At 3 A.M., AI Guesses and Facts Look the Same

Dep posted a Grok Bot night-shift setup on X: five bots, each with one job. The most counterintuitive part is that he recommends building the reviewer first—a bot that produces no content and is only responsible for finding mistakes. The other four write things, find leads, run code, and pull data; the reviewer at 07:00 opens every link, reopens every number, and compares code branches and summaries. Any link that can't be opened is an automatic failure.

Grok Bot gives each bot a cloud computer, browser, files, and a terminal; it keeps running after you close your laptop and can wake up on a schedule or trigger. Opportunity and risk come from the same source: it works while you sleep, and it also makes mistakes while you sleep. Dep notes that most people spin up seven bots on day one, connect all their logins, and have routines wake every few minutes. By Thursday, six are ignored, one did something no one approved, and the quota is gone.

Dep's original post

The five bots have very narrow roles. Reviewer checks the other bots' output and returns pass or fail plus a reason; it does not secretly fix things. Scout finds changes in your field over the past night, at most 10 items, each with a link; anything it didn't open is not reported. Coder runs one Claude Code task each night, leaves a branch and a summary, and does not touch the main branch. Metrics pulls real numbers and gives a table plus an explanation of changes; numbers that haven't been reopened are not reported. Lead writes the 07:30 briefing, limited to one screen, and does not do the expert bots' work. The build order is reviewer, scout, lead; after it survives a week, add coder and metrics.

Paste the rules block into each bot first. Only reversible actions count as safe; small actions can also be permanent, and big actions can be left as drafts only. If a request is waiting for a yes at 3 a.m., let it wait; the night shift must finish while you're away. The app keeps only the last 20 runs; your own logs are the evidence at 9 a.m. Research bots have no cap and can drown you by day four. A commenter pointed out that 'can't open it = failure' is the key default; otherwise 'can't open it' quietly turns into 'looks fine,' and bad data gets into the 7 a.m. briefing.

Scheduling is deliberately slow. 01:00 coder, 02:00 scout, 06:00 metrics, 07:00 reviewer, 07:30 lead. Gipp calculated that a 5-minute routine wakes 8,640 times a month, about $400 at Grok 4.6 API pricing; hourly is $33. Same bot, same model, same prompt—only the clock changes. Start slow; speed up only if you're actually missing things. Use triggers instead of clocks when you can.

For code tasks, after the October update Grok Bot automatically selects the backend model, and Opus 5.5 is on the list, but there is no picker and you can't force it. If you want Opus, go into the terminal and run Claude Code using your own Claude subscription, and open a night branch in the code repo. Allow read, edit, test, lint, commit; push and new dependencies must ask; deny recursive delete, force-push, hard reset, .env, raw network calls. Deny rules take precedence; broad allow rules can't build a wall. All bots share that computer—logins, files, and logs are shared—so connect only the accounts the task needs.

Pitfalls in the order you'll hit them: too many bots on day one—three surviving is better than seven dying; clock too fast—a 5-minute routine burns quota even on a night when nothing happens; nighttime approvals—something waiting for your yes at 3 a.m. waits until you wake up; shared connections—every bot can use the same set of logins; no model picker—you can't control when Opus 5.5 enters Grok Bot; other vendors will read your data, and the model handling the task will read what the bot reads; coder scope creep—one task per night, and if the branch is too big to review by breakfast, it's too big.

If you're mostly asking questions and reading answers, ordinary chat is enough. If you can't describe a task clearly to a new employee, fix the description first. If you have only one recurring task, this setup isn't worth it yet; start at three or four. In the comments, someone did the opposite and made the bot clock out at 23:00. The reasoning holds: what it finds at 3 a.m. can wait until morning, when your head is clear. Dep's reply: at 3 a.m., guesses and facts look alike; he'd rather wake up to 10 verified items than 30 maybes.

The conclusion of this setup is short: five bots, one job per role, build the reviewer first, schedule slowly, and hard-code the guardrails. In the morning, look only at one screen of failures and passes. First make the bot checkable, then let it do more work.