I Made You Something

We just wrapped our fifth Claude Cowork workshop over at VCI, and it was outstanding. Same thing happened that happens every cohort — people walked in unsure what Cowork even does and walked out with builds running.

Two more coming in September. Cowork the first week, Claude Code the final week. If you're not in the free VCI community yet, get in there. That's where the dates drop first.

But that's not why I'm writing this part.

While we were teaching that cohort, I ran a short introduction to something called loop engineering — and the response was strong enough that I turned it into a real deliverable and I'm giving it to you free.

👉 Grab the Loop Prompt Writer skill + Install SOP here

Here's what it actually does, because "loop engineering" sounds like another buzzword and I don't want you to skip it.

You already know Claude can run without you. /goal, /loop, scheduled routines, background agents. The problem was never getting a loop started. The problem is stopping it. A loop with a fuzzy finish line either quits early or burns your budget all night doing the wrong thing. Tell an agent "make the tests pass" with no guardrail and you'll come back to a project where it deleted the tests. Technically compliant. Completely useless.

The Loop Prompt Writer is a Claude Code skill that fires before any of that runs and forces the five questions people skip:

• What does DONE actually mean? Something checkable — "npm test exits 0 and the build succeeds," not "the code is good"

• What's the give-up condition? Attempts, time, or budget. Every loop needs one

• What is it allowed to touch? And what must it never touch

• Where does it report back? A PR, a file, a message — where do you look

• What does failure cost? That's what sets the supervision level

Then it picks the right mechanism for you — /goal versus /loop versus routines versus worktrees, with the real limits on each — and writes the stop condition properly. It also builds in the Doer-Verifier pattern: a separate agent grades the work, because an agent checking its own homework will always tell you it's done.

Install is two minutes. Paste one prompt into Claude Code and point it at the folder. The SOP walks the manual path too, if you'd rather see how skills actually work under the hood.

Now the part I actually want you to hear.

Over the next few weeks you're going to get buried in vocabulary. Loop engineering, graph engineering, harnessing, Grokbot, whatever ships next Tuesday. Half of it will be genuinely useful. Half of it will be people putting new names on things you already know how to do.

Don't get caught up in the hype. Everything you've learned to this point still works. None of it just got invalidated. What's changing isn't the fundamentals — it's the techniques the top operators are using to squeeze more output out of the same tools you already have. That's it. That's the whole gap.

My job is to keep closing that gap for you before it costs you money. So we start with loop engineering, and I'll keep translating the rest as it lands.

Grab the SOP. Read it. Install the skill. Then go build something that finishes on its own while you sleep — and actually stops when it should.

See you inside VCI. - https://vcinc.ai/ (Free Community)

And if you want to build with me in person: AI Unleashed Intensive, Cancún, September 30 – October 2.

I went last year. It wrecked me in the best possible way — actually transformational, not the word people use when they mean "nice." All-inclusive, great hotel, better food. This year's lineup includes Molly Mahoney, Austin Armstrong, and a few more names you'll recognize. I'm personally running nightly hackathons plus a full demo day, so you leave with something you built, not a notebook full of notes you'll never open.

Early bird pricing is live right now and it climbs the longer you sit on it.

👉 Get your ticket →

Come build with me.

🛠️ 🚀 HighLevel Just Gave You an Army of Agents (And You Can Rent Them Out)

You know that pile of client work that isn't hard, just endless? The social posts. The onboarding. The "hey can you check if that invoice went out." The stuff that eats four hours a week per client and generates exactly zero dollars.

HighLevel just dropped AI Super Agents, and that pile is now a product you can sell.

This isn't a chatbot. Chatbots wait for someone to talk to them. A Super Agent goes and does things inside the CRM — with access to 400+ actions across the platform. Social planner alone has 61 actions it can fire on its own. And halfway through recording the walkthrough, HighLevel shipped a second update on top of it: you can now schedule these agents to run on their own

Turning it on (2 minutes, and most people won't find it)

It's buried in Labs, so here's the path:

• Agency view → Settings → Labs → Sub-Accounts

• Find AI Agent Studio – Super Agent, then use the bulk action to Enable Access for every client at once

• Hit save, drop into any sub-account → AI Agents → Agent Studio

That's it. You now have agent creation for your whole book of business.

What you can actually build with it

Building an agent is dictation, not configuration. You talk, it writes the whole thing — goals, brand positioning, content pillars, conversation flows, tool permissions. It turned on web search and image generation by itself.

