In partnership with
Hey folks,
Normally people shouting "prompting is dead" on YouTube makes you do a double take. This one caught my eye for a different reason.
Boris, the guy who built Claude Code, has been saying the same thing:
I don't prompt Claude anymore. What I mostly use now is loops. I create loops. They do my entire job.
Here's the gap nobody was filling. Everyone's saying build loops. Almost nobody shows you what a loop actually is, or how to set one up.
So I tested the idea, and wrote up the part most people skip.
Let's dive in..
The context
Loop engineering is a small shift with a big payoff.
Right now, when you use AI, you are the loop. You prompt it, wait, read the answer, spot what's off, prompt again, wait again. Nothing moves unless you're sitting there pushing it.
You're not really using the AI. You're babysitting it.
A loop hands that job over. You give it a goal and a definition of "done," and it runs its own cycle: try, check its own work, fix what's wrong, try again, until it hits the goal.
You're no longer in the middle. You set the target and come back to a finished result.
This only works now because the models got good enough to grade their own output. A year ago they couldn't be trusted to run unsupervised. This generation can self-correct well enough that babysitting every step is mostly wasted effort.
Who this is for: anyone using Claude Code (it also works in Codex). It sounds like a developer tool, but the highest-value loops aren't coding tasks, and there's now a guide aimed squarely at non-engineers. If you can write a clear brief, you can run a loop.
What you'll need:
Claude Code. Note: /goal and /loop don't run in Claude chat or Cowork, so you do need Claude Code
A task with a clear finish line ("done" you can actually describe)
15 minutes to set up your first one
Be honest about the limits:
If you don't define "done" well, the loop runs forever and burns through your tokens
Self-review is good, not perfect. A loop can be confidently, calmly wrong
It lives in Claude Code, not the tools you may already use. If you only touch Claude chat or Cowork, there's a small setup step first
The walkthrough
Every loop has 4 pieces:
a goal (what you want)
a checklist (what "done" means)
an inspector (the AI grading its own work)
a budget (a cap so it stops)
Get those 4 right and the rest is syntax.
Step 1: Write a goal, not a prompt
A prompt says "do this thing." A goal says "do this thing, and here's how you'll know you're finished."
In Claude Code you use the /goal command. Here's one for auditing a website. Notice how specific "done" is:
/goal Audit my website https://beyondthebuzz.ai/ like a conversion consultant I'm paying $5,000. Write the findings to website.md. You're done when you've listed at least 10 specific issues. Every issue must have a severity, the exact location on the page, and a fix I could ship today. Rank the top 3 by revenue impact. Every item must reference something actually on the page. No generic advice. Stop after 12 turns."At least 10 issues," "reference something actually on the page," "stop after 12 turns." Those aren't style notes. They're the conditions the loop checks itself against.
Pro tip: if you can't write a tight goal, ask the AI to write it first. "I want to review my website. Write me a /goal with a defined set of deliverables." Then run the one it gives you.
Step 2: Put it on a loop
/goal on its own runs once. Add /loop and it repeats on an interval.
There are 2 flavors, and Claude picks for you:
Inner loop: one mission where the AI cycles internally. It scans, checks, fixes, re-checks, then marks it done. One goal, many internal tries.
Outer loop: a job that keeps running on a schedule. "Every morning, scan my 3 competitors' channels and log what's outperforming." It runs the whole task again over days.
The everyday version most readers will use: point a loop at your Downloads folder to sort files on a schedule, at your bank CSVs to categorize spending, or at a Slack channel to compile the week's blockers. Same mechanism, no code.
Step 3: Cap the budget before you walk away
This is the step people skip and regret. A loop with no limit will happily run all night and torch your subscription or API tokens.
Set a ceiling every time: "Stop after 12 turns." "Stop after 10 iterations." "Stop when the checklist is complete."
Watch out: the classic beginner mistake is an open-ended loop with a vague goal. It never decides it's finished, so it never stops. A clear "done" and a hard turn cap are your 2 brakes. Use both.
Step 4: Schedule it for when you're away
/loop only runs while your computer is on. /schedule runs on set times.
Simple way to hold the 3 apart: /goal is the mission, /loop is the interval, /schedule is the exact time. A daily competitor scan at 6am is a goal, on a loop, on a schedule. That's your first "working while I sleep" automation.
Step 5 (if you code): the build-and-review loop
If you vibe-code, you can chain 3 loops into a near hands-off pipeline: spec, build, review.
Spec: you give a rough idea, the agent interviews you (about 15 questions), then creates detailed tasks.
Build: a loop reads each task and builds it out, over and over, all day.
Review: another loop opens a browser, tests each build, clicks around, then opens a pull request.
The only manual work is the 2 ends: a few ideas in the morning, and a quick approve at night. Everything in between runs itself.
The system
Here's why this is more than a party trick. Loops compound with 2 things you may already use: memory and skills.
Give the agent a memory file so it stops forgetting between sessions, and its loops get sharper every run.
Add skills (reusable instructions for jobs like "write like me" or "review a landing page") and a loop can improve the skill as it goes. Picture a loop that scans writing you admire and updates your "write like me" skill on every pass.
Treat agents like employees, not magic.
A new hire with no onboarding, no context, and no clear brief will flounder. Same with an agent. Define the goal, give it the context, tell it how often you want the work, and let it run.
Push it far enough and you can run your whole life like a small company: a channel and an agent for finances, one for health, one for content, each doing a real job.
This builds on an earlier issue. Stop asking your AI for things pushed a related idea: describe the outcome, not the steps. Loop engineering takes it further. Set the outcome, set the interval, walk away.
Your voice. Every platform. No writing required.
You ghost your own socials by Wednesday. SureThing learns your voice and ships native posts to LinkedIn, X, Instagram, and TikTok, without you writing a thing.
Try this now
Open Claude Code and run one loop on something low-stakes. I'd start with the website audit: you get a useful result and you can't break anything.
/goal Audit my website https://beyondthebuzz.ai/ like a conversion consultant I'm paying $5,000. Write the findings to roast.md. You're done when you've listed at least 10 specific issues, each with a severity, the exact location on the page, and a fix I could ship today. Rank the top 3 by revenue impact. No generic advice. Stop after 12 turns.Swap in your own URL. Give it 10 minutes. You're looking for one thing: a fix specific enough that you could ship it today. Most people get at least one.
Links worth your click
Stop asking your AI for things, the scheduled-agent issue
👉️ Know someone who's still copy-pasting into a chat box all day? Forward it. When one person subscribes from your link, you unlock the Beyond the Buzz Prompt Vault: 10 copy-paste prompts that do real work.
Cheers, Tim
