🌐 日本語
Part 2: Understanding the Context Window
NOTE
Learn the structure of the LLM's "thinking space."
The physical basis for why the structural problems in Part 1 occur is here.
Window constraints are not product-specific. Claude Code is used as a concrete example of injection.
Three Concepts to Fix First
Token, Context, and Context Window are prerequisites for every later design decision. Each is defined on its own page.
| Document | Content |
|---|---|
| Token | The LLM's processing unit—neither character nor word |
| Context | Everything passed in one inference; history is resent each turn |
| Context Window | The ceiling, and the fact that the safe range is smaller |
| How the three relate | Relationship diagram for unit, content, and ceiling |
What You'll Learn in This Part
| Document | Content |
|---|---|
| Chat / Session | Why Context grows over time |
| What the LLM "Sees" | The view inside the window |
| Injection Timing | When each setting enters Context |
| Context Budget | How to allocate finite tokens |
Why This Sits Between Part 1 and Part 3
Part 1 shows what goes wrong. Part 2 shows the structure of the thinking space. Part 3 onward shows where to put controls.
Copying settings without knowing runtime constraints hides why those limits exist. This Part is that physical layer.