VS Code Is My AI Daily Driver. Claude Code and Codex Are My Agents.

This was supposed to be a post about Claude Code.

That would have been true enough to publish, but not true enough to matter.

I do use Claude Code. A lot. I also use Codex. I use both inside VS Code, and I still use the standalone harnesses when the shape of the work calls for it. I used Antigravity for a while too, mostly because multi-model access is useful when you are comparing approaches instead of getting one answer. Gemini is still in my stack, but it is clearly third chair for me right now: helpful in specific moments, rarely the place I start.

The honest version is simpler and more important:

Not because I have become a developer in the identity sense. I have not started waking up excited about semicolons. I use VS Code because it is the place where my real work lives close enough for AI to touch it.

Files. Repos. Markdown. Client notes. Website content. Terminal commands. Git history. Build checks. Project instructions. Drafts. Scripts. Receipts. All in one place.

That is the difference.

The Model Matters Less Than The Workbench

Most people still think about AI as a model choice.

Should I use Claude, ChatGPT, Gemini, Perplexity, Copilot, Codex, or whatever new thing just landed in the feed this morning?

That question matters, but it is not the first question anymore.

The better question is: where does the AI sit in relation to the work?

If the AI is sitting in a browser tab, it can advise. It can rewrite what you paste into it. It can brainstorm. It can summarize. It can be useful.

But it is still outside the work.

You are the integration layer. You copy. You paste. You rename. You upload. You download. You check whether the answer fits the file that actually needs to change. You become the fragile connection between the model and the business.

An AI workbench changes that.

When I work in VS Code, the agent can see the folder. It can inspect the actual files. It can search the repo. It can read the project instructions. It can run commands. It can test the change. It can show me a diff. It can leave a record. It can work inside the same constraints I am working inside.

That does not make it magic. It makes it operational.

And operational is where the value is.

My Actual AI Stack Right Now

My current stack is not elegant enough for a clean affiliate roundup. It is a working stack, which means it is a little messy. If you want the broader picture of the tools I run a business on, that lives in my marketing stack breakdown. This post is narrower: it is about where the AI actually sits.

The standalone Claude Code and Codex harnesses still matter. Sometimes a focused session in a narrower environment is exactly what I want. Sometimes I want an agentic workspace that is not tangled up with whatever else I have open. Sometimes the harness itself has a feature or rhythm that fits the task.

But those are modes, not my center of gravity.

If I need another opinion, I can get one. If I need to change the system, I need the workbench.

Why VS Code Wins For Daily Work

VS Code wins because it is boring in exactly the right way.

It is not trying to be a magical AI portal. It is a workspace.

On the left, I have the files. At the bottom, I have the terminal. In the middle, I have the document or code or content I am working on. I can search across the project. I can preview Markdown. I can compare diffs. I can commit changes. I can run the build.

Then I can put an AI agent in that environment and say, “Help me work here.”

That one word, here, is doing a lot.

Not in an abstract chat. Not in a pasted excerpt. Not in a tab that has no idea what repo I am in or what conventions this project follows.

Here.

This is why VS Code feels stronger than the native harnesses for much of my daily work. The harnesses can be excellent. The models can be excellent. But when the task depends on local context, file structure, existing conventions, project memory, and a reviewable path from idea to change, the environment wins.

It also reduces a surprisingly common failure mode: spatial confusion.

AI agents can be smart and still be wrong about where they are. Humans do this too. I have had sessions where the model was capable, the diagnosis was right, and the work still drifted because the tool was pointed at the wrong folder or only saw one slice of the workspace.

VS Code helps because the workspace is visible. The repo is visible. The path is visible. The wrong-file problem does not disappear, but it gets easier to catch.

The agent is not a floating genius. It is a worker in a room.

I want the room labeled.

The Non-Developer Problem Is Not Code. It Is Confidence.

A lot of marketers see VS Code and immediately think, “That is not for me.”

I understand that reaction. The file tree looks technical. The terminal looks hostile. Git sounds like something you are supposed to learn before you are allowed in the building.

But the practical barrier is smaller than it looks.

Most non-developers do not need to become software engineers to get value from this environment. They need enough fluency to be safe.

They need to know what a folder is. They need to understand that Markdown is just structured text. They need to know that Git is a safety net, not a secret society. They need to know that a terminal command is not inherently dangerous, but some commands absolutely are. They need to know how to ask an agent to explain what it is about to do before it does it.

That is learnable.

And once it clicks, the gap is enormous.

The post you are reading went through that second loop. I wrote it in VS Code, asked an agent to fit it into this site’s content collection, read the diff it produced, ran the build, and only then did it go live. No copy and paste into a CMS. No separate window. The draft, the format, the checks, and the publish all happened in the same room.

You still need judgment. You still need taste. You still need to decide whether the changes are good. The point is not that AI replaces that. The point is the same one I keep making about getting real value out of AI instead of producing slop: the tool absorbs the repetitive work so your attention goes to the parts that actually need a human.

But you are no longer doing the mechanical part like a machine.

That is the point. The human should not be the machine.

AI Training Has To Move Past Prompting

Most AI training is still prompt training.

That is not useless. Better prompts help.

But if prompt training is the whole offer, it leaves the biggest value untouched.

Businesses do not only need better prompts. They need better work environments for AI.

They need project instructions that explain the business. They need permission boundaries. They need file hygiene. They need shared folders and source-of-truth docs. They need review habits. They need “do not publish without approval” rules. They need a way to preserve what was learned instead of losing it in yesterday’s chat.

That is the real implementation problem.

When I think about helping companies use AI now, I am less interested in teaching someone a clever phrase to paste into a chatbot. I am more interested in helping them install the workbench.

That is a very different offer than “learn AI.”

It is closer to: let us build the place where AI can safely do useful work.

Where SYNTAX Fits

This is also where SYNTAX starts to make more sense.

SYNTAX is not just a collection of prompts. If that is all it becomes, it will not be worth much.

The value is the operating system around the work.

Routing. Playbooks. Voice rules. Quality passes. Source checks. Human review. Publishing handoffs. Measurement. Memory. A way for one session to improve the next one.

The model is one part. The harness is one part. The workbench is one part.

The system is what makes them compound.

That is why I am less interested in arguing whether Claude is better than Codex or whether Gemini is underrated. Those questions change constantly.

The more durable question is: do you have a system that can absorb better models when they arrive? That is the whole bet behind how I built SYNTAX: keep the operating system stable, let the models underneath it get better.

If the answer is no, the next model upgrade will give you a temporary thrill and then you will fall back into the same scattered workflow.

If the answer is yes, every model improvement makes the whole operating system more capable.

That is where this gets interesting.

Frequently Asked Questions

The distance between idea and execution is shrinking. The distance between a messy folder and a clean deliverable is shrinking. The distance between “we should update this” and “here is the diff, here are the checks, here is the publish checklist” is shrinking.

That is what I feel every day in VS Code.

Not because it is perfect. It is not.

The agents still make mistakes. They still misunderstand instructions. They still get overeager. Sometimes they need to be stopped. Sometimes the right move is to close the laptop and think like a human for a while.

But even with all of that, the shift is real.

AI stopped being a better autocomplete box for me when it entered the workspace where the real work lives.

That is why VS Code is my daily driver.

Claude Code and Codex are my agents. Gemini is in the room when I need it. Antigravity has a place. The native harnesses have a place.

But the center is the workbench.

And if you are trying to understand where AI is actually going to change business operations, start there.

Not with the model.

With the room you put it in.