OpenAI 将其提示词建议整合为一份面向日常用户而非开发者的指南。重点围绕四个基础模块、实用护栏以及 Codex 工作流，而非 API 参数或模型调优。
该指南发布前不久，OpenAI 推出了 ChatGPT Work，这是一款基于 Codex 技术和全新 GPT-5.6 模型的独立产品，能够花费数小时处理复杂项目，跨应用和文件操作，并生成完整的 Excel 或 Word 文档。
该指南将常规 ChatGPT 界面与 Codex 纳入同一框架，反映了两者正融合为一款产品。其语气与 OpenAI 近期面向 GPT-5 和 GPT-5.5 的开发者文档有所不同，后者侧重于 API 参数、推理努力级别和复杂的提示词模式。这份面向终端用户的指南去掉了所有技术细节，但保留了相同的核心理念：从小处着手，明确表达需求，仅在必要时添加规则。
提示词包含四个可选基础模块
OpenAI 将提示词结构化为目标、上下文、输出格式和边界。这些均非必填项。该公司表示，简短的提示词通常就能奏效，只有在处理较大任务时才需要填写全部四项。
该指南建议以结果为导向，而非罗列步骤序列。文档写道：“当流程本身至关重要时，才描述流程。否则，给 ChatGPT 留出搜索、比较信息和调整方法的空间。”目标受众或格式对输出的塑造作用远大于详细的指令。
约束优于分步脚本
OpenAI 建议不要为每一步编写脚本，而是设置一两条硬性规则来阻止不期望的行为。例如：“保持已批准的日期和预算数字不变”以及“将消息准备为草稿。不要发送它。”
同样的“少即是多”逻辑也适用于上下文。仅附加那些确实会改变答案的资料来源。该指南列出了电子表格、PDF、图片、网络搜索和共享项目文件等选项，以及 Google Drive、Gmail、Slack 和 GitHub 的插件。对于高风险工作，OpenAI 建议让 ChatGPT 自行验证其输出，例如检查每个行动项是否都有负责人和截止日期。
Chat 处理快速任务，Work 处理繁重工作
这份指南划清了 Chat 与“Work”的界限：Chat 用于快速提问和改写，而“Work”则负责那些需要调用多个来源、进行修改或生成报告等较大交付物的任务。Work 任务会消耗更多积分，但当它们能节省时间或支持重要决策时，就物有所值。对于重复性任务，OpenAI 建议先手动优化提示词，然后再将其自动化。
用户不必在第一次提示时就做到完美。后续追问是优化输出的预期方式。跨会话通用的偏好设置应放在“设置 > 个性化”中的“自定义指令”里。任何特定于任务的内容则保留在提示词中。
Codex 新增引导、排队和沙盒模式
对于编程助手 Codex，OpenAI 引入了两种在运行过程中影响任务的方式。“引导”功能可向当前运行添加一条消息并重新引导其方向。“排队”功能则为下一次运行排列一条消息。在 CLI 中，Enter 和 Tab 键可作为快捷键使用。
Codex 在限制文件和网络访问的沙盒内运行命令。如果任务需要超出这些限制，Codex 会请求批准。两个斜杠命令有助于处理多步骤项目：“/plan”命令让 Codex 在做出更改前先分析代码并提出方案，而“/goal”命令则设定一个更高级别的目标，Codex 会在多个步骤中遵循该目标。对于代码审查，用户可以在本地运行“/review”命令，或在 GitHub 评论中提及“@codex review”，并可选择性地指定审查重点，例如“review for security vulnerabilities”。
OpenAI has consolidated its prompting advice into a single guide written for everyday users, not developers. The focus is on four building blocks, practical guardrails, and Codex workflows rather than API parameters or model tuning.
The guide arrives shortly after OpenAI launched ChatGPT Work, a standalone product built on Codex technology and the new GPT-5.6 model that can spend hours on complex projects, operate across apps and files, and produce finished Excel or Word documents.
It covers both the regular ChatGPT interface and Codex in a single framework, reflecting how the two are converging into one product. The tone is a bit different from OpenAI's recent developer docs for GPT-5 and GPT-5.5, which focused on API parameters, reasoning-effort levels, and elaborate prompt schemas. The end-user guide drops all of that but keeps the same core idea: start small, say what you want, and only add rules where you need them.
Prompts get four optional building blocks
OpenAI structures prompts around a goal, context, output format, and boundaries. None are required. A short prompt often works, and filling in all four only makes sense for bigger tasks, the company says.
The guide recommends leading with the result, not a sequence of steps. "Describe a process when the process itself matters. Otherwise, leave ChatGPT room to search, compare information, and adjust its approach," the document reads. A target audience or format shapes the output far more than detailed instructions.
Constraints beat step-by-step scripts
Rather than scripting every move, OpenAI recommends one or two hard rules to block unwanted behavior. Examples: "Keep the approved dates and budget figures unchanged" and "Prepare the message as a draft. Don't send it."
The same less-is-more logic applies to context. Only attach sources that will actually change the answer. The guide lists spreadsheets, PDFs, images, web search, and shared project files as options, along with plugins for Google Drive, Gmail, Slack, and GitHub. For high-stakes work, OpenAI suggests asking ChatGPT to verify its own output, for instance, checking whether every action item has an owner and a deadline.
Chat handles quick tasks, Work handles heavy lifting
The guide draws a line between Chat for quick questions and rewording, and "Work" for tasks that pull in multiple sources, make changes, or produce larger deliverables like reports. Work tasks burn more credits but pay off when they save time or support important decisions. For recurring tasks, OpenAI suggests refining the prompt manually first, then automating it.
Users don't need to nail the first prompt. Follow-ups are the expected way to refine output. Preferences that carry across sessions belong in "Settings > Personalization" as "Custom Instructions." Anything task-specific stays in the prompt.
Codex adds steering, queuing, and sandbox mode
For the coding assistant Codex, OpenAI introduces two ways to influence tasks mid-run. "Steer" adds a message to the current run and redirects it. "Queue" lines up a message for the next one. In the CLI, Enter and Tab serve as shortcuts.
Codex runs commands inside a sandbox that restricts file and network access. If a task needs to go beyond those limits, Codex asks for approval. Two slash commands help with multi-step projects: "/plan" tells Codex to analyze the code and propose an approach before making changes, while "/goal" sets a higher-level objective Codex follows across multiple steps. For reviews, users can run "/review" locally or mention "@codex review" in a GitHub comment, with an optional focus like "review for security vulnerabilities."