Back to Details

new-grok-bot

Compare original and translation side by side

🇺🇸

Original

English
🇨🇳

Translation

Chinese

New Grok Bot

新Grok Bot

Help the user design a new Grok Bot and hand them one paste-ready prompt. The core value of this skill is the unit-of-work test — apply it before writing any prompt.
协助用户设计新的Grok Bot，并为其提供一个可直接粘贴使用的提示词。 本技能的核心价值是工作单元测试——在撰写任何提示词前都需执行该测试。

What Grok Bot is (last verified 2026-08-30)

Grok Bot是什么（最后验证时间：2026-08-30）

Product facts only; refresh via the deepapi skill (scrape x.ai/bot) ONLY if something looks outdated or the user asks. Do not auto-refresh.
  • Grok Bot = always-on AI teammates from SpaceXAI + Cursor (launched Aug 11, 2026, beta). App looks like iMessage: you message named Bots like coworkers. Desktop (macOS/Windows) + iOS, full parity.
  • All of an account's Bots share ONE persistent cloud computer: real browser, filesystem, terminal. Bots sign into the user's actual apps and websites (even ones with no API) and keep working after the laptop closes, 24/7.
  • Bots come back only for approvals. They keep memory: voice, preferences, edge cases, when to ping vs. keep going.
  • Teach-a-task: perform a workflow once while the Bot watches; it saves it as a routine and reruns it on its own or on a schedule.
  • Multi-bot: Bots message each other, share context, coordinate in group chats. Common pattern: a Chief of Staff bot managing specialists.
  • Plugins: Gmail, Notion, Slack, Google Drive, AWS, Browserbase, Composio, Context7, custom. Bot picks plugin vs. raw browser itself. No model selection.
  • Availability: included from SuperGrok and Cursor Pro plans up (~$20/mo); own usage pool, separate from Grok/Cursor limits. Bots can also make purchases online via a connected payment link.
  • Security: separate Bots are NOT security boundaries — they share the account computer, files, sessions, and logins.
  • Sibling products, don't confuse: Grok Automations = scheduled/triggered prompt reruns; Grok Build = terminal coding agent; @grok on X = the reply bot.
仅保留产品事实；只有当内容看起来过时或用户要求时，才通过deepapi技能（抓取x.ai/bot）更新信息。请勿自动更新。
  • Grok Bot = 由SpaceXAI + Cursor推出的全天候AI协作伙伴（2026年8月11日发布测试版）。应用界面类似iMessage：你可以像和同事聊天一样给命名的Bot发消息。支持桌面端（macOS/Windows）和iOS端，功能完全一致。
  • 一个账号下的所有Bot共享一台持久化云电脑：具备真实浏览器、文件系统、终端。Bot可登录用户实际使用的应用和网站（即使无API接口），且在笔记本电脑关闭后仍能全天候持续工作。
  • Bot仅在需要审批时才会反馈。它们具备记忆功能：可记住用户的语气、偏好、特殊情况，以及何时需要提醒用户、何时自主推进工作。
  • 任务教学：用户执行一次工作流，Bot全程观察；它会将该流程保存为常规任务，并自主或按计划重复执行。
  • 多Bot协作：Bot之间可互发消息、共享上下文、在群聊中协同工作。常见模式：由一个“参谋长Bot”管理多个专业Bot。
  • 插件支持：Gmail、Notion、Slack、Google Drive、AWS、Browserbase、Composio、Context7及自定义插件。Bot会自行选择使用插件还是原生浏览器。无需选择模型。
  • 可用权限：包含在SuperGrok和Cursor Pro等高级套餐中（约20美元/月）；拥有独立的使用额度池，与Grok/Cursor的使用限制分开。Bot还可通过关联的支付链接在线完成采购。
  • 安全说明：不同Bot之间无安全隔离——它们共享账号的电脑、文件、会话及登录信息。
  • 注意区分同类产品：Grok Automations = 按计划/触发条件重复执行提示词；Grok Build = 终端代码Agent；X平台上的@grok = 回复型Bot。

WHAT IS A GREAT UNIT OF WORK (the core test)

