Cliff Marquez

Building with AI daily and sharing what I learn as I go. Operator, not a
guru. Business, ideas, life — and dog lover.

[Cliff Marquez @cliff_marquez]

I’ll share the free GitHub starter kit at the bottom — but first, here’s why I built it.

Most people think they need a smarter AI.

What they actually need is less friction.

You don't need next month's model or the newest agentic framework.

You need to stop losing 30 seconds every time you can't find the brand asset folder.

You need to stop opening a new chat and re-explaining your business from scratch.

You need to stop breaking flow because you have to manually switch windows to approve an image.

The bottleneck isn't the model.

It's the hundred tiny friction points you've accepted as normal.

Here's what I mean:

You're building out 20 blog posts.

You batch the keyword research. Done.

You batch the outlines. Done.

You batch the copywriting. Done.

Then you hit images.

And suddenly you have to stop.

Open your image tool in another window.

Midjourney, GPT Image 2, Gemini — doesn't matter which.

Generate.

Wait.

Download.

Check it.

Go back to your doc.

Paste it in.

Repeat 20 times.

And here's the thing — wiring it up to the API doesn't fully solve it either.

The AI can generate every image for you, but you still have to look at each one.

Copy you can read and approve right in the terminal.

Images force you to leave it.

That one friction point just killed your momentum.

Not because the AI couldn't write.

Because your system made you context-switch.

This is what people miss about working with AI:

Speed isn't about the model.

It's about the environment.

If your folder structure is a mess, you waste time hunting for files.

If your AI doesn't know where your brand voice lives, you waste time re-prompting.

If your images live in one place, your copy in another, and your assets in a third, you waste time stitching it all together.

The solution isn't more tools.

It's better structure.

I run everything through a simple folder system that mirrors across my entire operation.

Google Drive matches my local files.

My cloud storage matches my Drive.

My email folders match my brand structure.

My AI knows where everything lives because I built it a map.

When I need to generate 20 blog posts, here's what happens:

Batch 1: Research.
AI pulls keywords, competitor analysis, and SEO data.
It outputs markdown briefs into the research folder.

Batch 2: Writing.
AI reads the briefs, checks the brand voice file, and writes the posts.
It outputs drafts into the content folder.

Batch 3: Images.
AI generates image prompts based on the content.
I batch-approve them in one session.
They go into the assets folder.

Batch 4: Assembly.
AI pulls everything together, formats for web, and outputs final files ready to publish.

No context switching.

No hunting for files.

No re-explaining what the brand voice is.

The AI isn't smarter.

The system just has less friction.

Here's the part that changed everything for me:

I stopped thinking like a prompter and started thinking like a system architect.

Instead of asking, "How do I get better output?"

I started asking, "Where is the friction?"

Where do I have to do something manually that breaks my flow?

Where do I have to open a second window?

Where do I have to re-enter information I already gave?

Where do I lose momentum?

Then I built small tools to remove those points:

- Voice dictation so I can brainstorm without typing
- Color-coded terminals so I know which agent is doing what
- A dashboard that shows all active operations
- A way to pass context between machines without starting over
- Mobile access that doesn't depend on third-party apps

None of this is complicated.

It's just organized.

And organization is underrated.

Because when you can see where everything is...

When your AI can see where everything is...

When your folder structure matches your mental model...

When your tools talk to each other...

You stop building.

You start operating.

You can run three brands at the same time because you're not drowning in chaos.

You can batch 20 blog posts because the system handles the assembly.

You can stay in flow because nothing forces you to stop and hunt.

The future of AI productivity isn't better models.

It's better environments.

Most people are trying to make AI smarter.

I'm trying to make my workspace frictionless.

That's the difference.

If you want to build something like this, I made you a starting point.

I took my own system and cleaned it down to a free starter kit:

- The folder structure my AI works inside
- A matching set of ready-made Google Drive folders
- One file you drag into Claude that checks the kit, interviews you, and builds your workspace
- 12 copy-paste prompts that run the system day to day, including a monthly friction audit where AI reads your own work history and shows you where you're losing time

It's on GitHub.

Free.

No email.

No opt-in.

Nothing to sign up for.

https://t.co/Laj54vNEbz

Read the first file, feed the second one to Claude, and you're running.

If it saves you even one of those 30-second hunts, it did its job.

[Cliff Marquez @cliff_marquez]

Quick note between chapters. Not a full kit, just the thing to set up before the next one.

You've now handed Claude a few keys: Cloudflare, your storage, your registrar. Where are they actually sitting right now? If the answer is "in a text file somewhere," let's fix that first.

Put them in a password manager. I use Bitwarden. Here's exactly how I run it, nothing fancy:

Logins. The browser extension fills my passwords everywhere. If you do one thing tonight, get every important login in there.

API keys. I save each one as a secure note, named the same way, like cc-mailgun-api, cc-twitter-api. So when Claude needs a key months later, I search, grab it, paste it. They live in the vault, not scattered across random files.

One honest note, because I don't want to teach you wrong. Pasting a key to Claude does put it in Claude, and you can absolutely avoid that: keep your keys in a hidden .env file and have Claude read them on demand. Claude itself prefers that, it would rather not hold your secrets. It's the more careful path, and if that's how you want to work, do it. I understand security, and I respect it.

Me, I trade a little of it for speed. Opening a file, finding the key, pointing Claude at it, that's friction I don't want fifty times a day. I'd rather keep the key in Bitwarden, click copy, and paste it straight to Claude, and put my guardrails on the API side instead: a spending limit with alerts on the account, an IP whitelist where I can, and a scope so the key can only do one job. Not the most secure setup, and I know it. It's what's worked for me. Pick your own spot on that line.

I pay for the Premium tier, $19.80 a year, about twenty bucks. Not required, but here's why I do it: some sites make you use two-factor, and instead of a separate authenticator app, Bitwarden generates those codes for me too. Logins, keys, and my 2FA in one place.

