the cure for the AI hangover

The absolute slowest way to use AI is to type at it all day.

You end up trapped in 'prompt-jail,' babysitting a machine that just hands you more work.

But what if you didn’t have to hold its hand?
What if your AI actually did your work, automatically, and you could get your life back?

The smartest builders in the world have completely spoonfeeding their AI and how they did it just broke the internet.

It’s a self running system and they call it loops.

So I built a free, open-source tool to help you make loops (AI that works for you). Even if, and especially if, you have no idea what the heck a loop is.

It’s called Loop Kit (scroll to the bottom to get it) and it’s here to free you from your desk.

Let me explain. Over the last few months, I’ve shipped a lot. Dozens of open-source projects, triple that for private projects. A dozen sites. My own SaaS. Free tools. Internal tools. Tools I built with founders. Content. Head spinning?

You get the picture. But by May, I felt trapped.

I was jumping from prompt to prompt across my mac, my phone, anything with an internet connection, just to keep my bot army running. I was blasting through work, but I had become a full-time prisoner. For every task I automated (and I automate A LOT), 5 new ones emerged in its place. Like an AI hydra rearing its ugly head at Captain Matt-Merica.

I call this the AI hangover.

And it’s not just me. I talk to founders all day. And I hear the same story every single day. But I hear two sides of it:

Turns out, the masterminds behind Claude Code and Openclaw, were feeling the same way, but they found a way out.

Peter Steinberger, the creator of OpenClaw, posted a single take that did millions of views: you shouldn’t be prompting your AI anymore, you should be building loops that prompt it for you. Boris Cherny, who built Claude Code, said it even simpler: his job isn’t writing code anymore. It’s writing loops, and letting the agent ship the work.

What do they know that we don't?

Here's what.

The secret is firing yourself as the middleman.

If you keep trying to play the prompt game on hard mode, you are quietly letting your most valuable asset, your own time, dwindle down to absolute zero.

The first loop I built was for SEO articles. Tedious yes. But I’ve seen enough AI slop writing to know what I didn’t want.

So I took my SEO kit, a thing I'd already launched and helped a bunch of founders implement for themselves, and spun up a loop version just for our new company site.

It grades its own work hard. Makes its own photos. Checks the page before anything goes live.

To break it down really easy:

The first (non automated) pass I did with the AI and it took me about 70 minutes working casually from my phone.

And here's what I want you to hear. 50 of those 70 minutes wasn't the wiring.

It was teaching.

The same way you'd coach a new hire after their first draft. I told it the tiny things I cared about. The proof I want it to show off. The phrases I never want it to use. My voice. My standards.

The kind you only want to do once. So I made sure it was once.

But now when the loop runs it has a judge that checks every future draft against that bar. And that bar gets tighter over time.

Once the job finishes, it kicks off a new one on its own. Nobody presses go. Then it just runs.

And when I turned the loop on for my second batch I had 5 more articles ready to roll.

All reviewed, and ZERO edits needed by me. I approved it one-shot. From my phone. Cuddled on the couch with my wife watching the new Yellowstone sequel. Yup. I can be that guy now.

That's the promise in one picture. The work got made well, while I was doing something else.

Reply LOOP with the job you keep redoing and I'll send you the loop shape for it. Free for the first few.

Automating repeatable work is easy. I do this all the time.

But if you run a business, you know you are constantly juggling new projects. A new brand. A new content format. A new SaaS feature.

And here’s the truth: you can’t hand off 'new' to an AI blindly. New requires your actual judgment.

This is why self-running loops only work when you build what’s called a Judgment Gate.

You automate the plumbing. That part is easy. But the judgment, the hard part, is where you have to bake yourself into the process.

But here is where things get crazy... Machines don't know what magic feels like. They don’t know what a banger social video ‘feels like’. Not yet. So you have to write down exactly what 'good' looks like, and sharpen it every single round.

Think about it. Your first time doing anything usually sucks. AI's no different. Its first attempt never blows me away. It's kind of just meh.

The loop is how "meh" becomes "yup, ship it" without you sitting there every round. Or even being there to get it started.

People ask me this all the time, so let me be crystal clear. I’m not scared of AI. I jump on new tech fast. But I am ruthlessly careful about what I actually ship.

Because the more I build, the more I realize one undeniable truth.

If you build a skyscraper on a shaky foundation, it tips.

This is exactly why I Loop Kit is built with something called Checkpoints.

All that is is a brutal, hard bar that the AI has to pass before the work moves on. Think of it like a bouncer. The written standard is the VIP list. The bouncer is the gate. If the AI doesn't meet the standard, it doesn't get in.

People think gates slow you down. They don't. Gates are the exact thing that allows you to build tall.

But here is where the panic sets in... Every single week, I hear the exact same two worries from founders.

These are fair questions. I know, because I used to ask them myself. But after building dozens of these systems, the answer’s simple.

Now, I know the idea of 'Checkpoints' can sound a bit abstract. So let's get completely tactical.

How do you actually design a Checkpoint for your own AI system? Before you build anything, you need to ask yourself this one million-dollar question: What is the absolute worst-case scenario if this specific task goes wrong?

If you’re automating an email newsletter, your Checkpoint isn't 'Did the AI write 500 words?' Your Checkpoint is 'Does this email sound exactly like the brand's voice without hallucinating fake links?'

If you are automating code, your Checkpoint isn't 'Did it run?' It is 'Did it pass all five of our security tests without throwing an error?'

Whatever that worst-case scenario is, you write a hard, brutal rule to prevent it. That rule is your Checkpoint. That rule is your bouncer."

But here is the specific loop that completely changed my week.

