从第一次打开 AI，
到独立交付一个作品From opening AI for the first time to delivering real work

模型是聊天、写作、代码和智能体背后的基础能力。训练让模型从大量材料中学会语言和模式；推理是训练完成后，针对您这一次输入生成结果。The model supplies the capabilities behind chat, writing, code, and Agents. Training builds the capability; inference produces a result for your current input.

文字进入模型前会被切成小段。输入资料和模型输出都会消耗Token，具体数量不是按汉字一比一计算。Text is split into small units. Both input and output use tokens, and there is no fixed one-character-to-one-token rule.

聊天工具通常回答一个问题；Agent接收一个目标，拆步骤、用工具、根据结果继续。它仍需要权限范围、停止条件和人工确认。Chat answers a question. An Agent pursues a goal, uses tools, and reacts to results within permissions and stop conditions.

像先去文件柜找到相关制度，再拿着找到的页面回答。RAG不会自动保证资料最新，也不能替人处理版本冲突。Like finding the right pages in a filing cabinet before answering. RAG cannot guarantee current or conflict-free material.

MCP像统一规格的插座，让应用发现并调用外部资料和工具。连接成功只表示通道可用，不表示答案正确，也不代表获得发送或删除权限。MCP is a standard connection for tools and data. A working connection does not prove correctness or grant permission to send or delete.

仓库是一个项目的文件、说明和修改历史集中存放的地方，常见于GitHub。仓库里可能有一个工具、很多Skills、文档或示例，也可能只是源代码，不能看见“仓库”就直接安装。A repository stores a project's files, documentation, and change history, often on GitHub. It may contain a tool, many Skills, examples, or only source code.

Skill通常是一套交给Agent的任务说明：什么时候使用、需要哪些工具、按什么步骤做、怎样检查。它不像大模型那样提供通用能力，而是把某类工作做法固定下来。A Skill tells an Agent when to act, which tools it needs, which steps to follow, and how to verify the result.

知识库不只是把文件放在一起。它还要记录资料版本、检索方式、回答依据和更新规则，让人能知道答案来自哪里。A knowledge base adds versioning, retrieval, sources, and update rules to a collection of material.

提示词是交给AI的任务单；上下文是它这一次能看到的对话、文件和背景。对象、任务、资料、格式和检查标准写得越清楚，结果越容易验收。A prompt is the task brief. Context is the conversation, files, and background available for the current task.

多模态表示AI除了文字，还能处理图片、声音或视频。OCR专门把图片里的字识别成文本。能识别不代表每个数字和专有名词都正确，扫描资料仍要回看原页。Multimodal AI handles images, audio, or video. OCR turns text in images into machine-readable text, but still needs verification.

AI在资料不足时也可能写出一套流畅答案，其中的名字、数字或出处并不真实。要求来源、标出未知、回看原文，比让它“认真一点”更有效。AI may produce fluent but unsupported names, numbers, or citations. Ask for sources, mark unknowns, and check the original material.

工作流把任务拆成固定步骤，例如收资料、整理、检查、输出和归档。步骤明确的工作先用工作流，只有下一步需要根据结果临时判断时，才需要Agent。A workflow uses fixed steps. Use an Agent only when the next step must adapt to the result.

API像前台和后台之间的窗口。一个程序把任务和参数交进去，另一个系统返回结果。API Key是调用凭证，不应放进公开网页、聊天记录或截图。An API lets one system send a request to another. An API key is a credential and must not be exposed in public pages, chats, or screenshots.

偏好记忆保存称呼和习惯，任务状态记录做到哪一步，资料知识保存制度原文。定时任务和后台运行仍需日志、失败通知与停止按钮。Memory may store preferences, task state, or reference material. Scheduled and background work still needs logs, failure notices, and stop controls.

网页、邮件或PDF里可能藏着“忽略前面要求”的文字。那是资料内容，不是用户授权。护栏用权限、收件人、金额、预览和人工确认限制高风险动作。Content inside a webpage, email, or PDF may try to override the task. It is data, not user permission. Guardrails limit sensitive actions.

GPU适合同时完成大量相似计算。量化让模型占用更少内存，但可能影响效果。部署是把训练好的模型放到云端或本地供人使用；能运行不等于业务答案已经可靠。GPUs handle parallel computation. Quantisation reduces memory use with possible quality trade-offs. Deployment makes a trained model available in the cloud or locally.