(There's a developer-grade "Secrets Manager" where your AI pulls keys automatically. I keep it simple with notes. And Bitwarden isn't the only good one. 1Password and others are solid. Use whatever you'll actually keep locked.)

Next chapter is your first key that actually does something: getting email working. And now it'll have a home the moment you make it.

[Cliff Marquez @cliff_marquez]

Everyone's chasing long-term memory for AI.

We're heading there. No doubt.

But here's what nobody talks about: memory gets stale.

It fills up with old context. Dead projects. Workflows you moved past months ago. The system gets slower, dumber, noisier.

Here's what works for me. Not elegant. Not one click.

End of every month, I sit down with Claude and do a full review. What's stale? What's outdated? What folders and handoffs are adding noise instead of value?

I archive. I delete. I reorganize.

Sometimes new processes come out of that session. Sometimes I realize a folder belongs at root instead of buried somewhere it made sense six months ago.

AI will automate this eventually. But the call on what you've outgrown — that's still you. And it takes real time.

How are you keeping your AI systems clean?

[Cliff Marquez @cliff_marquez]

THE SHORT VERSION:

A domain is the one piece of your whole setup you actually own. You rent the server. You subscribe to the software. The domain is yours, and it's about ten dollars a year.

I buy mine at Porkbun. Cheap, WHOIS privacy included free, and a clean API. That API is the point: hand Claude the registrar key and "check if this domain is free, buy it, and point it at my server" becomes one sentence.

Then point the domain at your Kit 3 Cloudflare account, and your domain, DNS, tunnel, and CDN all run from one control panel.

Full breakdown below. Copy-paste prompt at the bottom.

—

Part Five. Your domain, the one piece you actually own.

The last few parts built your foundation: Cloudflare as your control panel (Part Three), and your storage and CDN living inside it (Part Four). This part gives all of it a real address.

Here's the thing about a domain. It's the one piece of your stack that's genuinely yours. Servers you rent. Software you subscribe to. The domain you own outright, and it's the cheapest business asset you'll ever buy. Everything else points at it.

Where I buy them: Porkbun. Not a sponsor, just what I use. About ten dollars a year for a .com, with the two things I actually care about baked in.

One, WHOIS privacy, free. When you register a domain, your name, address, and phone number go into a public database called WHOIS unless privacy is on. A lot of registrars charge to hide it. Porkbun includes it. Your home address shouldn't be one search away because you bought a domain.

Two, a clean API. That's the part that matters for how we work. The throughline of this whole series: you hand your AI a scoped key, and the boring stuff becomes a sentence. Same move as the Cloudflare key in Part Three and the storage key in Part Four. Give Claude your registrar key and "check if this domain is available, buy it, and point it at my server" is one instruction, not twenty minutes clicking through a checkout dodging upsells.

(Namecheap and GoDaddy work too. Namecheap is a solid, popular pick. GoDaddy reopened its API to single domains in 2026. I just land on Porkbun for the price and the free privacy. Different tool, same job, use what you like.)

And since you're already in Cloudflare from Part Three: they sell domains too, at cost, right inside the same account. Honestly that's the easiest path of all. Buy it there and it's already wired to everything, nothing to point. Totally fine to do it that way. I keep mine at a separate registrar on purpose, for one reason: your domain is the one thing you truly own, and I don't want it living inside the same account that runs my infrastructure. If anything ever happens to that account, a lock, a billing dispute, a mistake, a domain held somewhere separate can be re-pointed elsewhere in minutes. It's your escape hatch, and I don't keep the escape hatch inside the building. If you'd rather have it all in one place, buy it on Cloudflare and you're done.

One thing worth getting right first: which domain to buy, because this first one is your hub. If you already have a business or a brand idea, easy, buy that. If you don't, buy your own name. Either way you want one primary hub with a real presence on it, plus the basics like a privacy policy and terms. Here's why that matters more than it sounds: the moment you start signing up for the tools we've been using, some of them ask for your website and your terms before they'll hand you a key. A hub connected to your business or your name is what clears that, and you want it to already exist. If you do have a company, point this at the company. It becomes the home base your other sites hang off, and it's what platforms want to see when you apply for API keys and ask "what's your brand."

You don't need to be incorporated for any of this. A business name you run under your own name works fine. Personal or business, either one is golden. I'm not telling you how to structure it, and none of this is legal advice, it's just what's worked for me.

Then you connect it to what you already built. Point its nameservers at the Cloudflare account from Part Three. Cloudflare gives you two (they look like https://t.co/oPCH2rpZlQ and https://t.co/Vh8qP4WOUm), and if your registrar makes you set them by hand, Claude gives you the exact two to paste and tells you where they go. Now your domain, DNS, tunnel, and CDN all live in one control panel, and your AI has the keys to all of it. The domain points at Cloudflare, Cloudflare's tunnel points at your server, your site goes live. That's the chain.

A quick word on money, because this is a turning point. The domain is the first thing you actually buy, and from here real building costs a little. Not much, but this is the line where you stop setting things up and start building something that's truly yours. That's not a catch, it's the whole point. Owning a ten-dollar domain beats renting your presence on someone else's platform every time.

If you want to do it, I made the whole thing a copy-paste prompt. Go to https://t.co/WMHLB0soZw, copy the domain-setup prompt, paste it into Claude, and it walks you through buying the domain and pointing it at your foundation. New here? Start with Kit 1 (terminal setup) first, so Claude can actually run.

Next part: the small server your site actually lives on. Then we put something online.

[Cliff Marquez @cliff_marquez]

Hope you all get some good use out of it. If you build your own, let me know how it works out for you.

[Cliff Marquez @cliff_marquez]

The short version:

Build your own system-level voice dictation.

Use Deepgram for the transcription.

Then hand Claude the setup instructions and let it build the install for your machine.

You'll need a Deepgram API key — search for their free credit when you sign up. It lasts a long while at this kind of use.

