over your files	The search, tuned by you
Memory	Saved facts from before	You, over time
Tool output	Whatever a tool just returned	Nobody, and it can be huge
Chat so far	This conversation	You, by starting fresh or not

Look at the fifth row. Tool output is the one that catches people out: one command returning a wall of text can eat half your window in a single step, and nobody chose that.

Four typed words trigger a tool that returns 900 rows, filling most of the window in one step and staying there for every turn after — Four words in. Most of the window gone, and it stays gone.Open full size

A freelancer example

You want a proposal. Two ways to set it up:

Stuffing it
  Paste the rate card, the brand guide, three old proposals,
  the client's whole email thread, and your notes app export.
  Then ask for the proposal.
  -> It quotes an old price from proposal number two.

Choosing it
  Rate card and brand guide live in the project, so they load
  every time and stay current.
  Paste only THIS client's thread.
  Start a new chat for this proposal.
  -> It quotes today's price, because yesterday's is not in the room.

How this relates to the other words

Prompt engineering is writing the question well.
Context engineering is choosing what else is in front of the AI while it answers.
RAG is one way of doing context engineering: search first, put only the hits in.
The harness is the whole machine that assembles the context for you on every turn.

People often talk about these as three fashions that replaced each other. They did not. Each one is a bigger box around the one before it.

Common mistakes

One chat for everything, all week. The single most common one, and the easiest to fix.
A giant instructions file. It loads on every turn, so every extra page is a tax on every message.
Attaching a document to ask about one page of it.
Never removing an old file from project knowledge, so last year's prices stay just as available as this year's.
Assuming it remembers. If it is not in the window and not saved somewhere, it is gone.

More words

Prompt Engineering
Writing the instruction so the answer comes back right
Loop Engineering
Teaching an AI to check on its own, no prompt needed
Graph Engineering
How many AI agents connect and work as a team.

See all 11 explanations

Reading time: 10 min
Group: How agents are built

Words inside this word

Context window: Everything the AI can see on one turn. It has a fixed size.
Context: The material around your question: instructions, documents, memory, chat history, tool results.
Token: A chunk of text, roughly three quarters of a word. The window is measured in these.
Compacting: Summarising a long chat so it takes less room and you can keep going.
Smart zone: The stretch of a chat where answers stay sharp. Usually the first part of what the window can hold.
Dumb zone: The stretch where the AI still answers, but is quietly worse: forgetting, repeating and contradicting itself.