queue a long AI job, then walk away

Lenny's AI Builders - LAB 0014

the prompt i use so an agent keeps working for hours, without me there to babysit it.

i think the AI companies built a casino.

so remember this: queue it, then walk away.

why it matters

long jobs stall when the agent stops to ask you things.

you come back hours later and it's been waiting on one question the whole time.

a few plain lines in your prompt fix that. the agent keeps going, checks its own work, and leaves you one page to review.

what you get

one prompt for any job that takes hours.

• no babysitting. it doesn't stop to ask you things.
• a clear finish line. you tell it what done looks like.
• one review page. what it made, what it couldn't finish, what it needs from you.

the real example: a 4-hour run

BuilderBench is my own benchmark. the model runs on extra high, at standard speed, in a four-hour window.

in this episode i put GPT-6.1 Sol through it. it scored 41.75 out of 100. the recorded task cost was $14.71 - an API-equivalent estimate in US dollars.

my take on the landing page it built: good enough, but it doesn't really have any taste. Opus 5.5 is still the winner for me.

try it

1. pick one big job you'd normally sit and watch.
2. write a short checklist of what done looks like.
3. paste the prompt in the next section and add your checklist.
4. start it, then walk away.
5. come back and read the review page first.

copy this prompt

this is the prompt i read out at 12:43. swap in your own checklist.

This job will take hours. Work until it's done.
Here is what done looks like: [insert your checklist]
Don't stop to ask me things.
If you're stuck, write down why and move to the next part.
Check your own work against the checklist before you finish.
When you finish, build me one review page: what you made, what you couldn't finish, what you need from me.
I'll be back in four hours.

you'll know it worked when

you come back to one review page, not a question.

it shows what it made, what it couldn't finish and what it needs from you.

each item is checked against your checklist.

a bad result: it stopped early to ask you something, it skipped the checklist, or it says everything is done with nothing to show.

limits and dates

the key lines are "don't stop to ask me things" and "i'll be back in four hours". in my experience they work especially well with Claude. that's my experience, not a test.

BuilderBench scores are my own ratings of real work. GPT-6 Sol (45.5) and Opus 5.5 (68.3) were scored on older versions of the bench - a reference, not a ranking.

the run was set up on 30 September 2026. the scores were finalised on 6 October 2026.

LAB 0014 went public on YouTube on 7 October 2026 at 10:56 am Perth time.

resets: the effect is real, the intent is unknown. the casino take is my opinion.

the rest of this LAB

• the AI casino: Codex resets. OpenAI keeps refilling everyone's Codex usage at random. one tracker has counted 58 resets. people bet real money on Polymarket on how many come each week. the effect is real, the intent is unknown.
• the dark psychology behind it. Skinner's pigeons from the book Hooked, why the wait is the hook, and why losing hurts more than winning feels good. my take: it's probably not that deep.
• Claude Code vs Codex. David Ondrej says Claude Code gives you 16-24x the usage of Codex. his maths, not checked by anyone.
• the Pope vs AI art. Pope Leo XIV says it's urgent to tell human art apart from what machines produce. my take: AI is a tool we create with.
• Apple vs every AI agent. Apple plans new Full Disk Access controls for Mac developers.
• Hinton on the intelligence explosion. Geoffrey Hinton says the intelligence explosion may be close.
• Theo vs BridgeMind. a public fight over how NerfBench measures models.

sources

• Polymarket: how many Codex resets, week of Sep 28
• a Codex reset tracker (willreset.com)
• David Ondrej: Claude Code vs Codex usage
• Pope Leo XIV on AI art
• MacStories: Apple's new Full Disk Access controls
• Geoffrey Hinton on the intelligence explosion
• Theo on NerfBench
• BridgeMind: how NerfBench works

next step

watch LAB 0014 on YouTube - the practical starts at 10:32.

want the next one? get the next field notes. optional - these notes stay free to read.