The Drop | You Are the Loop.

This is how you stop it.

Jul 14, 2026

TL;DR: You're currently the machine that keeps your AI running: checking, correcting, re-prompting. A loop flips that. Two real Claude Code commands (/goal for finish-line work, /loop for rhythms), five beats, one memory file, and the checking happens without you. Start with one repetitive task this week.

The person who built Claude Code at Anthropic said something recently that stopped me mid-scroll: he doesn’t prompt Claude anymore. Loops prompt Claude for him. His job now is writing loops.

Read that again. The man who built the tool stopped using it the way the rest of us do.

And before you file this under “tech people doing tech things”: this is the biggest shift in how AI actually gets used right now, and almost nobody outside developer circles is doing it. Which means you, reading this, can be early. Ajde.

First, let me show you the problem you already have

Here’s your current workflow with AI, if you’re honest:

You type a prompt. AI produces something. You check it. Something’s off. You explain what’s off. It fixes that, breaks something else. You paste the problem back. It tries again. You check again.

Notice who’s doing the checking, the judging, the deciding-what’s-next, every single round?

You. You are the loop. The AI does the typing, but you’re the machine that keeps the whole thing running. That’s why “AI saves me so much time” often feels like a lie you tell yourself while babysitting a chatbot at 11 PM.

My Baka never re-explained čipka to her own hands. The pattern was set once. The hands checked their own stitches. She could talk, drink coffee, watch the neighbors, while the work verified itself. That’s the difference between doing work and holding a system. She held systems. We hold threads.

Kava je na stolu — and so is every new post. Subscribe free for AI strategy, brand voice, and a Balkan perspective on building something that actually sounds like you.

What a loop actually is

A loop is a small system that prompts the AI for you, over and over, until the job is actually done. Not “produced.” Done. Every loop, stripped down, is the same five beats:

Find the work. Scan for what needs doing: files in a folder, open tasks, unread emails.
Do it. One item at a time, the way you would by hand.
Check itself. A second pass confirms the work is correct, not just generated. This beat is the whole magic.
Remember. It writes down what’s finished, so it never repeats work or loses its place.
Go again. Until nothing’s left. Then it stops, or pings you.

Beat 3 deserves its own paragraph, because it’s the one that fixes the problem that quietly burns everyone: AI confidently making things up. Ask AI for a researched brief and it hands you clean prose with sources at the bottom. Open the links. Some go nowhere. Some go somewhere that doesn’t say what was claimed. A single prompt can never catch this, because the AI stays confident it’s right until something actually opens the link. A loop opens the link. Every link. It never gets bored, and it never skips the tedious ones. That’s not a skill you or I can compete with at 11 PM.

“Isn’t that just a scheduled task?”

Fair question. No. Your computer has been able to run the same script every morning at 8 AM since before most of us had email. That’s a cron job: fixed script, same steps, zero thinking.

A loop has one thing a script never had: a decision-maker inside it. The AI looks at the current situation, picks the next action, does it, checks the result, and then decides: keep going, try again, undo, stop. A script can’t look at a broken result and figure out a different approach. AI can. This only became possible in the last stretch, when the models got good enough to make real judgment calls mid-job. That’s why “almost nobody does this yet” and “you can start this week” are both true at the same time.

The two commands (yes, they’re real, yes, you can use them)

In Claude Code, this isn’t a hack you duct-tape together. There are two built-in commands, and which one you reach for depends on the kind of loop you need:

/goal is the loop with a finish line. You describe what “done” looks like, and Claude works turn after turn, on its own. After every turn, a second copy of Claude quietly checks: are we there yet? If not, it tells the first one why, and work continues. The moment the goal is met, the loop stops itself.

For a content business, that looks like:

/goal every article in my drafts folder has a Pinterest 
description, an Instagram teaser, and a Substack Notes 
version. Check each file after writing. Stop only when 
every article is covered.

You come back to a finished content batch, not a chat history.

/loop is the loop with a rhythm. No finish line, just a beat. You tell it how often and what to do, and it re-runs the same job on that rhythm.

For a content business, that looks like:

/loop 1d check my Substack inbox and Notes replies. 
Draft a short, warm response for each new comment in 
my voice. Save the drafts in a file for me to review. 
If someone asks about working with me, flag it 
separately at the top.

Every day, the same quiet round: comments gathered, drafts waiting, leads flagged. You review with your morning kava instead of chasing notifications all day. The 1d means daily; 30m would mean every 30 minutes. You can also just write “every morning, triage my inbox” and Claude schedules it itself.

Rule of thumb: finish line → /goal. Heartbeat → /loop. Most strong loops start with /goal, but the ones that change your daily rhythm are usually /loop.

The quiet hero: the state file

Here’s the detail most people will skip, and it’s the one that separates a loop that works from a loop that wanders: tell it to keep a memory file. After each finished item, the loop writes the item name and status into a simple file (call it LOOP-STATE.md, call it whatever). Every new run starts by reading that file first.

Without it, every run starts from zero. With it, the loop picks up exactly where it left off, even days later, even on a schedule. It’s the difference between hands that remember the pattern and hands you have to re-teach every morning. Baka’s hands had a state file. It was called forty years of repetition. Yours gets to be a text file. Nema žurbe, but also: what a shortcut.

Three honest things before you build one

One-off tasks don’t need a loop. If the job is a single answer, a plain prompt is faster. Loops earn their setup cost on work that repeats or has many pieces.

Loops cost more. A loop that checks itself and retries is running the AI several times per item. On a subscription plan, you’ll hit your usage limit faster. That’s not a flaw, it’s the price of not doing the checking yourself. Decide consciously.

Vague work doesn’t belong in a loop. “Think of a better content strategy” is not a loop. If you can’t describe what done looks like, the problem isn’t the tool. Go figure out the goal first. Pomalo means deliberate, not fuzzy.

Start this week

Pick one task you keep doing by hand, the kind with lots of small pieces. Write what done looks like, where the work lives, and how the loop should check itself. Run it once while you watch every step. Only when you trust it, put it on a schedule.

The first time you wake up to work that finished overnight, correctly, verified, you’ll understand why the person who built the tool stopped prompting it.

Prompting is doing the work. Loops are managing the worker. Baka never did the stitches one by one while staring at her hands. She set the pattern and let the pattern run.

Tako se to radi.

I hope you liked today’s Drop.

Ajde, until next week

Anita

Discussion about this post

Ready for more?