Context Engineering: What Happens When AI Can't Fit Everything in Its "Line of Sight"
> Source commit: 9621a05 | easy-learn-ai project daily update
---
1. A Story About a Tray
Imagine walking into a busy restaurant. A waiter carries a tray. The tray is small—at most four or five plates. Now the kitchen needs to send out ten dishes at once, but the tray can only hold a few. The waiter has to decide: which dishes go first? Which can wait? Which can be stacked to save space?
When an AI works, it faces the same kind of "tray."
Every time you talk to an AI, what it "sees" isn't just your latest message. It also sees your chat history, files you've uploaded, and system-level rules and background knowledge—this entire bundle is the "information in front of it" for the task at hand.
The problem is that this "front" has limited capacity. Just as the tray can't hold ten dishes, the AI's context can't hold too much. Give it too little, and it lacks material and makes things up; give it too much, and it loses focus, gets distracted, or squeezes out important information.
This craft of carefully arranging what's in front of the AI is called Context Engineering.
---
2. Why Doesn't It All Fit?
To understand context engineering, you first need to know what the AI's "front" actually is.
Think of the AI as a student taking an exam. On the desk in front of him, he can only spread out a limited set of reference materials. The size of that desk is called the context window—the maximum amount of text the AI can process at once.
Early AI models had tiny desks—a few pages at most. Today's desks can fit a whole book, but they're still finite. Worse, the bigger the desk, the harder it is for the AI to find things—like searching for a book in a giant library: the bigger the library, the more time it takes.
And the desk doesn't just hold what you provide. It includes:
- Your question — the main character
- Reference material — documents you upload, web pages retrieved
- Chat history — everything discussed so far
- System prompts — hidden instructions on how the AI should respond
---
3. Too Little vs. Too Much: Two Predicaments
Context engineering is essentially about balancing two extremes.
Too little: the AI starts hallucinating
With too little context, the AI enters a "making things up without material" state, known professionally as hallucination—the AI confidently states falsehoods.
Imagine asking someone who has never been to Beijing: "Which hotpot restaurant in Sanlitun is the best?" Without any data and unwilling to admit ignorance, they'll invent a plausible answer: "I think it's called 'Old Beijing Instant-Boiled Mutton House'—supposedly famous." In reality, that restaurant may not exist.
AI does the same. When its context lacks necessary background, it fills the gap with general knowledge learned during training—knowledge that is often imprecise or outdated.
Too much: the AI loses the thread
The other extreme is overloading. Dump an entire manual, dozens of pages of chat history, and a dozen reference documents into the context, and the AI can't tell where to look.
It's like asking someone to read ten documents simultaneously and then answer a specific question right away. Their attention scatters; they miss key details or mix up information from different files.
Research shows that when context is too long, the AI's efficiency in using information drops significantly—it tends to attend to the beginning and end of the context while the middle gets ignored. This is known as the "lost in the middle" phenomenon.
---
4. Four Basic Moves: Write, Select, Compress, Isolate
So how do you solve this? The industry has distilled four basic strategies—like the waiter's four techniques:
1. Write — store it outside first
Some things aren't needed now but may be later. Rather than occupying precious tray space, jot them on a "sticky note" and retrieve them when needed.
AI can do this too: it can write intermediate conclusions or temporary search results to external storage (a file or database) and read them back later. Its context keeps only the most relevant current information.
2. Select — put in only what's needed
Don't stuff in all the material; pick only the parts most relevant to the current question.
It's like a doctor's consultation. A patient's full medical record might be hundreds of pages, but the doctor pulls out only the test reports and medication history relevant to the current symptoms—not decades of complete records.
In practice, this is usually achieved through Retrieval-Augmented Generation (RAG): first search for the most relevant document fragments based on the user's question, and put only those fragments into the context instead of the entire document library.
3. Compress — turn the long into the short
As chat history grows and fills the window, old conversations can be summarized into brief digests, freeing space for new content.
Imagine a very long meeting. Every hour, a secretary condenses the discussion into a "here's what we've agreed on so far" summary. Later participants only need that summary, not dozens of pages of verbatim transcript.
AI can do the same: periodically compress accumulated context into summaries, preserving the core information in shorter text.
4. Isolate — keep things separate
If one tray can't hold all the dishes, get more waiters, each carrying a tray with a few dishes.
In the AI world, this is the multi-agent architecture: split a complex task into subtasks handled by different AIs, each with its own context window, without interference. Then merge the results.
---
5. The Sweet Spot of Context Engineering
Good context engineering is neither "as little as possible" nor "as much as possible." It pursues a sweet spot—just enough information for accurate judgment, without overwhelming the model.
Finding it involves several factors:
Recency — recent events matter more than ancient ones. Place the newest information at the end of the context (the AI is more sensitive to endings); compress or remove old information.
Relevance — directly relevant material goes in; marginally relevant material waits outside; irrelevant material is cut entirely.
Density — express the same meaning more concisely. Remove filler, repetition, and decoration; keep only the core facts.
Structured presentation — organizing information into clear formats (lists, tables, hierarchies) makes it easier for the AI to understand and use than walls of text.
---
6. Why Is Context Engineering Trending Now?
The term has only been talked about frequently in the past year or two—not because it wasn't needed before, but because of two changes:
First, AI got more capable. Early AI struggled with basic conversation, so nobody expected it to handle complex context. Today's AI is much smarter and can do far more—and people have realized that what limits it is often not its "brain" but whether the information in front of it is right.
Second, applications got more complex. Simple Q&A chatbots have evolved into complex workflows requiring retrieval, tool use, multi-step reasoning, and long-term memory. These scenarios demand far more from context management.
There's a good saying: "Prompt Engineering is how you talk to AI; Context Engineering is what you show AI." The former is rhetoric; the latter is curation. Combined, they unlock AI's true potential.
---
7. Summary
The essence of context engineering is the craft of information curation. Its core question: within the AI's limited "line of sight," what should go in, what should stay out, and how should it be arranged?
Like that restaurant waiter's tray—a good waiter never leaves it empty (guests go hungry) nor piled to overflowing (things get spilled). He carefully arranges what each trip carries, what goes first, what gets replenished, ensuring every dish arrives at the right moment.
In the AI era, every one of us can become such an information curator.
---
*This article is based on the new "Context Engineering" module added in commit 9621a05 of the easy-learn-ai project.*