Context window

Meaning

The context window is the maximum length of text a model can read at once. It is measured in tokens and ranges from tens of thousands to hundreds of thousands depending on the model. Everything in the prompt, the system instruction, the story bible, previous chapters, the writer's request, and the answer the model generates, has to fit inside it. The context window is not memory. A model remembers nothing between requests. A bible shown in the last request does not exist in this one unless it is included again. When a writing tool says "the AI remembers your work," it means the tool re-inserts the bible on every request.

Why it matters

A long novel is always larger than the context window. A 200-chapter web novel exceeds a million characters, and no model reads that in one pass. So the central design question of a writing tool is what to put inside the limit. The conventional priority is: a short style sample, the bible entries for the characters, places and open hooks in this scene, the last 1,500 or so characters of the previous chapter, and the request for this scene. That combination must finish within a few thousand characters to leave room for the answer. When the context overflows, the tool truncates from the front and the model writes without knowing what was cut. Many continuity errors, the left-hander from chapter 3 suddenly holding a sword in the right hand, originate here. SparkLore's story bible selects only the entries that appear in the current scene for exactly this reason.

Example

Composition of one request (illustrative, rough Korean-character counts)

Item	Size
System instruction (inserted by the tool)	600
Style sample (excerpt of the writer's prose)	800
Bible excerpt (Yuri, Doyun, the bankbook, two open hooks)	700
Tail of the previous chapter	1,500
This scene's request	300
Input total	3,900
Room for the answer	the remainder

The key is excerpting 700 characters of bible for this scene instead of inserting the whole 20,000. Insert everything and there is no room for the previous chapter; the model then knows the style but not where the story is.

Common mistakes

• Believing the AI remembers the work. It does not. It knows only what is in this request.
• Inserting the entire bible every time. It fills the limit and pushes out the preceding text. Only the entries the scene needs.
• Asking for a long answer with a long input. Input and output share the same limit. To generate a whole chapter, shrink the input.