The goal is simple: press one key anywhere, talk instead of type, and get your thoughts into the AI before they shrink.

—

Part One went out on the 17th.

This one took longer than I wanted.

I'm working on getting the cadence tighter — but this is Part Two, and honestly, it might be the one I'd set up first.

In Part One, we set up the terminal — getting Claude running with the permissions and the flow so it can actually do work without you babysitting every keystroke.

This one is smaller and quieter.

But it matters more than it sounds.

When you work with AI all day, the bottleneck stops being thinking and starts being typing.

You already have the instruction in your head.

The nuance.

The context.

The messy middle of it.

And then you sit there pecking it out one key at a time, and half of it falls away before it ever reaches the screen.

Talking is three to four times faster than typing.

But the real win isn't speed.

It's that talking moves at the speed your thoughts actually move.

When I can say the whole paragraph out loud and watch it land as text, I hand the AI the full picture instead of the compressed, lazy version I would have bothered to type.

For a while, I used the dictation built into the operating system.

It was fine for a sentence.

But it was walled into certain apps, it choked on product names and tools I use every day, and the accuracy wasn't good enough.

So I went down a level — to the layer actual products are built on.

That layer is Deepgram.

Deepgram is a speech-to-text engine.

You stream audio to it and get text back in real time with very high accuracy.

A lot of transcription tools are built on top of engines like this. You usually don't touch that layer directly.

But you can.

You sign up, get an API key, and now you have that transcription quality working for you.

Except it's yours.

It works everywhere.

And you control it.

"At the system level" is the part that matters.

This isn't dictation locked inside one app.

It's a key you press anywhere.

Your terminal.

Your browser.

Your notes.

A text box on some random website.

You talk, and the words appear wherever your cursor is.

One tool.

Every app.

The shape of it is simple.

You press a key.

Your mic streams to Deepgram.

The text it sends back gets pasted wherever you're focused.

Then you press the key again to stop.

Or it stops itself after a stretch of silence, so it isn't sitting there listening to — and billing for — an empty room.

A few things separate "it works" from "I actually trust it."

These are the parts I would not skip.

First, a vocabulary list.

Every speech model invents words for names it doesn't know.

Deepgram lets you feed it a short list of the proper nouns you say constantly — your company, your tools, the people you work with — and it stops guessing.

That single thing took my transcripts from "mostly right" to "I don't have to re-read every line."

Second, a silence cutoff.

Streaming costs money per minute, so have it shut off when you go quiet.

Set it short.

Thirty seconds is enough.

A forgotten session should not drain your credits.

Third, mic gain.

This cost me an afternoon.

If your mic is boosted too hot, it clips.

And clipped audio turns into garbage that looks like the model got dumber.

It didn't.

The audio did.

Fourth, a little indicator.

A dot that turns green when it's listening and red when it's not.

Simple, but important.

You should always know when the mic is hot.

Now, I run all of this on Linux, which is the least common of the three setups.

So I'm not going to pretend my exact scripts will drop straight onto your machine.

But here's what makes this era different:

You don't build it by hand.

The same architecture works on Mac and Windows, and you set it up the way you set up everything now.

You hand Claude the spec and tell it to build it for your machine.

So that's what I made for you.

There's a setup file that goes with this — a plain spec you can hand straight to Claude.

Give it to Claude.

Answer its questions.

You'll need a Deepgram key, and the free credit lasts a long while when you're using it this way.

A few minutes later, you'll have a key you can press in any app to talk instead of type.

It's a small thing.

That's the point.

Fundamentals usually are.

You set it up once, it disappears into the background, and the distance between a thought and the AI acting on it gets a lot shorter.

That's the whole game.

Every fundamental in this series is one more piece of friction pulled out from between you and the work.

This is the one I'd pull first.

[Cliff Marquez @cliff_marquez]

The setup file is free on the kit page. No opt-in, no email. Grab and go: https://t.co/ADm8PolBtp

Copy it, hand it to Claude, and it builds the whole thing for your machine. (Kit 1 — the terminal setup — is right there too.)

[Cliff Marquez @cliff_marquez]

When Anthropic dropped their new model Fable, the first thing that came to mind wasn't benchmarks or pricing.

It was Robert Miles.

For the young whippersnappers: Robert Miles was a trance producer in the mid-90s. His track "Children" was everywhere. But the album was Dreamland. And track 3 was called "Fable."

That era of trance was something else. I was in Miami during the peak of it. George Acosta was spinning. Tiësto was breaking through. Paul Oakenfold. Paul van Dyk. Every night was a pilgrimage.

And the music? It was so deep and emotional. Words couldn't begin to describe it. It was melody and feeling and connection all at once. Something you felt in your chest before your head could even catch up.

The unity and friendships I made during that time were incredible. People I still think about today.

Robert Miles passed away in 2017 at 47. Cancer. Rest in peace, man. If you've never heard Dreamland start to finish, give it a shot. It holds up. "Children" was the hit, but the whole album is worth your time.

Check out the jam. Tell me what you think.

You'll probably end up playing this in your head every time you launch Fable from now on.

https://t.co/z9NsBZLIJc

[Cliff Marquez @cliff_marquez]

The Short Version

Cloudflare isn't just your security layer anymore. It's becoming your one control center. Part Three set up the foundation: the account, the tunnel to your server, the locked doors. Now we add the next layer in the same place. Where your files live. Your CDN.

It's called R2. Think of it as a hard drive in the cloud that both your websites and your AI can reach. You give Claude one key, set once, that doesn't expire, and it manages your files for you.

This isn't about ditching Google Drive. Keep it. This is a second kind of storage, built for a code-first way of working, for the stuff you want Claude to operate on directly.

It's perfect for hosting your images. Drop them in, serve them from https://t.co/wA9v0LzQzd, and they load fast anywhere.

And yes, this costs money. Building does. But you're already paying to work this way, and what this piece buys you is worth far more than the price.