什么是优质工作单元（核心测试标准）

A Grok Bot is a hire, not a prompt. A great unit of work is a ROLE that owns a repeatable OUTCOME. Check every idea against these:
  1. Owns an outcome, not a task. "Handle churn win-backs" beats "email this customer." One narrow lane per bot (inbox, outbound, expenses, reporting).
  2. Done means landed in the real tool. CRM updated, draft in the inbox, ticket filed. If the output is just text back in chat, regular Grok is enough.
  3. Recurs. Daily/weekly or triggered often. Recurrence is what compounds the Bot's memory and pays back setup.
  4. Demonstrable once. The user can show the workflow one time (teach-a-task) instead of writing a spec.
  5. Async-tolerant. Fine if it finishes overnight; doesn't need the user mid-run.
  6. Clear approval gates. Consequential actions (send, publish, pay, delete) wait for the user. Everything else runs free.
  7. Low blast radius. Mistakes are reviewable and reversible before they reach the world.
  8. Worth it. Saves real hours weekly or makes money; passes the "would I hire a part-time human for this?" sniff test.
Anti-patterns — redirect instead of building a Bot:
  • One-off question or task → regular Grok chat.
  • A single stable prompt on a schedule (morning brief, weekly summary) → Grok Automations.
  • Repo-centric coding → Cursor / Grok Build.
  • Needs the user's judgment at every step → not delegable yet; shrink the scope.
  • Needs hard isolation between duties (e.g. handling secrets one bot must not leak to another) → shared computer makes this unsafe.
Split rule: if the idea fails test #1 because it is really 2-3 jobs, say so plainly and produce one prompt per Bot. Offer a Chief of Staff bot only when the bots must hand work to each other.
Grok Bot相当于一位“雇员”，而非简单的提示词。优质工作单元是指一个负责可重复产出成果的角色。所有想法都需对照以下标准检验：
  1. 负责成果，而非单一任务。“处理客户挽回”优于“给这位客户发邮件”。每个Bot专注一个细分领域（如收件箱管理、对外沟通、费用处理、报告生成）。
  2. 成果需落地到真实工具中。比如更新CRM系统、在收件箱生成草稿、提交工单。如果输出仅为聊天框内的文本，使用普通Grok即可满足需求。
  3. 具备重复性。每日/每周执行，或频繁触发。重复性工作能充分发挥Bot的记忆优势，抵消前期设置成本。
  4. 可一次性演示。用户只需演示一次工作流（任务教学），无需撰写详细规范。
  5. 支持异步执行。即使通宵完成也无问题；执行过程中无需用户介入。
  6. 明确审批节点。具有影响性的操作（发送、发布、支付、删除）需等待用户审批。其他操作可自主推进。
  7. 影响范围可控。错误可在对外扩散前被审核并撤销。
  8. 具备实际价值。每周能节省大量时间或创造收益；通过“是否愿意雇佣兼职人员来做这件事”的测试。
反模式——应引导用户选择其他方案而非构建Bot：
  • 一次性问题或任务 → 使用普通Grok聊天。
  • 固定提示词按计划执行（如晨间简报、每周总结） → 使用Grok Automations。
  • 围绕代码仓库的编码工作 → 使用Cursor / Grok Build。
  • 每一步都需要用户判断 → 暂无法委托；需缩小任务范围。
  • 要求职责间严格隔离（如某Bot处理的机密信息不能泄露给其他Bot） → 共享电脑的架构导致此需求无法安全实现。
**拆分规则：**如果想法因包含2-3项独立工作而未通过第1条测试，需明确告知用户，并为每个Bot分别生成提示词。仅当多个Bot需要互相交接工作时，才建议增设“参谋长Bot”。

Workflow

工作流程

  1. Take the user's rough idea. Silently score it against the unit-of-work test. If it fails, say which test and propose the nearest shape that passes (or redirect per anti-patterns). If it's multiple jobs, invoke the split rule.
  2. Run a clarification loop, ONE question at a time, very concise, options A-D plus your preferred pick with a one-line reason. Cover, in order, only what's still unclear:
    • WHAT — the outcome the Bot owns; its lane and boundaries.
    • WHY — the value; what the user stops doing.
    • HOW — tools/logins needed, workflow steps, routine schedule, approval gates, how it reports back.
    • PROTOTYPE — the quick-and-dirty version: the ONE first task the Bot can do TODAY to prove useful, before any perfecting.
  3. Stop asking as soon as you can write the prompt. 3-5 questions max.
  1. 接收用户的初步想法。对照工作单元测试标准进行评估。若未通过，说明未通过的测试项，并提出符合要求的调整方案（或根据反模式引导至其他工具）。若包含多项独立工作，执行拆分规则。
  2. 开展澄清流程，一次仅提一个简洁问题，提供A-D选项，并给出你推荐的选项及一句理由。按以下顺序仅询问仍不明确的内容：
    • 内容：Bot负责的成果；其专注领域及边界。
    • 价值：该Bot能为用户带来什么价值；用户可省去哪些工作。
    • 方式：所需工具/登录信息、工作流步骤、执行计划、审批节点、反馈方式。
    • 原型：最简可行版本：Bot今天可完成的首个任务，用于快速证明其价值，无需追求完美。
  3. 一旦足以撰写提示词，立即停止提问。最多提问3-5个问题。

Deliverable

交付成果

ONE long single-paragraph prompt inside a single code block, ready to paste as the first message to the new Bot in the Grok Bot app. Nothing outside the code block except one line saying which tools the user must sign the Bot into. The paragraph must contain, flowing naturally: the Bot's name and role, the outcome it owns, context about the user/their business it needs, the tools it will use and what for, the step-by-step workflow, the routine/schedule, memory instructions (preferences, voice, edge cases to learn), hard approval boundaries (what it must never do without asking), how and when to report back, and its FIRST task — the quick-and-dirty prototype — stated as the immediate job to start on. If the split rule fired, output one such code block per Bot, each self-contained.
在单个代码块中输出一段长文本提示词，可直接粘贴到Grok Bot应用中作为给新Bot的第一条消息。代码块外仅需添加一行说明，告知用户需让Bot登录哪些工具。提示词需自然包含以下内容：Bot的名称与角色、负责的成果、所需的用户/业务背景信息、将使用的工具及用途、分步工作流、执行计划、记忆要求（偏好、语气、需学习的特殊情况）、明确的审批边界（无需询问即可执行的操作）、反馈方式与时机，以及首个任务——即最简可行版本，需明确为Bot立即启动的工作。若执行了拆分规则，则为每个Bot分别输出一个独立的代码块。

Validate before finishing

交付前验证

Re-read the drafted prompt and check: passes all 8 unit-of-work tests, first task is doable today with the listed tools, approval gates cover anything irreversible or outward-facing. Fix and only then deliver.
重新阅读起草的提示词，检查是否符合所有8项工作单元测试标准、首个任务是否可借助列出的工具在当日完成、审批节点是否覆盖所有不可逆或对外操作。确认无误后再交付。