At 11 p.m., Xiao Lin is revising his resume to switch to a career as a large language model product manager, using a chat tool to refine his work experience against target position and recruitment requirements. After several successful rounds, the tool rewrote deleted experience and forgot his repeated instruction: "Do not package me as a technical R&D personnel." This occurs because he consumed a large number of Tokens, the unit large language models use to process text. A Token can be a Chinese character, part of a word, or a punctuation mark, number or space, and different models split text differently, so Token is not equal to word count. Input questions, uploaded resumes, job descriptions, chat history, and model responses all consume Tokens. The context window, measured in Token, is like a truck with limited capacity; as conversation lengthens, the model may compress or discard older information, gradually ignoring earlier requirements. If Xiao Lin initially said "highlight operational experience" but later discussed products, technology and data, the model may only remember the latest information, causing the resume to drift off track. Similar scenarios occur daily, such as sending fifty-nine WeChat messages in a row with a sentence "Don't bring your computer tomorrow" in the middle, which may be forgotten.