There's a complaint going around about AI work...
We're a year into the AI hype now, and people are finally admitting the thing they were afraid to. "Wait... am I actually working any less?"
All these new AI tools and their "BEST EVER" models, but many are realizing they're working the same hours, at the same desk, with the same sore neck at 6pm.
The only thing that got faster is the typing.
If that's you, we need to update your AI workflow.
This issue could be turned into a whole consulting service on it's own. You're about to learn a process that you can charge $1,000/hr for and that's not hype.
Loop engineers are the new strategy consultants. But I digress...
A chatbot is just a vending machine.
You walk up, punch in a request, and it drops an answer in the tray. Useful? Sure. But you still have to carry the snack back to your desk and eat it yourself.
You're STILL the one doing the work. You just have a faster snack dispenser. A vending machine with a turbo button is still a vending machine.
The 2027 way of using AI isn't a better vending machine. It's turning your AI into an actual employee (or team of employees). We do this by optimizing for "loops."
- In 2024 we had "prompt engineering."
- In 2025 we had "context engineering."
Now, I want to introduce you to loop engineering.
The Basics of Loop Engineering
You may have noticed that when you ask your AI chatbot of choice to do something, it does all this weird behind the scenes work...
Loading... Calling tools... Eval... Checking again... Great now I can deliver to the user...
In the simplest way I can describe, loop engineering is the understanding of what's going on in that background process and optimizing that process for longer running tasks.
Here is an example (greatly simplified) loop engineering prompt for writing an SEO/AEO optimized blog post...
GOAL: Produce a blog post on [TOPIC] targeting the keyword [PRIMARY KEYWORD] that
is optimized for both Google search and AI answer engines.
Run this loop until ALL exit conditions pass. Max 4 passes.
STEP 1 — DRAFT
Write or revise the post. Requirements:
- Answer the primary question in the first 2-3 sentences (AEO snippet bait)
- One H1, logical H2/H3 structure
- Primary keyword in title, first 100 words, one H2, and conclusion
- Include 3-5 related questions as H2s (People Also Ask style)
- 1,200-1,800 words
STEP 2 — SELF-EVALUATE
Score the current draft against this checklist. Output PASS or FAIL for each:
[ ] Direct answer appears in first 50 words (AEO)
[ ] Primary keyword density between 0.5%-1.5% (not stuffed)
[ ] Every H2 is a real question or clear subtopic
[ ] At least 2 sections are quotable as standalone answers
[ ] Has a definition, a list, AND a step sequence (formats AI engines extract)
[ ] Meta title under 60 chars, meta description under 155 chars
STEP 3 — DECIDE
If ALL items PASS → output final post + meta tags, then STOP.
If ANY item FAILS → list the specific failures, then return to STEP 1 and fix
ONLY those failures. Do not rewrite passing sections.
Begin pass 1.
If you're clever you might be thinking "But I still have to copy/paste that work myself. This is still just a vending machine."
Like I said, that was a greatly simplified loop engineering prompt. It was a demonstration of the concept. But this concept can be used for creating multiple files, connecting to multiple tools, triggering multiple skills, and completing full projects.
One of your requirements could be "- Publish this to my Kit account using the Kit MCP and send it to my email list."
These are the core elements of a loop...
- GOAL: The AI needs to understand what "done" looks like.
- Max pass count (4-10) so it can't spin forever wasting tokens or account usage limits.
- Act → Evaluate → Decide: These are the 3 steps YOU think through to make sure the AI knows what to do, what to include, how to grade itself, and how to know when "good enough is good enough."
- Hard exit condition (all checks pass) so it converges all of its work into the final output.
Don't worry, Claude knows how to write these loops for you.
You can always just attach this issue as a PDF to Claude to teach it as well. You could even attach this newsletter to help Claude make you a "loop writer" skill.
Ideally that skill would interview you using the AskUserQuestion feature, to ask about your goal, max passes, evaluation criteria, and your hard exit condition.
Combining Tools For Weeklong Even Monthlong Running Loops!
Obviously, your standard chatbots can only run while you're sitting there in the chair. You could use scheduled tasks or routines, but even still that's not quite going to give you that "set it and forget it" feeling.
A real overnight loop is a small stack of four pieces working together, and not one of them is YOUR computer.
- A connector (MCP) so the agent can reach into your tools and actually make changes, not just talk about them.
- The loop instructions, the part you just read, the actual loop engineered prompt. Could also be a skill file or multiple skill files triggered by the loop prompt.
- Memory so it remembers any metrics and every change it made. I use and recommend a tool called GBrain. Without memory, your loop starts from zero every run and never compounds.
- Cloud hosted agent. I use Hermes, running on a $24-a-month digitalocean droplet. A scheduled job (a "cron job," which is just a timer that says "run this at 6am") kicks off the loop while you sleep. Your laptop can be closed, heck your power could be out from a hurricane and this loop keeps running.
Are you beginning to see the power of loop engineering now?
How this could be setup so that your blog just blogs for you, or how your Twitter just tweets for you, and optimizes itself for lead gen?
The Bad News About Loops
Loops aren't free. In fact this is being called the most expensive use case for AI. But that's because people are doing it WRONG!
But first, let me be fair to my opponents. An agent burns about four times the tokens of a normal chatbot conversation, because it's not answering once, it's working through a problem again and again.
Point a whole team of agents at something and you're suddenly looking at roughly fifteen times the tokens.
That sounds like a lot until you start thinking like an ad buyer. A game for which I am uniquely qualified.
I've spent tens of millions of dollars on paid ads across 14+ networks in my 20 years in this game, and I've babysat live accounts by hand at 6am more mornings than I'd ever admit to my wife.
The game is token earnings > token cost
You just need to make sure you're using loop engineering on tasks or projects that earn more than the tokens cost. Same as we used to make sure that our payroll expenses were less than our total revenue.
The only thing that's changed are the tools.
A loop that runs that same grind all night, for the price of a few bucks in tokens, is the best deal I've seen in marketing in a long time. Tokens are still FAR CHEAPER than human labor... yours included.
Token spend can be optimized too, by setting different models to do different tasks.
For today's premium section we are going to build a loop engineering process that automates your paid ad campaigns.
The loop we will build together will:
- Connect to your Meta or Google ads
- Find the winners and losers
- Write new ads
- Decide what to pause and what to boost in order to fully optimize your paid ad results until it's reaching your profit goals.
Continue with Build Notes+
This is the free preview. The complete step-by-step guide and supporting files are available to Build Notes+ members.