Here's what came out of testing:

• A social media manager that researched angles, wrote the posts, generated the images, and scheduled them to Facebook, Instagram, TikTok and LinkedIn — from one paragraph of dictation

• An onboarding agent triggered by a form submission: client fills out the form, the agent researches their website, and drafts their voice and chat agent setup for them

• A community moderator that keeps the group on topic, looks up contacts, updates tags, and escalates to a human when someone actually needs one

• An invoice checker that watches for late payments and manages invoices without you opening the billing tab

• Scheduled research — set a trigger for 10 AM daily and it goes hunting for content topics before you've had coffee

The triggers stack too: chat started, form submitted, lead tagged, appointment booked, or a straight calendar schedule. And there's no cap. Build twelve of these things, drop them in a folder, let them run.

Two honest notes, because it's beta and it acts like it. First, it will occasionally tell you it did something it didn't do — it confidently claimed it scheduled posts to a social account that wasn't even connected. Verify everything for the first two weeks. Second, feed it a knowledge base before you expect quality. Services, pricing, years in business, tone of voice, the FAQ your team answers on repeat. Ten minutes of knowledge base is the difference between a usable post and generic filler.

The part that should actually get your attention: HighLevel has confirmed these will be addable to snapshots, and memory is coming so agents remember past interactions with each customer. When that lands, you stop building agents one at a time and start deploying an entire agent stack the moment a new client signs.

Which means the agencies setting this up now — while it's still buggy and still in Labs — are the ones with a tested, priced, documented offer ready the day it goes stable. Everyone else will be watching a YouTube video about it in six months, wondering why their onboarding still takes four hours.

Go turn on Labs.

 🤖One Agent That Picks the Cheapest Model That Gets It Right (Work Cheaper and Faster)

Right now you're doing one of two things every time you start a build. Either you pick a model out of habit and hope it's the right call, or you don't pick at all and just pay top rate for whatever's loaded — including for the throwaway tasks.

Both are guesses. And one of them is expensive.

So I built the thing that decides for me: a Claude Code agent and skill wired into OpenRouter, which reads live model leaderboards, matches the model to the task, and routes the job to whatever gets the result for the least money. It refreshes itself weekly, so when a new model lands it's already in the running.

Fable 5 still orchestrates — it's my lead nine times out of ten. It just stopped personally doing every piece of labor on the project.

The build, start to finish

The whole thing is one dictated prompt plus four links. That's the build.

• Tell Claude what you want — a skill and agent that acts as a model selector, reads OpenRouter's documentation and the leaderboards, logs time and cost per model run, and updates itself on a schedule. Dictate it. Don't type it.

• Hand it the sources — openrouter.ai/models, OpenRouter's own rankings board, an outside leaderboard like AI Arena, and OpenRouter's API models file (that last one is already formatted the way an AI wants to read it)

• Get your key — create the OpenRouter account, add credits, spin up a workspace, generate an API key. Set a credit limit on the key while you're there so a runaway loop can't drain the balance

• Drop it in .env — Claude Code writes the placeholder for you when it builds the agent

• Register the weekly refresh — one PowerShell command (run as administrator) creates a scheduled task that re-pulls the leaderboards. On a Mac, tell Claude you're on a Mac and it hands you the Mac version

Two things worth stealing here that have nothing to do with model routing.

First: I didn't describe the models to Claude. I gave it the actual documentation URLs and let it read them. Most people build skills from memory and then wonder why the output is thin. Feed it the source and it builds from facts.

Second: when the PowerShell command threw an error, I copied the error, pasted it back into Claude, and it told me what I'd done wrong in about four seconds. Nine times out of ten it's a copy-paste mistake, not a real problem. Stop debugging things by hand.

Then I ran all four against each other

Same prompt to every model, running in parallel: build a 30-second explainer video on how to install Claude Code, using Remotion. Log the cost. Log the time.

Four finished videos. $1.09 total.

Here's what I expected: the expensive models win, obviously, that's why they're expensive. Here's what actually happened.

Fable 5 did produce the cleanest output — smooth, accurate, sharp graphics. It also cost the most while writing 40% less code than Opus 5. Opus was the slowest of the four at 119 seconds and came in eight cents under Fable. Sol 5.6 landed right in the middle on both and made a design choice on one screen I honestly liked better.

And Kimi 3 — at 6.5x cheaper and 4.5x faster than Fable — turned in a video I would have published. Its one real miss was quoting an older install command.

Line them up and the quality gap is there if you're hunting for it. Nobody hunts for it. Your client definitely doesn't.

Which is the whole point: not "cheap models are secretly better." They're not. It's that the premium model is only worth premium money on the jobs where the difference actually shows up — and most jobs aren't those jobs.

