Designing the Vision of AI
I understand the meaning of context engineering, which designs the range, location, and compression method of information that AI can see when generating answers.
Strange Moments in Long Conversations with AI
When you talk with AI for a long time, strange moments arrive.
Conditions you clearly stated earlier suddenly get answered differently by the AI. It asks again about background you've already explained, misses the format you asked it to follow, and starts saying things that differ from the criteria you set initially.
If you think of it like a person, you want to say it forgot.
But the more accurate expression for LLM is slightly different. It's not that it forgot, but rather it doesn't see enough or doesn't prioritize it sufficiently when creating an answer right now.
AI Receives a New Desk Each Time
What AI sees when creating an answer is not the entire conversation. Each model has a range of text it can see, and this range is called the context window.
It's easier to understand if you imagine a person writing an answer while only looking at the papers on their desk. If something isn't on the desk, that information doesn't directly enter the process of creating an answer right now. AI doesn't pull out conversation from memory—it looks at the text on the desk right now and creates the next answer.
So when old conditions get pushed out or important criteria become blurred in a long conversation, AI behaves as if it forgot. The desk changed.
What to Keep and What to Remove
When context is insufficient, answers become generic. But when context is too much, answers become blurry.
If unrelated materials, outdated conditions, conflicting instructions, and unimportant conversation history all go in together, AI can waver about what should be the standard.
bad-context:
- entire meeting transcript 30,000 characters
- already-discarded decisions
- past discussions conflicting with latest goals
good-context:
- current goals
- latest decisions
- key materials to reference
- desired output criteria
The difference isn't the amount of information. It's whether the information that should affect the current answer right now is organized.
Context engineering is not about putting in more information. It's about keeping the information you need.
Position Is Also Design
Context is not just about what to put in. Where you place it also matters.
When important conditions get buried in the middle of long input, the model sometimes fails to sufficiently reflect that information. Just as people better remember the beginning and end when reading long documents, the position of information can affect the direction of AI's answers.
That's why important instructions and criteria are often placed at the front, and final output conditions are fixed once more at the end.
Organized practically, it looks like this.
- Put absolute rules to follow at the front
- Fix the final output format once more at the end
- Divide long reference materials into sections
- Distinguish old information from latest information
- Add titles to high-importance contexts
When Conversations Get Long
As conversations get longer, you want to keep all the content as is. But in reality, important and unimportant things accumulate together.
Carrying a long conversation as is can become a burden rather than memory.
What's needed then is history management. Instead of simply appending old conversations, summarize and compress only the content necessary for current decisions. Keep only information you'll use again, like goals, established criteria, prohibited directions, and remaining questions.
[SUMMARY]
What's been decided so far:
- Write as independent short books, not as a series
- subject is a bookshelf, not a series
- Keep slug if possible
[CURRENT TASK]
Design the structure of the prompt-context-engineering article
[DO NOT]
- Don't use expressions dependent on preceding or following articles
- Don't write like a simple API document
This kind of summary doesn't reduce conversation. It makes what AI needs to look at again clearer.
What to Place Before AI
Context engineering is not about throwing more information at AI.
It's about deciding what AI should see right now, what it doesn't need to see, which information should be placed up front, and which information should be compressed.
AI is not a being that remembers everything. It creates answers based on what's placed before it right now.
That's why designing context is close to designing AI's field of vision.