Every single day, I have to review a mountain of work. Static ads. Videos. Internal projects. Written copy. Before this, that process stole 2+ hours of my day. I was grinding through items, repeating the exact same opinions, and becoming the ultimate bottleneck for my entire team.

So, I made the one move that actually mattered. I took my personal standards, built them into a Loop, and gave it to my team.

Now, the machine applies my opinion and my bar to the work before I ever even see it. They check the work against my standard, without me having to be in the room. It auto-runs. It learns. It gets sharper every single time.

Here is what happens when you build this.

You stop being the bottleneck. That massive block of wasted time completely disappears. You actually get your day back for the high-leverage things that require a human, instead of doing the exact same manual review over and over.

If you want to stop getting dragged into the weeds... Reply 'LOOP' in the comments and tell me the one review task you hate doing most. I will personally reply and tell you exactly how I would build a Loop to automate it for you. Free for the first few.

So, what actually is Loop Kit? It is three specific skills, packed into one single install. Built for Claude Code, Codex, or any agent that reads skills.

Now, I originally built this for myself. But after talking to hundreds of founders, I realized we’re all on the exact same journey. We’re all just trying to build our brands and figure out these tools without losing our minds. So, I’m open-sourcing the exact system I needed, and I am giving it to you.

It is broken down into a simple, three-part architecture:

Think of this as the front door.

You don’t need to know architecture jargon. You don’t need to write complex risk codes. It just asks you one or two plain-English questions, and immediately hands you a system you can actually run. You just answer the questions, and it maps the loop.

This is for when you already know exactly what you want.

It maps the entire loop, writes the plan, and gives you a prompt for a no-stakes dry run. It produces the smallest, safest version that actually gets the job done. Map it. Plan it. Dry-run it. Ship it.

This is the part that actually lets me hang out with my kids at night.

Before any loop gets scheduled, it asks one brutal question: Is this safe to run, and where will it hurt you?

If anything is unclear, it completely stops instead of plowing ahead.

The default is 'don't,' never 'send it.' It’s the ultimate safety net running underneath everything you build."

Under all three skills is one stubborn little check. It's the whole trust layer.

Before I let any loop run, it has to answer four dead-simple questions. The same ones I'd ask a new hire before handing them the keys.

Can't answer all four? Then it's not a loop. And the kit will tell you so, and hand you a checklist instead.

This is the part I'm proudest of, because these are the exact ways a good loop quietly goes bad. loop-kit names them so it can catch them.

→ It says "done" when it isn't. A soft "done" that hides a half-finished job. That’s the Ralph Wiggum loop.
→ It keeps running long after nobody on the team understands it anymore, and the unread output piles up. That's comprehension debt.
→ You quietly stop checking the output and just trust it. That's cognitive surrender.
→ You leave it running unattended and it becomes a door you left open. Passwords leak, permissions creep. That's the security tax.

There's one number I watch on every loop too. If the loop keeps handing you changes and you keep saying no, it isn't helping.

So I ask: am I accepting more than half of what it proposes? Under half, the loop is losing. Kill it or fix it.

The old way is you, in the chair, every run, forever. Every draft, you. Every review, you. Every handoff between Codex and Claude and whatever's next, carried by hand. You.

Add it up and it's 2+ hours a day on reviews alone for me. Plus the 50 minutes of teaching baked into every new thing I build. Plus the prompt I write to make the prompt to build the system. Forever.

loop-kit gets you the system once and gives you your life back.

You teach it your standard one time. It holds that standard on every run after, while you're somewhere else. The old way costs you the rest of your year, an hour at a time. This costs you one good afternoon of setup.

For free.

Loop kit is open source. MIT. Free.

You get:

→ three skills: loop-start, loop-builder, loop-doctor
→ the safety engine: the four-question, stops-when-unsure check that every loop has to pass
→ the named failure modes (Ralph Wiggum loop, comprehension debt, cognitive surrender, the security tax) baked into the audit
→ a pattern library index of proven loop shapes, with titles, sources, and links
→ supporting scripts for onboarding, plain-language translation, pattern selection, and a self-audit
→ reference playbooks: archetypes, loop composition, spec and launcher templates, onboarding, red flags, example loop packets
→ install.sh for Claude, Codex, or both, plus doctor.sh
→ tests, a golden-case baseline, MIT license, and a SOURCES notice

No upsell. No catch. MIT license. Fork it.

54% of Americans are sick of AI. They’re experimenting every day, and they’re exhausted.

That’s the AI hangover.

It's the exhaustion of babysitting a problem-machine. It's getting bloated 30-page documents you didn't ask for. It's not trusting the work, and ending up re-reviewing it all yourself anyway.

But here is the undeniable truth... This sickness is just a sign of premature systems. The bet on AI wasn't wrong. My belief in it hasn’t changed.

Because right now, there is a team out there completely reinventing how they work. And they’re going to out-ship and out-sell you a thousand times over.

The entire game is getting the most signal from the noise. You have to build systems and Loops where you get massive output for the absolute least amount of time.

That’s exactly why I built Loop-Kit to cure my own AI hangover. It’s the smartest thing you can do for your bottom line.

Go big,

Matt"

P.S. loop-kit is the on-ramp layer. It's what gets AI running for you safely. Point your loops at the work you've already systematized: Frontrun for market voice, Short-Form Idea Engine for video ideas. loop-kit is the layer underneath that makes those run on repeat without you in the middle.

P.P.S. If you want my team to map your first loop with you, reply "loop" and tell me the job you keep redoing. I'm taking a few a week.

P.P.P.S. Star the repo if you grab it. It tells me to keep building.