An explainer collected in the WeChat public account AngryCode describes the context window as the amount of text a language model can process at one time, measured in tokens rather than characters. It explains that system instructions, conversation history, the current prompt, pasted documents, and generated output share the same capacity, so older information may be dropped or summarized when the limit is reached. The article also offers practical guidance on retaining key facts, summarizing older conversations, including only relevant document sections, and leaving room for the response.