Full breakdown below. Copy-paste prompt at the bottom.

———

Part Four. One place for everything, now including your files.

I want to zoom out for a second, because this is where the picture comes together.

The whole point of what we're building is one control center. Instead of your stuff scattered across ten different accounts and dashboards, it all lives in one place, Cloudflare, and your AI has the keys to it.

In Part Three we set up the foundation there. The security layer. Your Cloudflare account as the control panel, the tunnel that lets you reach your server with nothing exposed, and the locked doors on your private dashboards. That's the "keep it safe and reachable" layer.

This part adds the next one, in the same account. Where your files actually live. Your CDN.

Let me explain it simply, because you don't need to be technical for this. That's the whole idea. You hand the technical part to Claude.

It's called R2. Forget the jargon and think of it as a hard drive in the cloud. A place to keep your files: your images, your downloads, your backups. Two things make it special. First, it lives in the same Cloudflare account as everything else, so it's one more thing in your one control center instead of another separate silo to manage. Second, it's built to be reached by code, which means it's built to be reached by Claude.

One honest note so you don't overspend, because this part confused me at first. The technical name for this kind of storage is object storage. Amazon built the first big one, called S3, and it became the standard everyone else copied. R2 is Cloudflare's version of that same idea, and it speaks the same S3 language, so all the same tools work with it. I use R2 because it sits in the same Cloudflare account as everything else, and it doesn't charge you to pull your files back out. But here's the part that matters for your wallet: if you already have an S3 bucket, or storage from someone else that speaks S3, you do not need to go get R2 on top of it. Point Claude at what you already have. R2 is just the one I would pick if you're starting fresh.

Now let me be clear about something, because I don't want you to hear this wrong. This is not "quit Google Drive." Keep your Drive. It's genuinely great at what it's for: your everyday documents, organizing things visually, sharing with people, and its own tools like NotebookLM and Google's AI, which are excellent. None of that is going anywhere, and I'm not here to talk down any of it.

R2 is a different tool for a different job. Google Drive is built for a person clicking around in a browser. That's perfect for you as a human, and a little fiddly when you want an always-on AI to reach in and work. R2 is built for code, so Claude connects once with a permanent key and just works, quietly, in the background. Drive is for you. R2 is for your agent. You'll use both, and that's exactly right.

Here's how you actually use R2. You give Claude one key, and from then on it handles your files. "Put these images in my storage." "Back up this folder." "Pull that file down." It just does it. No login screen in the way, no connection to keep re-fixing.

The clearest example is images. Say you're building a site and you've got a bunch of photos. If you pile them onto your web server, your site gets slow and clunky. Instead, you drop them into R2 and serve them from something like https://t.co/wA9v0LzQzd. Now they load fast from anywhere in the world, and they don't slow your site down. That's what a CDN is, and you just got one by turning on this layer.

One honest aside, because I use it myself. There's a tool called Cloudinary that does a different, narrower job. It automatically resizes and optimizes your images on the fly so your pages stay fast. It's good, and I use it. But it isn't the same thing as R2. Cloudinary transforms images, R2 stores and serves your files. You don't need it to start. R2 gets your images hosted and fast today, and I'll do a whole part later on image optimization. One layer at a time.

Not everything is public, though, and that matters. Your images, sure. Those are meant for the world, so they go on your public cdn address. But your backups and private files stay locked. When you need to hand one to a specific person, you send a link that expires on its own. Claude sets all of that up. You just tell it what's public and what's private.

Now step back and look at what you've got. Your security, and now your files, both in one Cloudflare account, both controllable by your AI with keys that don't expire. One place. One brain running it.

Now the real-talk part, because it matters. This costs money. If you're serious about building, that's not a surprise, it's the deal. You're already paying to work this way, and every real piece you add has a price. This one happens to be small and predictable, but the price was never the point. The point is what it gives you. Your data under your control. Your images served fast. Your AI running all of it without you in the loop. You can stand up things today, by yourself, that used to take a team or a budget you didn't have. That kind of leverage is the best money I spend, and I spend it gladly.

If you want to do it, I made the whole thing a copy-paste prompt. Go to https://t.co/WMHLB0soZw, copy the storage-setup prompt, paste it into Claude, and it sets up your storage, your key, and the connection so your AI runs your files. New here? Start with Kit 1 (terminal setup) first, so Claude can actually run.

Next part: connecting your own domain, so everything you've built gets a real address. Then the small server your site runs on, and we put something online.

[Cliff Marquez @cliff_marquez]

For some it's early, but for me it's late.

The night's almost done. And before it goes, I step back and look at things.

All this chaos. All this building. This strange, wonderful moment we're all living through. I take a breath and just feel grateful for the chance to be in it at all.

There's a lot out there. A lot of noise, a lot of things pulling at you.

People will tell you it's all fighting for your attention. I don't see it that way. Some of it deserves your attention, some of it doesn't. You're the one who decides which is which.

The one thing you can count on is that things change.

Sometimes you're flying high, unstoppable. Other times it feels like everyone's passing you by and you're at the bottom.

But regardless of the moment, one thing should stay constant. Your intention.

And how do you find that intention? How do you guide it?

You look within. Not out. Because most of what's out there won't help you find the answer anyway.

Business isn't going to change. It's going to shift.

The ones still standing years from now? They're the ones who moved with real intention. The intention to give back. To do good. To actually help the people in front of them. And maybe that's where the next breakthrough comes from. The next leap in technology, in science. Because they cared enough about their society to build something that actually matters.

As the tools get better, you'll see more fakes come to light. More people promising to be something they're not.

That's the gift, honestly. It gives you room to be humble. To be real. To hold your ground when others fold.

A soft heart, but a firm fist.

So the question I keep sitting with is intention.

Are you just chasing a dollar? Desperate for the next thing?

We all want things. There's nothing wrong with that.

