I Stopped Prompting AI One Task At A Time. This Works Better.

Study Guide

Overview

In this video, Nate B. Jones (AI News & Strategy Daily) argues that the way most people use AI, one task at a time, is the wrong frame for the work that actually weighs on us. Instead of prompting AI repeatedly and remaining the manager of every step, he proposes a structure he calls a loop of loops: recurring jobs that remember context, notice when something changes, hand off to each other, and stop where your judgment is required.

The promise of the video is conceptual, not a tool tutorial. Nate wants you to leave able to look at one repeated job in your own life and say, "that's a loop I could build." A full build walkthrough is promised for Friday.

The Core Vocabulary

Nate keeps the definitions deliberately simple:

  • A prompt is one request. It gets you a single answer.
  • A loop is one recurring job with memory. It saves you from running the same answer again and again.
  • A loop of loops is what happens when recurring jobs can notice each other, share what changed, and stop when they hit your boundaries. It saves you from becoming the full-time manager of all your automations.

Crucially, this is not a "nanny in a box" or an agent pretending to run your life. It is a set of small, remembered workflows that notice what changed, hand off context, stop in the right places, and bring you in when your judgment matters. As Nate puts it: agents are loop managers, and that is what makes them interesting.

From Prompt to Loop: The School Trip

"Make me a packing list for the school trip" is a fine prompt. The AI can suggest a raincoat, lunch, water bottle, sunscreen. But the prompt was never the whole job. You still had to:

  • Know the trip was happening at all
  • Find the school email
  • Remember whether it was the outdoor trip or the museum trip
  • Check the weather
  • Know whether the raincoat had been outgrown
  • And do it all again next month when the next trip appears

That recurring cycle, notice, gather details, compare against what you know, make the list, buy what's missing, is the loop. Without a system, the loop rests on your shoulders every time.

From Loop to Loop of Loops

Going a level higher, a loop of loops detects "there's a school trip coming" and wakes the relevant loops together:

  • The packing loop checks what you already own
  • The weather loop sees it will rain
  • The school schedule loop sees pickup moved earlier
  • The calendar loop notices the new pickup time now conflicts with a meeting you're in
  • The message loop drafts a text to the other parent, but stops before sending it

That last detail is the heart of the idea. Self-organizing is only trustworthy when the AI does not autonomously fire off messages while you scramble to unplug the computer. It is recurring jobs that notice each other and hand off context, then pause where a human decision belongs.

Why Apps Failed Us

Most useful work is not a single question; it is a recurring situation that takes memory. A customer needs follow-up, a draft needs revision, a database needs refreshing, a client needs a response today.

Apps gave us only pieces of the loop: your email holds the confirmation, the school portal holds the pickup change, the grocery app holds the list. But the loop does not live inside any one app, it lives between them. For the last twenty years, the wiring between apps (open the calendar, text the parent, look up the address, set the reminder) has been on us. Apps digitized every piece of life and quietly made you do all the connecting. The first genuinely useful agent is therefore not another app in a box; it is something that sits across the loop that carries your load.

Finding Your Own Loops: Loops Organize Attention

Nate stresses that his own setup (heavy on AI research) is only an example to illustrate the shape. Once you see the shape in one place, you spot it everywhere. He suggests interrogating your own life:

  • What current state am I carrying in my head for these recurring jobs?
  • Where do I check things for quality?
  • Where do I need to see multiple versions to trust something?
  • What is waiting on me? What is blocked? What changed since the last pass?

The key reframe: loops organize your attention. Your job is to make as many loops run cleanly as possible so you only get woken up when it really matters. Think of building a loop as allocating your attention once to create something durable.

Building Loops: Worked Examples

Nate emphasizes the gap between prompting and looping is "99% of your mental effort," and that loops are not intuitive, which is why he is so concrete.

The simplest loop

"Every Monday, tell me what's going on in the AI world. Give me the top 50 tweets, organized by theme, and do it again next Monday." That recurrence plus memory is a loop in its simplest form.

The simplest loop of loops

Add a second loop that checks multiple individual loops to find what's good. For example, a Twitter loop and a Google AI-news search loop (Claude or Codex can search directly, which is far easier than screen-recording a search). A loop of loops looks across both sources, aggregates them, and gives you one perspective on what matters today. Loops simply tell each other: here's what changed, here's what ran, here's what needs the human, here's what woke up another loop, here's what stopped.

The boring-but-real examples

  • Kids' clothing sizes: a prompt can suggest a size; the loop version remembers what you bought last time, notices the season coming, tracks the child's growth, and proactively says "you may need to size up," catching the problem before it happens.
  • The spinach example (from a friend): a prompt asks when you bought it; the loop version knows it went in your cart Sunday, and pops up Wednesday night to say you have about 24 hours left, so put it in tonight's recipe.

The point is not grocery management specifically. It is caring about the practical loops that "eat our real brains up and make us tired," far more than any benchmark score.

Keeping Agents Safe

This is explicitly not about never asking. An agent that never asks is dangerous. Nate's framing for safe loops:

  • Give the agent a world state to engage with (maybe it researches it)
  • Provide a trigger and a set of reasonable choices to reason through
  • Give it an action you feel comfortable with, then let it loop and do it again

If something recurs, you can loop it. To design one well, ask question-quality questions: What can it do safely? What should it ask? What record should it leave behind? How would it get smarter next time? And, importantly, what other recurring jobs should know about this? If the trip loop finds rain, what else needs to know? If the hiring loop finds a candidate, who needs to know? If the finance loop sees a travel charge, does the trip loop need it?

Loop of Loops as a Control Pattern

At the loop-of-loops level, the questions change, and Nate notes this is rarely covered well. A loop of loops is really a control pattern across loops. You don't get there by asking "where's the specific pain here?" That question is for designing an individual loop ("I have recurring pain; let me describe it well enough to delegate it"). The loop-of-loops question is higher level: what entire process am I willing to let go of and trust the AI to run? For example, delegating the whole "get the kids ready for summer camp" process. You only reach that level after thinking through the individual pieces of pain underneath it.

Getting Started Safely

If you want an early start on a loop of loops, pick a process where:

  • You're confident enough that if it runs off the rails, you can live with it and laugh
  • It's tedious but not high-impact, so the downside is low and the relief is high

Do not start with anything involving banking. A good first candidate: filling out product use cases (tedious), then turning those into Linear tickets, then having another loop tackle PRDs. That chains naturally into a loop of loops, and if it goes wrong, you have some cleanup but nothing catastrophic.

Key Takeaways

  • Stop thinking one task at a time. A prompt is one request; the real work is recurring jobs that take memory.
  • A loop is a recurring job with memory; a loop of loops is loops that notice each other, share what changed, and respect your boundaries.
  • The loop lives between apps, not inside any one of them. The connecting work has always been the human's burden.
  • Loops organize attention. Build them so you are only interrupted when your judgment truly matters.
  • Safety is built in by design: give the agent a world state, reasonable choices, and a comfortable action, and have it stop before irreversible steps (like sending the message).
  • Design with question quality: what's safe, what to ask, what record to leave, how it gets smarter, and which other loops should know.
  • Start tedious and low-stakes (use cases to tickets to PRDs), never banking, and laugh if it breaks.
YouTube