The Model Stack.

Three LLMs. Three jobs. One workflow.

Most operators reach for one model and try to make it do everything. ChatGPT writes the email, generates the image, drafts the contract, plans the dinner. The model is a Swiss army knife. The operator is the bottleneck.

The MAAD stack the models.

Three models, three jobs. Each one used for the work it is actually best at. The output of one becomes the input for the next. The operator runs the relay.

This is the workflow we ran live on the first MAAD Monday, building a t-shirt landing page from research to ship in under an hour.

/// THE STACK

Gemini// Research

Not search. Research. Deep web, current, opinionated. Pro mode with thinking enabled. Gemini hunts trends the way a dog hunts squirrels: it reads the entire internet in the background, synthesizes it, hands you what is alive right now versus what was alive last quarter. ChatGPT and Claude do not research like this. Their training cutoffs are stale, their browsing is shallow. Use them for what they are good at. Stop punishing them for what they are not.

ChatGPT// Visual

Image generation, mood boards, concept frames. The aesthetic layer. ChatGPT's image model is the strongest at translating a verbal brief into a usable visual. You hand it the trend (from Gemini), describe the form, get the asset back, iterate. The shirt that came out: a retro-futurism design called “Too Cool For This Timeline”. Chrome typography, wireframe textures, early-2000s nostalgia palette. Three iterations to a design we could put on a shirt.

Claude// Writing, structure, code

The prompt for the landing page. The copy on the page. The system around it. Claude's strength is structure: it builds prompts that build prompts, copy that follows a brief, code that does what you said and not what it imagined. The brief we gave Claude was natural language, not engineered: top-center headline, two-column body below, left column with product info and a checkout button, right column holding the image, supporting copy on why someone would buy. Claude returned a prompt ready to paste into a vibe-coding builder. No prompt-engineering theater. Descriptive language and a clear desired output is the entire technique.

/// TOKEN DISCIPLINE

A rule that will save you money and signal:

Do not burn your best tokens on ideas you will not ship.

Sonnet handles copywriting. Sonnet handles “give me a prompt for a website.” Sonnet handles drafts. Opus is for the build you are actually committed to. Claude Code routes automatically based on the work, but if you are picking manually, default to Sonnet and reach for Opus when the stakes are real.

The math: you have made thousands of ideas in your career. You launched five. Stop burning Opus on the 995.

/// THE SKATEBOARD HEURISTIC

We pasted the Claude-generated prompt into GHL's Vibe Studio. The build started. The spinner went infinite. Nothing came out. This is not a one-off bug. It is a property of the platform.

GHL ships the skateboard model. Wheels first, then bolts, then a board. Then they call it a car. Vibe Studio shipped without the ability to attach products directly to the page. The page can be built. The conversion path cannot. To actually take a payment from a Vibe-Studio-generated page, you build a separate checkout funnel and link to it. The interesting half ships first. The boring half that closes the loop ships later.

This is fine for the operator who needs everything-in-one-place to start. It is not fine for the operator who needs a specific thing to actually work on demand.

If a platform pitches breadth over depth, it is a skateboard.

Use it for what it does well today. Do not bet your business on what it will do later. The MAAD answer: own the stack. Vercel and Resend at the top of the funnel. Specialized tools in the middle. Your own database, your own deployment. No platform dependencies you cannot replace in a weekend.

/// THE RECEIPT

Three things to take into your week:

// 01

Run the relay this week.

Pick one thing on your build list. Open Gemini, ask the broad version of the question (in pro mode, thinking on). Take the output to ChatGPT for any visual. Take both into Claude for the deliverable. Notice how much faster it moves than asking one model to do all three.

// 02

Default to Sonnet. Reach for Opus only when you ship.

Set your editor or chat default to Sonnet. Before you escalate to Opus, ask: am I committed to launching this, or am I exploring? If you are exploring, Sonnet is enough. The best model is wasted on the 995 ideas you will not ship.

// 03

Run the platform audit.

Open a doc. List every platform in your stack. For each one, write the answer to two questions: what would I do tomorrow if their pricing tripled, and what would I do tomorrow if they shut down. If either answer is “I would have nothing,” that is a dependency to start replacing this quarter.

That is the signal.

P.S.The full sixty-minute build is unlocked, no login required: maadsociety.com/preview/mondays-001

▶WATCH THE FIRST MONDAY MAAD MONDAYS · $39.14/MO OPERATOR · $128.57/MO ›

This is Transmission 002. New transmissions broadcast every Friday at 9AM Mountain.