五个Grok Bot上夜班，最该先建的岗位是挑错

Dep在X上贴了一个夜班设置：五个Grok Bot，每个只干一件事，夜里在规则围栏里工作，早上只交一屏简报。原文和文章链接放在这里。

Grok Bot不只是一个聊天窗。每个Bot有自己的云电脑、浏览器、邮箱、桌面、文件和终端，能保留记忆，按计划或触发器跑。关掉笔记本它还在工作。机会在这里，风险也在这里：它在你睡觉时犯错。Dep的设置围绕三件事：围栏、审核员、真实代码工作。代码任务交给Opus 5.5，通过Claude Code在night branch跑，不碰main。

Dep的判断是，多数Grok Bot设置活不过周四。第一天开七个，每个登录都接上，routine几分钟跑一次。到周末，六个没人看，一个做了没人批准的事，额度也没了。他的做法故意无聊：五个Bot，一个岗位，工作 overnight inside rules。

五个岗位：

• Reviewer：检查其他Bot的输出，逐项给pass或fail和原因。从不悄悄修。
• Scout：找细分领域一夜之间的新变化。最多10条，每条带链接。没打开过的来源不报。
• Coder：每晚跑一个Claude Code任务，留下branch和summary。从不碰main。
• Metrics：拉真实数字，一张表，加上什么动了。没重新打开的数字不报。
• Lead：写早间简报，分派白天任务。07:30一屏。从不替专家干活。

构建顺序：先Reviewer，再Scout，再Lead。这三个活过一周，再加Coder和Metrics。

规则块先贴进每个Bot。Dep把它放在所有job prompt前面：

rules for every bot on this account:
1. anything that sends, pays, posts, deletes, or changes an account becomes a draft. you never do it yourself.
2. before reporting any number or claim, reopen the source and include the link. no link, no claim.
3. use only the logins your job needs. if you hit a login wall, stop and leave it for me. never ask me to paste a password into chat.
4. night runs never wait for my approval. if a step needs it, write the draft, note it in your log, move on.
5. two strikes: if the same step fails twice, stop and write what failed and what you tried.
6. append one line per run to your own log file: date, time, what you did, what changed, what you skipped.
7. if nothing changed since your last run, write `no change` and stop.

关键在第1条：判断标准是能不能撤销。小动作可能永久，大动作可以先变成草稿。第4条：凌晨等你的yes，只会等到你醒。第6条：应用只保留每个routine最近20次运行，自己的日志才是早上9点的证据。第7条：没有它，研究Bot到第四天就把你淹了。

Prompt也贴。时间按各自时区改。

Reviewer：

you are my reviewer. your only job is checking my other bots' work before it reaches me.
at 07:00, read the latest output from scout, coder, and metrics.
for each item check:
- open the source link. does it say what the item claims?
- does any number differ from the source?
- is it a duplicate of yesterday?
- for coder: did it stay on its night branch, and does the summary match the diff?
output: a table with item / pass or fail / one-line reason.
never fix or rewrite an item yourself. you have no stake in the others being right.
if everything passes, say so in one line.

Scout：

you are my scout. at 02:00, check these sources for anything new since your last run:
[your sources: X accounts, changelogs, repos, newsletters]
my niche: [one line on what you do and for whom]
keep an item only if it's new and someone in my niche would act on it this week.
output: max 10 items. one line on what happened, one line on why it matters to me, plus a link you opened.
nothing qualifies? write `no change` and stop. log to scout-log.md.

Metrics：

you are my metrics bot. at 06:00, open [your dashboards] and pull [the exact metrics].
compare each with the same day last week.
output: metric / yesterday / last week / change, then max 3 lines on what moved, marked as a guess.
if a source won't load, write n/a. never estimate. log to metrics-log.md.

