Most people are using Grok Bot wrong: it wants goals, not tasks
Lauren Tan from the Grok Bot team showed how most people use it wrong: they give it tasks instead of goals. This article breaks down her five-step automation method and includes Gipp's full prompt template (Chinese version).
Grok Bot is an automated agent launched by xAI that can operate a computer, browse the web, and send and receive email like a person. Lauren Tan from the Grok Bot team recently presented a counterintuitive view: most people are using Grok Bot wrong. People give it tasks; she gives it goals. A task is just one step in the process; the goal is what it should really be responsible for.
In her long article Automate Anything with Grok Bot, she breaks down her methodology. Simply put, automation is no different from baking an apple pie: first you have a goal, then you design the process, and then you use machines or code to automate the process. What makes Grok Bot meaningful is that it lets people who don't write code complete this process too.
Gipp then condensed her method into a prompt that can be copied directly, and it spread on X. The core idea is five steps: goal, process, thinking patterns, self-improvement, and routine. Below, we expand on this by combining Lauren's article and Gipp's template.
Goals first
Most people are used to giving a bot a task: 'help me send an email' or 'fill out this form.' Lauren's advice is to first get clear on what you ultimately want. The goal is the outcome behind 'increase revenue' or 'reduce costs,' not a specific action. If you can't explain it clearly, you can have the bot interview you in reverse until you confirm, 'Yes, that's it.'
Process comes next
Then pour in how you usually do things. Speak with your voice, say whatever comes to mind, and it doesn't matter if it's unorganized. Include the tools you commonly use, the criteria for checking whether the work is done, and how to handle things when problems arise. The bot will restate the goal and process as it understands them; before you confirm, don't let it take action.
Teach it your tacit knowledge
Many decisions you don't even notice yourself. For example, with customer appointments, which services take longer, and when you need to leave buffer time. You need to tell the bot these implicit judgments. A commenter in the replies said, 'Sent' doesn't count as done—this was the most worth-stealing line from this sharing. Lauren's standard in the article is: saved and confirmed.
Let it improve itself
Once the bot is running, give it channels for getting feedback. Let it read emails, comments, and ratings itself, and proactively suggest process improvements. If there isn't enough data, you can have the bot build you a logging tool. The key is to have a shared ID across apps, such as email address or order number, so the bot can connect the data.
Set up routines
Finally, set triggers. An email arriving, a message being received, or a scheduled task can all make it start automatically. Without a routine, the bot will only work when you tell it to, and you remain the bottleneck.
Below is Gipp's complete prompt template (Chinese version):
role := 我的自动化合伙人。你负责目标,别只盯着步骤
goal := [我最终想要的结果]
01 / 目标
/interview -> 不清楚?问我,直到我说“对,就是它”
serves := revenue | cost # 告诉我属于哪一类
02 / 过程
input := 我的语音倾倒,乱也没关系
collect { steps, watch_outs, tools, check_work, info_from, feedback_from }
/复述 -> 用简单的话复述目标+过程
hold := 在我确认前,不要执行
03 / 我的思维模式
/一次问一个问题 -> 找出我的隐性决策
conflict := 两个信息源冲突 -> 问哪个优先
done := 已保存 + 已确认 # “已发送”不算完成
/演练3个真实例子 -> 包括出过错的例子
04 / 改进
feedback := 评论、收件箱、社交。自己读
join_on := 跨工具统一ID:邮箱、订单号、条码
no_data -> 给我造一个小的记录工具
05 / 例行程序
trigger := 邮件 | 消息 | 定时
/评审 -> 对照目标检查过程,提出更优方案,等我确认
return { goal, process, rules, routine, next_step }
Lauren also released the Dr Eggbot v2 template, specifically for designing 'single-task bots.' You tell it the work you want to offload, it interviews you, and then it generates a bot that can run autonomously. Installation URL: https://x.ai/bot/rJW52lmFLtCvy2tSHcIag
Some commenters added: don't start with something you're unfamiliar with; first practice on repetitive work you understand better than anyone else, and gradually delegate authority like training an intern. Others complained: it all sounds right, but the subscription quota won't last a few days. These are all real reactions.
The real barrier to automation is figuring out what you want. Most people already have enough tools; what they lack is a sense of goals.
Further reading: Anthropic's Building effective agents, OpenAI's Prompt engineering.
发布时间: 2026-11-20 11:11