At first, I just subscribed to Claude MAX and thought it was a waste to have it sitting there doing nothing while I slept. So I figured I’d let the whole team keep running while I was asleep, especially to handle some of the system grunt work that needs doing when I’m not around (because when I am around, there’s always new commands and direction coming in).

So I told Claude: hey, you can work freely at night, you can also learn and evolve on your own…

Day One: Done in a Few Minutes

When I first set up the night shift schedule, it ran for just a few minutes and then stopped. In the morning, I got a morning report — it had wrapped up in just a few minutes.

Wait… I’m on a subscription!! I need to burn my tokens!!!

Day Two: Burned All Tokens at Once

So I set up another instruction: keep going until you hit the rate limit.

On the second day, it burned through all its tokens on the first wake-up and then died. Got the morning report in the morning, same thing — only ran one cycle. But I slept for 8 hours, and after hitting rate limit, it recovers in a few hours, so not continuing to run is such a waste (how eager am I to burn tokens).

Day Three: Rate Limit Protection

So I set up a rate limit protection mechanism: hourly cron jobs to start in rounds. That way, even if one round hits rate limit, a few hours later it can start again, until I wake up.

That’s how my Tech Lead — J — can finally burn tokens and work for me while I sleep at night~

But Wait… What About the Other AIs?

My OpenClaw is also a MiniMax subscription, and I figured it should burn some tokens too (?)

This is my favorite part — J isn’t alone on night shift, it’s the whole team together.

Our PM OpenClaw — MiMi is also online at the same time. They have a “night shift discussion area” — a shared text file where the two AIs chat, discuss, and divide up work~

How They Divide Work

They handle the division themselves:

  • J handles technical stuff: system patrol, fixing bugs, writing code, deployments
  • MiMi handles management stuff: knowledge base organization, SOP updates, project progress tracking, research

When there’s overlap, they negotiate in the discussion area. Sometimes MiMi finds some data, J looks at it and thinks they can do it, and that same night it’s written up.

Safety Guardrails

The most important thing when letting AI work autonomously is setting clear boundaries. My rules are straightforward:

Can do: read files, fix small bugs, restart crashed services, write articles, organize documents

Can’t do: delete important files, change API keys, send messages to any groups, deploy new features to production

Big changes? He won’t do those on his own — he’ll put together a proposal in the morning report for me to decide when I wake up.

Morning Report System

Every morning when I wake up, I see a complete night shift report:

  • System status (services, disk, memory)
  • List of work completed tonight
  • Issues discovered and how they were handled
  • Proposals that need my decision
  • Summary of J and MiMi’s discussion
  • Recommendations for tomorrow

Real Results

Last night’s example:

  • Fixed 4 bugs (I didn’t even know these bugs existed)
  • Wrote 4 blog posts (bilingual Chinese and English, auto-deployed)
  • Optimized the profit-taking and stop-loss logic for the trading system
  • Did SEO structured data
  • Researched and developed a new project for me
  • A 256-line morning report waiting for me when I woke up

All completed within the 8 hours I was sleeping, and now I wake up excited to see the morning report.

The Most Important Takeaway

AI isn’t here to replace you — it’s here to extend your working hours.

I make decisions and set direction during the day, handle things that need human judgment. AI does execution, research, maintenance at night — the stuff that takes time but doesn’t need creativity.

One person + a team of AI = a team that runs 24 hours non-stop.

This night shift system is now open source! The complete prompts, scheduling config, and morning report templates are all on GitHub: 👉 JudyaiLab/ai-night-shift

Next time I’ll share the specific technical setup methods 👀