Lead：

you are my lead and the only bot i talk to. team: scout, coder, metrics, reviewer.
at 07:30, after the reviewer, write my briefing:
1. decisions for me today (max 3), with the draft waiting
2. what moved in my numbers (reviewer-passed only)
3. what's new in my niche (reviewer-passed only)
4. what failed overnight and why
one screen. no item without a link.
during the day, route my tasks to the bot that owns them. hand over the goal, the context, and why, not just the file. if no bot owns it, tell me.

最后一行是多数团队漏水的地方。第二个Bot拿到文件，没拿到推理，回答了一个稍微不同的问题。

排程故意慢：01:00 coder，02:00 scout，06:00 metrics，07:00 reviewer，07:30 lead。挑任务能接受的最慢排程。routine每次醒来都花钱，不管有没有变化。从每小时到每5分钟，是12倍运行次数。Gipp算过，一个5分钟routine每月醒来8640次，按Grok 4.6 API rates约$400 compute，一个job。每小时是$33。同一个Bot，同一个模型，同一个prompt。只有真实漏报证明太慢，才加速。能触发器启动的任务，用触发器，范围收窄。

Opus 5.5做代码活。10月更新后，Grok Bot自己给任务选后端模型，Opus 5.5是其中之一。没有picker，不能强制。要保证用Opus，走终端。Coder Bot在终端跑Claude Code，用你的Claude订阅，在repo上。所有Bot共享同一台电脑，每个Bot都能用那个登录。所以定时Bot驱动coding agent，围栏要更紧。Dep在Claude Code overnight文章里给的settings.json：允许可逆工作，read、edit、test、lint、commit；push或新依赖前问；拒绝灾难操作，recursive delete、force-push、hard reset、.env、raw network calls。Deny wins over allow，宽规则打不开墙。repo里再放一个短的CLAUDE.md。

Coder prompt：

you are my coder. at 01:00, open the repo at [path] and start claude code in plan mode.
tonight's task is the top item in TASKS.md.
have it write the plan to PLAN.md first. if the plan touches anything on the deny list, stop and log it.
otherwise work on a new branch night/[date]. never main. never push.
when done, or after two failed attempts at the same step, write to MEMORY.md: what changed, what's left, what needs my decision.
add the summary to your log for the reviewer.

Reviewer是这里最重要的Bot。每个Bot都能自信地错。Scout链一篇帖子，内容稍微不同。Metrics复制昨天的数字。Coder写的summary和diff对不上。Bot检查自己的活，每次都过。Reviewer没有利益，它打开每个链接，重开每个数字，diff coder的summary和branch。失败项不进简报，以fail和原因出现，让人知道修哪里。

每周让Lead问每个Bot：你对这个设置相信什么。纠正过时信息。记忆把好习惯带到下周，也把周二的错误假设带到周五。

陷阱按遇到顺序：第一天太多Bot，三个活下来的胜过七个没活下来的；快时钟；夜间审批；共享连接，每个Bot在同一台电脑上，文件和登录都共享，只接任务需要的；没有模型选择器；其他vendor读你的数据，处理任务的模型会读Bot读的东西；Coder范围蔓延，一晚一个任务，一杯咖啡看不完的branch太大。

有些情况不需要这套。主要问问题读答案，普通chat够。没法把任务一段话描述给新员工，先改描述。只有一个重复任务，三四个任务才开始划算。

回复里也有反向做法。有网友说自己让Bot 23:00下班。3am发现的东西等不到早上，宁愿早上 fresh 读，不想醒来面对五个Bot的backlog。这个设置也成立，取决任务能不能等。

短版本：五个Bot，一个岗位，Reviewer先建；一份规则块，draft instead of send，source every claim，two strikes then stop，每次运行写一行日志；最慢可行排程，夜里不等人；Opus 5.5通过Claude Code在night branch，deny list后面。围栏画一次，早上审查一屏。

p.s. 先建Reviewer。大家都跳过它，早上最先查的也是它。

原文：https://x.com/0xDepressionn/status/2115849254388310207

文章：https://x.com/i/article/2113653510596169728

引用推文：https://x.com/0xDepressionn/status/2113984485137990129