Anthropic engineer: "You're not supposed to watch Claude Code work. You're supposed to wake up and review what it shipped." In 22 minutes she builds the entire workflow live on camera. Most people close their terminal and everything stops. This setup keeps shipping while you sleep. Watch the video,... then save the exact setup below👇show more

original link:

If your not reviewing AI code like it’s your junior dev you are doing it wrong

claude code gets much more useful when it keeps running without supervision

🙏🏼

You come back the next day to find Claude 90 seconds into its run, stuck asking for permission to read a file it wrote in the last session. Why do you watch Claude? It is sometimes a dumbass and needs to ask questions.

Yeah but trouble is that if you it alone it will be waiting for your permission to create a temp directory in .Claude folder the whole night.

close the terminal and everything stops is the failure mode nobody talks about. this setup fixes the overnight gap.

yessir

the "review what it shipped" workflow only works if your test suite actually catches the bad stuff. without that its just "wake up and debug what broke at 3am"

Yeah but it keeps asking me if it wants to proceed, or for permissions I’ve given 30x before If I don’t watch, I come back 5 hours later to find out it stopped after eleven minutes

That's wild. Honestly feels like we're moving from coding to just managing AI workers. Not sure if I'm ready to wake up to code I didn't write.

review what it shipped sounds easy until you realize reviewing means remembering what you asked for. the work didnt vanish, it moved from writing the code to holding the intent

"If the spec has OPEN QUESTIONS, stop and show them to me." How’s that going to work if I’ll be sleeping?

half the agent discourse is just people rediscovering background jobs

This will produce complete garbage code. The human needs to write the architecture/specs. Still, even rock solid specs will veer heavily if not batched and supervised.

watch the terminal or wake up to a shipped feature, two different workflows

Two ways same Claude

I need a prompt that just keeps clicking yes.

Banger

Of course an Anthropic engineer would say that is there title Marketing Engineer? 🤣🤣

Bullshit.yesterday it took few hours in fixing something it itself caused (regression). And regarding what it did when you wake up,it's normally stuck at some or the other permissions,even for those items which I have previously "Always Allow" ed. It just keeps running endlessly

Anthropic engineer tells you to use their product more? Yah ok! tell me any business person who will tell u to not use their product

@zodchiii how do you even track what it's doing while you're asleep? sounds like magic but i feel like i'd wake up to chaos 😂

Oh god these ai bros hugging anything anthrpic or openai say. Just to get some stupid impressions. Stay with cryptos guys, please.

@michaeltirvin This is impressive, and valuable in some spaces. The greater value is in collaboration, using the strengths of both human and AI. That is where we achieve what neither can get to alone.

And occasionally (not) Claude goes off the rails, blows past guardrails. Develops agents that build or supervise incompletely and there is no warning that the 36 agents you're using have an observability gap.

Imagine what it would be in 5 years... Amazing

Yeah if you turn off all the security checks .. and do all your planning up front.. a lot go risk there ..

"You don't watch Claude Code work. You just wake up and let the API bill jump-scare you."

If cloud just runs while you sleep implementing and bug fixing all of your apps, that would be so useful for so many people. I could technically buy 2 Claude Max subs, one for debugging at night and one for implementing during the day while I'm awake. That would be perfect.

22 mins of solid content to check out before sleep, appreciate it bro

They’ll do anything but teach people hot fricking easy is to keep your laptop running with the lid closed. Which is super easy, but they don’t make money out of it

It’s not quite there yet, 15-20 min unattended with a solid spec is good, 8 hours? You’d spend the next 16 making it coherent

wake up to either you have hit your weekly limit or an invoice for api usage what an experience

finally an ai workflow that doesn't just hallucinate and crash actually shipping while you sleep is the only metric that matters

17 scheduled agents ship across five domains while I sleep here. The catch nobody mentions is the token bill, because every loop from cold context runs about 4x a warm one. I warm them once a day and read the diffs at breakfast. What does she run the wake loop on?

The biggest mindset shift is treating AI like a coworker with tasks, not a chatbot with questions.

Token seller provides system for using more tokens 🥱