But watch what happens when it's only about you. With no regard, no care for anyone else. Life has a way of balancing that back out.

The more you want something, the more it seems to run. Let it go, and it comes back in abundance.

The journey is your own, even when what we're building is bigger than us.

I'm no one. Just a stranger on the internet.

So I'll leave you with this:

What's your intention?

You don't have to tell me. Just be honest with yourself.

The answer's already within.

P.S. Tomorrow I'll drop the next lesson on setting up your infrastructure to work with less friction. We've been going through how I set up my Claude setup. Today was Cloudflare, tomorrow we're covering R2 and S3 spaces. See you then.

[Anthropic @AnthropicAI]

We’ve received notice that the Department of Commerce has lifted export controls on Claude Fable 5 and Mythos 5.

We'll begin restoring access tomorrow, and will share an update soon.

We’re grateful to our users for their patience, and to everyone who worked with us on redeploying the models.

[Cliff Marquez @cliff_marquez]

Spent the morning thinking about where the jobs are actually going.

Not the fear stuff. The real pattern.

The AI apocalypse narrative gets clicks. I get it. But most of it is noise.

Here's what I actually see happening.

Low-skill service jobs will get displaced. That part is true. If the work is repetitive, rule-based, and doesn't require human judgment — the machine will do it cheaper and more accurately. Call centers. Data entry. Basic scheduling. The bottom rung of a lot of industries.

But here's the part that gets skipped.

Someone still has to manage the AI. Wire it into the business. Know what to do when it breaks. That person — the operator who understands both the tool and the business — is going to be in very high demand.

I think the real play is fractional AI contracting. Not building a bot and selling it to a dentist office and praying it doesn't break. Actually managing the AI for them. Getting into their client process. Understanding their operations. Being the person who makes the machine work and fixes it when it doesn't.

The businesses that don't have time to learn AI but desperately need the efficiency — they'll pay for that. Not for software. For results.

That's one lane.

Here are a few others I'm genuinely bullish on.

Trades. Plumbing. Electrical. HVAC. Construction. Landscaping.

Trade school enrollment has been declining for decades while demand has done nothing but grow. The average plumber is in their mid-fifties. When they retire, someone has to replace them. AI can optimize a plumbing business's scheduling and marketing. It cannot snake a drain.

I think we're about to see a boom in skilled trades that makes the last decade look quiet. Young people who skipped college for a trade, running lean AI-assisted businesses, are going to clean up.

Licensed professions. This is my favorite one.

Real estate agents. Mortgage brokers. Insurance agents. CPAs. Attorneys. Contractors.

Anything that requires a license to operate. Anything where the government says you need a stamp.

AI doesn't replace the license. It supercharges the person who holds it.

Imagine you get your insurance license. Your AI researches the market for you. Builds your lead lists. Handles follow-up sequences. Qualifies prospects. You get on one final call, close the deal, and the AI does the rest.

One person. One license. AI doing the repetitive work. The human doing the human work.

Same playbook works for mortgages. Real estate. Tax preparation. Legal document review. Any profession where the bottleneck is time and process, not skill.

The agencies that don't offer real value — generic marketing, templated content, no actual expertise — those are going to fade. The operators who layer AI on top of a real license and real process will flourish.

E-commerce stays interesting too. Not dropshipping junk. Real, creative e-commerce — products people actually appreciate, built by people who understand their customer. The fulfillment gets automated. The creativity doesn't.

The one thing I'm not bullish on is physical retail. Too much overhead. Too much real estate. Too much working capital tied up in inventory that might not move. There are exceptions. But as a default, I'd rather sell online and keep the margin.

Here's the honest part.

The coming shift isn't AI taking everyone's job. It's AI making it very clear whose job had real value and whose didn't.

If your value was showing up and executing a repeatable process — you're in trouble.

If your value is judgment, relationships, creativity, or a license the state won't let a machine hold — you're in a better position than you've ever been.

Position accordingly.

What industries are you betting on?

[Cliff Marquez @cliff_marquez]

THE SHORT VERSION (skip the rest if you just want the gist):

Put every domain you own in one Cloudflare account. That's your control panel.

Run a tunnel so your server reaches OUT to Cloudflare instead of exposing ports to the world. Now you can close the ports — and even SSH in through the tunnel with nothing open.

Use Access (a login gate) plus a service token to control who gets through the doors you lock — including your AI — with a one-click kill switch.

Set it up once, scope one key, and the CDN, email, and your next sites all click into the same foundation later.

Full breakdown below. Copy-paste prompt at the bottom.

—

Part Three. Time to lay the foundation.

The first two were about your machine — launching Claude in a keystroke, color-coding your terminals, talking instead of typing.

This one is different.

This is the layer everything else sits on. And it's the one I'd get right before you build anything you want the world to see.

Quick reminder of how I think about all this: I'm always hunting for friction. Where does the work slow down? Where am I repeating myself? Where's the thing that should be one sentence but turns into twenty minutes of clicking around some dashboard?

For everything that lives online — my sites, my servers, my files — almost all of that friction came down to one tool.

Cloudflare.

Not because it's exciting. Because it's the single front door and the single control panel for all of it. One account. Every domain. Every server. And because it's all controllable by API, my AI can run it for me.

Let me break down the pieces that matter.

First, the boring one that makes everything else work: put every domain you own into one Cloudflare account.

That's your control plane.

Once your domains live there, "point this one at my new app" is a sentence to Claude — not a hunt through a registrar. One place. One set of keys. One brain that can see all of it.

Now the piece I really want you to understand. The tunnel.

Normally, to put an app on the internet, you open a port on your server and expose it to the world. And the moment you do, you're a target — bots scan that port all day looking for a way in.

A Cloudflare Tunnel flips it around.

You run a tiny program on your server. Instead of the world reaching in, your server reaches out — it dials Cloudflare and holds the line open.

Traffic comes in over that line. Your app only ever answers on localhost. And now you can close the ports. There's nothing left exposed to scan.