There's a second use for this that has nothing to do with cost. Once you've got every model on tap, you can send a finished build to a different model for a second opinion — a full outside read on your own application, from a system that had no hand in writing it.

Pick the one task you run most. Route it. Check the invoice in thirty days.

 🧠 Claude Started Watermarking Your Copy (And 3 Other Things That Actually Matter)

If you're running AI-written content anywhere in your business — newsletters, blog posts, client deliverables, social — read this part twice.

Anthropic is now embedding machine-readable watermarks into text Claude generates. Not a visible tag. An invisible mark woven into the words themselves, applying across the Claude API, Claude app, Claude Code, Cowork, and Claude Tag, worldwide. It survives copy-paste and can persist through some editing.

They're careful about what it proves: a detected mark signals the content passed through Claude, but it isn't conclusive — and no mark doesn't mean no AI. So it's a signal, not a verdict.

And Anthropic isn't alone. This is a wave:

• Suno is rolling out audio watermarking and fingerprinting, plus download caps starting September 3rd — 20 songs/month on the cheaper plan, 60 on the $24 plan. That background track under your YouTube videos is now a metered resource.

• Spotify starts showing an AI persona badge on artist profiles in mid-September. Get flagged as an AI artist and you're out of editorial and algorithmic recommendations entirely. No playlists, no discovery, unless someone already follows you.

Here's the truth about what this actually means for your business: the era of quietly shipping raw AI output as your own work is closing. Not because of ethics lectures — because of infrastructure. The platforms are building detection into the pipes.

Which is fine, if you were already doing the work. Draft with AI, then rewrite it in your voice, add your numbers, cut what isn't yours. That was always the right process. Now there's a mechanical reason to actually follow it.

xAI shipped an agent app for people who don't want to see code

Grokbot dropped this week, and it's not a coding tool — xAI already bought Cursor for that. This is closer to a Cowork competitor: you spin up named agents and they figure out their own job from the name.

Name one "email triage" and it checks whether it has your Gmail, asks for authorization, and starts sweeping the inbox. Name one "research bot" and it offers you a daily 8:57 AM brief on what moved and why.

Two features worth your attention:

• Teach a task — it records your screen while you walk through a process once, then turns that demonstration into a repeatable skill

• Routines — recurring runs on a schedule, or fired by events like a new Slack message

Each agent also gets its own cloud computer with Chrome, a file manager, and a terminal, plus plugins for Drive, Calendar, Notion, Slack and more — and a mobile app to manage them from your phone while the work happens somewhere else.

Sound familiar? It should. HighLevel's Super Agents, scheduled Claude Code sessions, Grokbot routines — three different companies shipped the same idea in the same month. Agents that run on a clock instead of waiting for you.

The model flood is now a pricing story

Eight model announcements this week, and honestly most were marginal. Two are worth knowing:

• Grok 4.6 — near the top on knowledge-work and legal benchmarks, priced at $2 in / $6 out per million tokens against GPT 5.6 Soul's $5 / $30. Close capability, fraction of the cost. Grok 4.7 is reportedly 3–4 weeks out.

• Gemini 3.7 Flash — 75¢ in / $3.75 out through end of year, landing in roughly Fable 5 territory on intelligence at a fraction of the spend.

Also this week: Meta's Muse Glimmer and Nvidia's Nemotron 3.5 Lightning (both open-weight, both runnable on your own hardware), DeepSeek V4 Pro, and GPT 5.6 Soul with an ultra-fast mode pushing 750 output tokens per second on Cerebras chips — limited testers only, for now.

The pattern under all of it: capability is converging, price is collapsing, and the gap between the flagship and the discount model keeps narrowing. Which is exactly why the model router I built above stops being a nerd project and starts being a line item on your P&L.

Quick hits from Anthropic, since most of you live in Claude Code:

• Auto mode is now the default as of August 14 on Pro, Max and Team plans — it approves routine steps on its own unless something looks harmful. Prompt it, walk away, come back to finished work.

• The Claude Chrome extension now runs on Cowork — it can screenshot every open tab, then keep working on all of it and produce a report inside the Claude app.

• Claude Code sessions can message each other now. One session hands another a summary mid-task instead of you re-explaining context for the fourth time.

None of this week's news changes what you build tomorrow. All of it changes what you build for.

The watermarking, the AI badges, the download caps — those are platforms drawing a line between people using AI as a tool and people using it as a mask. That line is going to keep getting sharper, and being on the right side of it is going to be worth money.

Meanwhile Grok 4.6 at $2 in / $6 out, Gemini 3.7 Flash at 75¢, and open-weight models running on your own GPU mean the cost of intelligence is falling faster than most people are adjusting their pricing. In twelve months, "I use AI" stops being a differentiator entirely — because everyone will. What's left is who built systems around it and who just bought a subscription.

