The AI Builder Technical Stack

There are three kinds of people using AI at work right now. AI consumers open ChatGPT or Claude, ask a question, and copy the answer back. AI workers chain pre-built agents and tools to get more of their current job done. AI builders use general agents to ship custom agents and traditional software (dashboards, automations, internal tools).

The long-term leverage is in becoming an AI builder. To do it you need to know two halves of one stack: the AI side, and the traditional software side. None of this assumes you're learning to code. It's the tools you need to ship things.

I've taught over 100 PMs, marketers, founders, and operators to make this jump. Here's the orientation lecture.

The whole stack on one page

It's a one-time cost. Once you've got these names in your head, you play in the same realm as roughly 30 million developers, with access to the whole infrastructure they've built up. Read the table twice. The rest of this guide just zooms in on each row.

The AI stack

The four pieces that turn a text-prediction model into a working agent. This is the new half of the stack, the one that didn't exist 24 months ago.

Models

Anthropic is not the AI. Claude is not the AI. Those are products built on top of a model. The model itself is just predicting the next word based on what came before, massaged into something that looks like reasoning. The cleanest analogy: a genius with amnesia who knows nothing about the world. You have to feed it context every single time.

Different models have different strengths. The cleanest summary I've seen is Lenny Rachitsky's:

Mapped to the underlying models: Opus 4.7 (the model under Claude Code) is the best at knowledge work and strategic thinking. GPT-5.5 (the model under Codex) is the strongest at heavy coding tasks. Gemini Pro Preview (the model behind Google's tools) is unusually strong at design and image work, though it's falling behind on general intelligence.

You'll mix them. Use the right model for the job, not whichever brand you started with.

Harness

A harness is the thing that turns a model into an agent. It hands the model a toolset (search the internet, read code, write code, read logs) and a set of instructions for how long to keep working and how to verify it did a good job. That feedback cycle is the agentic loop, and you can jump in to interrupt or add context whenever you want.

When people say "Claude Code," they usually mean the harness. The model inside is Sonnet or Opus depending on the day. Switching between Claude Code, Codex, and Cursor is mostly switching harnesses.

Agent orchestrator

Once you have an agent, you need a place to sit in front of it: give it context, watch it work, check in. That's an agent orchestrator. The easiest mental model is a chat thread with agents, with multiple threads going at once.

There's no perfect tool today. Claude Code in the terminal (or embedded in Cursor) is fast and feature-complete. Claude Code Desktop has the prettier interface. Conductor is a nice abstraction on top of Claude Code, but because it's a wrapper it lags behind whenever the underlying tool ships something new. You can even hook agents into something like Linear and just move cards around. Pick the one that matches how you want to work.

We bounce between the terminal and Claude Code Desktop in class. Full deep dive on orchestration here.

Skills

Skills are how you customize the harness. Each one is a markdown document that says "when this kind of task comes up, here are the steps to follow and here are the tools to use." Drop a SKILL.md called "write-prd" in the right folder and the agent suddenly knows how you write PRDs.

A real one I use: a Google Search Console skill. After I publish a new article, I tell Claude to submit it for indexing. The skill is a one-page document with the steps and the script to run. Full setup here.

The traditional software stack

These are the five tools that already existed before AI showed up. AI doesn't replace them, you have to integrate with them. The good news: AI manages most of them so well that you mostly need to know what each one is for, not how to drive it. Once you learn this side, you have it forever.

Code editor

A code editor is where coding used to happen, before AI did most of the typing. Think Microsoft Word for code. Most popular today: VS Code and the apps built on top of it (Cursor is the AI-native one).

Now that AI generates most of the code, you have two real reasons to open an editor:

1. Saving API keys in a special file called .env (more on this in the APIs section).
2. Viewing markdown (.md) files. Markdown is the native language of AI output. Skills, plans, instructions, generated docs, almost all of it lives in markdown.

Git

Git is a small program that runs in the background and saves checkpoints as you build. Think of it like a video game save: if something gets messed up, you can revert to a known-good point.

Three words you need:

• Commit: save a checkpoint locally.
• Push: move your commits up to GitHub.
• Pull: grab someone else's changes from GitHub.

AI manages git better than most engineers. You almost never run a git command yourself, you just say "commit and push" at the end of a session and the agent handles it.

GitHub

GitHub is Google Drive for code. You upload your project to GitHub, and every commit you push goes with it. Two real uses:

• Teams sharing the same code.
• External tools (Claude Code, Vercel, etc.) reading your code so they can do something with it.

When GitHub asks about your "repo" or "repository," that's just a fancy word for a parent folder. All of coding is text files inside folders. GitHub is where you put those folders so other software can reach them.

Host

Once your code is in GitHub, you need somewhere to actually run it. That's a host: Vercel, Cloudflare, Railway, AWS. The host pulls the code from GitHub, runs it on their servers, and gives you a URL so anyone can use the app.

Hosts also handle the boring stuff so you don't have to:

• Spikes of traffic.
• Scheduled automations (cron, like "run this every day at 9am").
• Optimizing images so pages load fast.
• Blocking obvious attackers.

Your host is also the second place your API keys live (the first is your local .env). Whenever you push a new version, the host re-deploys it and gives you a fresh production link.

APIs and MCPs

APIs and MCPs are how your app or your agent connects to other software. The mental model: it's like logging into another app, but through code instead of a UI. Every connection needs either an API key or a sign-in.

An API key is a password. Treat it like one. If your image-generation API key leaks, someone can pretend to be you, run a million image generations, and stick you with the bill. So:

• API keys live in .env files. Never in code that gets pushed to GitHub.
• .gitignore is the file that tells git to skip your .env. Claude Code creates it for you.

MCPs are the newer flavor of the same idea, built specifically for agents. They're a standard way to plug an agent directly into another tool (Stripe, Notion, your database, your inbox). Full MCP guide here.

Terminal

The terminal is a code editor in a different form factor. Cursor and VS Code are visual: you see a file tree and click around. The terminal is text-based: you type a command to see what folder you're in, type another to move into a subfolder, type another to list its contents.

Why bother? Once you've practiced it, the terminal is faster than clicking, and it's easy to automate (edit ten files at once with one command). Models love it too: pure text in, pure text out, no visual ambiguity.

Inside Cursor, the terminal lives as a panel at the bottom of the editor. It's the exact same terminal application as the standalone one, just embedded for convenience. The classic way to use Claude Code is right there: visual file tree on the left, agent running in the terminal panel below.

The progression: from AI consumer to AI builder

Three rungs:

1. AI consumer. Open ChatGPT or Claude, ask a question, copy the answer.
2. AI worker. Use pre-built agents and AI tools to get more of your current job done.
3. AI builder. Use general agents (Claude Code, Codex) plus the software stack above to ship your own agents, automations, and dashboards.

Most knowledge workers are stuck between rung one and two. The leap to rung three is where the leverage starts. Once you can connect every box in the stack, you can build and push real software, agents and traditional apps alike.

What to do this week

1. Install Claude Code and a code editor (Cursor or VS Code).
2. Set up git and GitHub once.
3. Pick one small thing to ship and push it through the whole stack end to end.
4. Add skills, MCPs, and a host on your second project, not your first.

You don't need to read more guides before you start. You need to ship one thing badly and feel the loop click.

Want the in-person version? ClaudeFluent takes a room of operators from "I've never opened a terminal" to a working, hosted app in a single Saturday.