One tunnel can carry dozens of sites on a single box — each domain pointed at its own little local door.

Here's the part that got me: this works for SSH too. My servers don't have an open SSH port — it's firewalled shut. I get in through the same tunnel. There's no front door to pick, because I never installed one.

Now — one thing I want to be precise about, because I had to get it straight myself.

The tunnel is the road. It carries my public sites in, and it carries the SSH I use to push my sites live. But a road isn't a lock.

The lock is a separate Cloudflare feature called Access.

Access is a login gate that sits in front of a site before any traffic reaches it. I put it on the private stuff — my admin panels, my dashboards, the things only I should see — so they're not just sitting behind a URL someone could stumble onto. On my main agent server, I even put it in front of SSH.

And for letting my AI in, Access gives you a service token — a key made for a non-human. That's what my agent uses to get through the gate. Not a password sitting on a server. A scoped key I can delete in one click if anything ever feels off.

So the two work together, but they're not the same thing. The tunnel moves the traffic. Access decides who's allowed through the doors I choose to lock. Worth keeping straight.

So here's the actual move for today:

Create the Cloudflare account. Move your DNS in. Stand up a tunnel. And scope one key for your AI — wide enough to cover what's coming, locked to only your account.

That's the foundation.

And here's the nice part — everything else in this series clicks into this same account. The CDN for your images. Your email records. The next ten sites. You don't rebuild any of it. You flip each piece on when we get to it, and because the key's already scoped, it just works.

Set it up once. Let the rest pop in.

I learned this the slow way — a different setup for every project, ports open I'd forgotten about, keys scattered everywhere. Pulling it all into one foundation is the thing I wish I'd done first.

Now, the honest part.

I'm not a security engineer. I'm an operator, learning as I go, and I'm sure people who do this for a living would do some of it differently. So take this as what's working for me, not gospel.

My bias is speed over friction. Given the choice, I'd rather hand my AI a scoped key and let it do the work than slow myself down locking every door twice. I'm not careless about it — I keep the keys narrow, I keep the one-click kill switch, I lock things to specific IPs where it counts, and I cap what a key is allowed to spend. But when there's a tradeoff, I take less friction, so I can stay focused on what actually moves the needle: shipping.

And that tradeoff is the whole point. This is how I keep a handful of businesses running without a big team behind me — e-commerce stores, affiliate sites, landing pages. It's let me put up more in a month than I used to manage in a year. It won't be the right call for everyone. It's the right call for me.

This isn't the exciting part. The foundation never is.

But it's what makes everything after it calm instead of chaotic.

If you want to set this up yourself, I put the whole thing as a prompt you can copy — paste it into Claude and it walks you through it for your machine. Free, no email, like always:

https://t.co/WMHLB0soZw

One thing first, if you're newer to this: start with Kit 1 on that same page. It sets up your terminal so Claude runs without stopping to ask permission at every step. That's what makes handing it the keys actually pay off — otherwise you're babysitting clicks instead of building.

Next part: buying a domain cheap, and handing your AI the keys to the registrar too — so even that happens by talking.

[Cliff Marquez @cliff_marquez]

There's a specific kind of tired that comes from trying to keep up with something that refuses to slow down.

Every day there's a new model, a new tool, a new post from someone who supposedly built in a weekend what's taken you months.

And you do the math in your head.

Not on the work.

On yourself.

How far behind am I?

I've felt that all year. I still catch myself doing the math some mornings.

What I've slowly come to see is that the overwhelm was never really about how fast things are moving.

It was two quieter things.

I had no filter for what actually deserved my attention.

And I kept measuring my messy middle against everyone else's highlight.

That second one is the quiet thief.

What we see online isn't the work.

It's the trophy shot.

Nobody posts the months of wiring, the dead ends, the version that didn't ship.

So we hold our behind-the-scenes up against their highlight reel and decide we're slow.

We're not slow.

We're just watching the wrong tape.

Here's what actually steadied me:

Underneath all the new, the fundamentals haven't moved an inch.

There's an old line:

Give me six hours to chop down a tree and I'll spend the first four sharpening the axe.

The tools in front of us are brand new.

The discipline that makes them work is ancient.

So I stopped trying to be everywhere.

I picked one thing to make sharper.

The agent I'm building.

The system that catches what matters, sorts it, and hands it to me clean.

Most of my time goes into that blade.

Not wider.

Sharper.

The rest is the part only I can do:

The judgment.

The call.

The actual move.

And because the blade is sharp, that part is light.

The foundation doesn't fight me.

It hands me the thing, and I say yes or no.

The heaviest weight I carry on a hard day isn't the workload.

It's the quiet story that I should be further along by now.

That one's self-imposed.

And it's the only shackle I can actually take off.

The opportunity in front of all of us is real.

And it isn't going anywhere.

You don't have to catch every wave.

You just have to be sharp enough to ride the one that's yours.

Build the foundation.

Sharpen the one tool.

Stop racing a highlight reel that was never the whole story.

What's the one thing you keep meaning to sharpen?

And what would it free up if you finally did?

[Cliff Marquez @cliff_marquez]

Anthropic dropped something new on June 23rd called Claude Tag.

It's an always-on Claude that lives inside Slack, and the internet is treating it like the second coming. I read the coverage. I read the docs. And I want to give you the version that actually matters — not the hype, not the enterprise pitch, just what it is and who it's for.

The short version

Claude Tag lets you tag @Claude in a Slack channel, give it a task, and watch it work through the thread in public. Everyone in the channel sees it. Anyone can pick up where the last person left off.

It remembers context across days. It can pull information from other channels if you let it. And in "ambient mode," it checks in on its own — flags stalled threads, surfaces relevant updates, follows up when something falls through the cracks.

Slack's GM called it "multiplayer AI." That's not just marketing — it's the actual shift. The work moves between people and Claude without anyone re-explaining context.