Build the systems.

A Guy Vibe-Coded a Video Game in a Few Hours to Prove Vibe Coding Can't Replace Game Developers

Matt Wolfe spent a weekend building a 3D roguelike about a librarian chasing children who won't stop knocking books off shelves. The point was to settle an argument on X, where somebody keeps posting one-level demos and declaring the entire game industry dead.

He called it The Librarian 2: Look Ma, I Can Read Good.

Here's what came out of it:

• 10,400 lines of code across 33 modules — with zero asset files. Every texture, material, character, sound effect and piece of music generated in code. Built in Claude Code with Opus 5, then handed to Codex with GPT 5.6 Soul for fine-tuning.

• Claude Code tested its own game. It opened the build in a browser, took screenshots, spotted the camera punching through the ceiling, and fixed it. Then rebalanced the difficulty on its own after deciding boss pressure was dominating the chaos meter. It's playtesting itself now.

• The handoff trick is the actual lesson. Before switching models, he had Claude write a full design markdown file explaining the entire game. Codex read it and picked up exactly where Claude left off. Steal that for literally any project.

• Braden the Bully broke the game. A child NPC so fast you cannot catch him — sprinting laps outside the library while the chaos meter climbed to 95%. Then he simply vanished into nothing. The game-over screen reads: "The building has been surrendered to the children."

• A tornado spawned inside the library and knocked over four books. A volcano grew out of the center of another level. A Karen appeared to announce she'd be leaving a review.

• The four "different" locations were the same room in different colors — except the grocery store, where a falling banana makes you slip and a mushroom makes you grow like Mario.

And the verdict, after all of it: game developers are nowhere close to cooked. As the CEO of Take-Two put it, coding was never the bottleneck. It's the thousand tiny judgment calls about what goes in and what stays out — and AI is still terrible at those.

Which honestly tracks. AI wrote 10,400 lines in an afternoon and still couldn't figure out that a tornado should knock over more than four books.

Go watch the whole thing on Matt Wolfe's channel. Braden alone is worth it.

💲Deals at a Glance

🚀 Claude Code Workshop — 4 Days, 8 Hours, $67: Build apps, AI agents, and automations from scratch — hands-on, start to finish. 👉 Register Now

🎓 NEW — The Automated Marketer Skool Community (LIVE NOW): The room where the building doesn't stop. Twice-a-week group calls with me, plus courses on vibe coding anything — HighLevel, Claude Code, Lovable, AI, and automations. Introductory pricing for a limited time, with a lifetime option on the table. Two ways in:

• 👉 Join the Free Community — free access, free course, daily community

• 👉 Go VIP — $47/mo intro — hot-seat support, advanced templates, inner-circle access

🏗️ Agency Build-A-Thon — Free: A complete AI agency built from zero on GoHighLevel — domain, white-label SaaS plans, a client-ready website, voice + chat AI, the whole back end — recorded live, every step in order, nothing re-staged. Your start-to-finish blueprint if you sell to clients (or want to). 👉 Watch Free

🚀 Business Build-A-Thon — Free: Two real businesses built online with AI in one afternoon — domain, website, automations, and the AI that answers the phone, all from a blank screen. For launching your own business, not clients'. 👉 Watch Free

🛠️ Done-For-You — Built by Me and My Team at Rapid Active Marketing: Don't want to build it yourself? Skip the learning curve and have us build the whole thing for you — GoHighLevel setups, websites, funnels, automations, and AI voice + chat widgets — done right, by the people who actually teach this. We take on a handful of clients at a time, on purpose. 👉 See What We Build

💰 The Best GoHighLevel Affiliate Offer on the Planet: Everything you need to win with GHL, in one place. 👉 Get Started

🔥 Vibe Coding Incubator — FREE: Hands-on vibe coding, live weekly trainings, app templates, support, and Lovable how-to's. 👉 Join Free

😲 GoHighLevel Masterclass on Udemy — $20: 5,000+ students, 4.7★, 18+ hours of content. 👉 Grab It

10x the context. Half the time.

Speak your prompts into ChatGPT or Claude and get detailed, paste-ready input that actually gives you useful output. Wispr Flow captures what you'd cut when typing. Free on Mac, Windows, and iPhone.

Try Wispr Flow free

Remember, there’s always something new coming from the Automated Marketer YouTube Channel—new content drops every Tuesday & Thursday. Subscribe so you never miss the next big idea.

Keep watching, keep learning, and keep moving forward!!

Nuno