Sounds great. And it is. But here's what most coverage skipped.

Who this is actually for

Claude Tag requires a Team or Enterprise plan.

Not Pro. Not individual. The whole thing — channel tagging, shared org identity, admin-controlled tools, per-channel scoping, ambient mode — is gated behind an org subscription.

The old "Claude in Slack" integration that worked with personal accounts is being retired on August 3, 2026. If you're on a Pro plan using Claude in Slack today, you'll lose it in about five weeks.

Anthropic hasn't clarified what individual users get after that date. The docs say Claude Tag is "available on Team and Enterprise plans." Full stop.

So if you're like me — individual subscriber, running your own setups — Claude Tag probably isn't for you. Not yet. Not in its current form.

Why the architecture still matters

Even if you can't use it, the design is worth understanding.

Old Slack AI integrations worked under your personal identity. Your permissions. Your billing. When you logged off, the AI stopped.

Claude Tag flips that. It works under an org-level identity. The admin sets what it can access per channel. The work keeps going whether you're online or not. If someone leaves the company, the channel's Claude doesn't lose its memory.

That is not a chat bot with a Slack integration. It's an AI that belongs to the team.

This is where the industry is heading. Even if you can't deploy it today, it's worth understanding why it's built this way.

What it can actually do

For teams that can use it:

- Draft and review PRs from Slack threads
- Monitor deployments and post status updates
- Answer questions by searching docs, wikis, and connected repos
- Summarize long threads for people who join late
- Schedule follow-ups and check in when a task stalls
- Flag cross-team dependencies before they become blockers

Anthropic's own product team used an internal Claude Tag build to generate 65% of their own code — including parts of the tool itself. Dogfooding numbers are always suspicious, but that one is real.

What it costs

Channel usage is billed to the org. DMs are billed to your personal account. Admins set per-channel spend limits and get alerts at 75% and 95%. No surprises if it's set up right.

The old Claude in Slack retires August 3rd. If your team is using it, the clock is ticking.

The part I like

Anthropic didn't sell "AI replaces your team." They sold "here's someone who shows up where you already work, remembers context, and doesn't disappear."

They made admins the gatekeepers instead of individual users. They scoped tool access per channel. They built spend controls.

Those are the decisions you make when you're shipping something real.

My take

Claude Tag is the right idea. Shared AI that lives in the channel, works for the team, doesn't depend on any one person being online — that's where this is going.

But right now it's built for orgs. If you're a Team or Enterprise customer, it's worth a serious look. If you're an individual subscriber, check whether you're affected by the August 3rd cutoff — because the old Claude in Slack is going away, and it's not clear what replaces it for Pro users.

For my own work, I'm not switching. I run Hermes as my personal agent, and Claude Tag fills a different role — organizational, not personal. A truck and a motorcycle both get you places, but you don't pick between them. You pick the right one for the trip.

If you're setting this up

A few things from someone who's been running agents in Slack for months:

- Start on-demand. Ambient mode is powerful but noisy if you're not selective about which channels get it.
- Review channel history before letting Claude read it. Old messages become its context.
- Scope tools per channel. A support channel should not have deploy access.
- Branch protections on AI-generated PRs are not optional.
- The migration deadline is August 3rd. Don't wait.

That's the real picture. Not the marketing. Not the breathless coverage. Just what Claude Tag is, who it's for, and what it means.

[Cliff Marquez @cliff_marquez]

I rebuilt a reporting tool with an AI agent last week instead of adding another subscription.

That got me thinking about what this means for the software we all pay for every month.

I believe the monthly recurring model is going to slow down or shift. With AI, people can build their own back-ends and dashboards that plug straight into their systems. One-time purchases with optional support start to make more sense again.

The big established players will not disappear fast. Airtable, Shopify, Notion, and WordPress already have users locked into their workflows. Migration takes real effort, and most teams will stick with what works until the gap gets too wide.

Still, their growth will probably stall. If I were running one of those companies, I would start offering core pieces at a one-time price while demand is still there. Let people buy the parts they need and wire them into whatever they build themselves.

Smaller tools look more exposed. ActiveCampaign and even QuickBooks feel like they could lose ground quicker. Accounting needs to be reliable, but once AI can handle the math with enough consistency, the monthly lock-in loses its edge.

The giants with real resources might do better by moving into infrastructure. Data centers, hardware, the actual platforms that AI-built tools run on. Or they could sell access to the proprietary layers at a flat cost and let others build on top.

This is just my read on where things are headed. It could play out differently.

What software are you still paying for monthly that you would buy outright tomorrow if the option existed?

[Cliff Marquez @cliff_marquez]

I went looking for a cleaner answer on AI and jobs.

The internet version is usually simple:

AI kills jobs.

The data is not that clean.

PwC's 2026 AI Jobs Barometer found that AI-specialist job postings rose 68.9% from 2024 to 2025, while total job growth rose 8.6%.

That does not look like a market only collapsing.

It looks like a market creating selective demand in new places.

SHRM's 2026 AI research points in the same direction. Among organizations using AI, only 7% of HR professionals reported AI-related layoffs.

But 57% reported new upskilling or reskilling.
39% reported shifts in worker responsibilities.
24% reported new roles being created.

That feels closer to the real story.

Not "AI replaces everyone."

More like:

AI changes the shape of the job.
Then the labor market argues with itself for a while.

And Stanford's 2026 AI Index shows the confidence gap pretty clearly: 73% of AI experts expect AI to have a positive impact on how people do their jobs, compared with only 23% of the public.

That gap matters.

Because if you are a worker, especially early-career, this does not feel like a clean productivity story. It feels like the rules changed while you were still learning the old ones.

So I would not describe this as painless.

The entry-level squeeze is real.
The uncertainty is real.
The anxiety is rational.

But the better read is not broad jobs collapse.

It is labor-market reorganization.

Specialists are being pulled forward.
Existing roles are being rewritten.
And workers are being asked to reskill faster than most companies know how to train.

AI is not just deleting work.

It is rearranging work.

Which is less dramatic than the panic version...

but probably more useful if you are trying to make decisions.

[Cliff Marquez @cliff_marquez]

This weekend I gave my AI agent a home on a VPS.

Now I can talk to it from Slack.
It can run commands, draft posts, and read through my knowledge base.

Not perfect.
Not magic.

But definitely the first time it felt less like a chatbot...
and more like an operator.

[Cliff Marquez @cliff_marquez]

A little reflection for whoever needs it today.

It’s easy to get stuck in the rat race.

Always looking outward for the next solution.

The next tool.

The next answer.

The next “perfect” way to do something.

But I really believe a lot of the answers are already inside us.

When we slow down and actually think, they start to show up.

Sometimes the answer is simple:

Do more of what’s working.

Quietly stop doing what isn’t.

Build around how you actually operate.

And what a time to be alive for that.

We have so many tools at our fingertips now. So many ways to execute. So many ways to take an idea and turn it into something real.

The real shift, at least for me, is this:

Stop chasing “the best way” and start building systems that flow with who you are.

We’re at a point where your code, your tools, and your workflows can bend to how you think instead of forcing you to bend around them.

That’s powerful.

Take what only you can offer.

Productize it.

Systemize it.

Put it out there.

We get hung up waiting for the one perfect thing to be the end-all-be-all.

Maybe the better move is to be more fluid.

Test. Learn. Adjust. Keep going.

Letdowns don’t have to be failures. They can be lessons.

Successes are lessons with rewards.

And sometimes the hardest part is getting out of our own way.

So keep focused.

Keep close to your core.

Keep building...

[Cliff Marquez @cliff_marquez]

Hey everyone — been a few days.

I’ve been heads-down putting some projects together, and I want to start sharing a little multi-part series:

Small, actually-doable tips to make your AI workflow smoother.

Quick context on me:

I come from the business side — building sites, integrating tools, outsourcing, marketing, and figuring out how to make systems run better.

And I genuinely think what AI is putting in our hands right now is incredible.

I’m a systematic person by nature. I’m always watching for flow and friction.

Where does the process slow down?

Where do I repeat the same thing over and over?

Where does one tiny annoyance break momentum?

What’s changed is that instead of complaining about a pain point, or waiting for someone else to fix it, I can often just solve it myself with AI.

That still amazes me.

So I’m going to start sharing little pieces of how I’ve set things up to run as smoothly as I can.

Hopefully some of it helps you speed up your own workflow too.

First one: how I run my terminal.

I’m in Claude Code almost all day on the $200 plan. I keep Codex around too on the $20 plan for the occasional small job.

Most of my work is orchestrated — one window directing another. Every now and then I’ll hand a smaller piece to Codex, but most of the heavy lifting is Claude.

What the setup in the screenshot does is simple:

It gives me two-letter shortcuts to launch Claude Code, already past the permission prompts.

It also color-codes my terminal sessions so when I’m running two instances — an orchestrator and a worker — they share the same color and never blur together.

That’s the whole trick.

Small, but it clears the kind of friction that used to break my flow.

I’m a big believer in being organized from the get-go, and color is a big part of that.

It works with how we’re wired.

A quick glance tells me which window goes with which. No thinking required.

The screenshot is on Ubuntu, but the same idea works on Mac, Windows, or Linux.

Before:

Opening terminals manually.
Typing the same startup steps.
Double-checking which window was doing what.
Losing small bits of momentum every time I switched context.
After:
Two-letter launch commands.
Permissions handled.
Matching colors for related sessions.
One glance and I know exactly what belongs together.
That’s the kind of tiny improvement that compounds.

If you want it, I put a page on my site where you can just copy the prompt.

No email. Nothing to sign up for.
Paste it into Claude and it builds the setup for you.

Copy it here:
https://t.co/wJqCVbPafE

Or grab the Word doc: https://t.co/M8CjjR4rF2

If you want to follow along as I share more of these, subscribing is there too — but no pressure either way.

Best of luck.

I’ll try to be more consistent with these for you.

[Cliff Marquez @cliff_marquez]

Anthropic’s most advanced AI models, Fable 5 and Mythos 5, just got pulled from customer access after a U.S. government directive.

Here’s what happened in simple terms:

The government cited national security concerns and ordered Anthropic to suspend access to the models by foreign nationals.

Anthropic says the practical effect is that they had to disable the models for all customers, at least for now.

The concern appears to be tied to a reported “jailbreak” involving Fable 5.

But here’s where it gets messy:

According to Anthropic, the demonstrated bypass was narrow and specific.

It was reportedly used to identify a small number of already-known, minor software vulnerabilities.

Anthropic also says other publicly available models can already find those same kinds of vulnerabilities without requiring that bypass.

That raises a much bigger question:

If an AI model can do something that security researchers already do every day…

And if other models can already do the same thing…

Should that capability be enough to shut down access?

Or does banning one model just push the same behavior to other tools that remain available?

This is the tension nobody has fully solved yet.

The same AI capability that helps defenders find security flaws can also help attackers look for them.

The same model that can accelerate research can also create new risk.

So the question isn’t simply:

“Can this be misused?”

Almost every powerful tool can be misused.

The harder question is:

“What standard should be used before a government can restrict access to an AI model?”

Because if the standard is “this model can find vulnerabilities,” then that standard could apply to a lot of modern AI systems.

And if the standard is “someone found a jailbreak,” then almost no frontier model is safe from that argument.

Maybe the government is seeing something Anthropic hasn’t shared publicly.

Maybe the risk is bigger than the company is admitting.

But based on what’s public, this feels like a major precedent.

Not just for Anthropic.

For every AI company trying to release advanced models.

What do you think?

Is this the right move for safety?

Or does it set a standard that makes it almost impossible to build and release powerful AI systems?

https://t.co/pVEtaCdy0e