The Best Part of ChatGPT Codex Is Not the Code

The project that changed my opinion

Over the past few weeks, I have been helping my mother improve her e-commerce business, Divino Candles.

Divino Candles is a small-batch home fragrance brand built around hand-poured candles, cotton-braided wicks, considered fragrances, and products designed to look as beautiful in a room as they do while burning. It is a personal business, which means every product, photograph, order, and customer genuinely matters.

The store already had plenty of products. The problem was that they were not being presented or sold like premium products.

ChatGPT helped me work through almost every part of that.

It reviewed the website and identified where the experience felt cluttered or inconsistent. It helped restructure the homepage, improve product descriptions, organize collections, correct product types, evaluate product photographs, and make the material claims more transparent.

It researched competing candle brands, helped us refine Divino’s positioning, suggested conversion improvements, and gave direct feedback when something did not feel premium enough.

It also helped us plan and configure Shopify automations, including abandoned-cart recovery, post-purchase communication, fulfillment-based follow-ups, and requests for customer feedback. It assisted with integrations, checkout questions, email workflows, product listings, and the dozens of smaller decisions that appear when you are operating a real store.

This was not one magical prompt.

It was an ongoing working relationship where ChatGPT retained the context of the brand, understood what we had already changed, and helped move the store forward one decision at a time.

I am not going to pretend ChatGPT deserves all the credit for the orders we received. A business still needs good products, trust, marketing, and execution. But it materially improved how the store looked and operated, and those improvements helped us move from endlessly preparing the store to actually converting customers.

Plugins are useful until the real website gets in the way

The Shopify plugin is usually the best place to begin.

It provides structured access to the store and can handle many tasks more reliably than clicking through menus manually. This is how I prefer AI tools to operate when possible.

OpenAI describes plugins as bundles that can combine reusable instructions, connected services, and specialized tools. That makes them much more useful than the old idea of a plugin being a glorified shortcut. They give ChatGPT an understanding of both the system and the workflow it is expected to follow. OpenAI’s documentation explains how skills and plugins work.

But integrations are not perfect.

I have had situations where the Shopify plugin returned a 404, could not reach a particular page, or simply did not expose the setting we needed. Remote browser sessions can also run into Cloudflare checks or other bot protection.

Previously, that would have been the end of the AI-assisted part of the task.

ChatGPT would explain what I needed to do, and I would take over from there.

Computer use changed that.

When the structured integration cannot complete something, I can ask ChatGPT to use my local browser. It can open Shopify, navigate through the actual interface, find the relevant setting, make the change, and then revisit the flow to verify that the trigger, conditions, and actions were configured correctly.

That fallback is enormously important.

OpenAI’s own guidance says computer use is intended for situations where command-line tools or structured integrations are insufficient. It can visually operate desktop applications, browser interfaces, and workflows that span multiple apps. That is precisely how OpenAI documents Computer Use.

The plugin handles what the system exposes cleanly. Computer use handles what is only available through the interface.

Together, they can complete far more of the actual task.

The interface deserves more credit

The biggest advantage may not be any single model, plugin, or feature.

It is the way ChatGPT brings them together.

I can begin by asking for strategic feedback on the store. ChatGPT can research the market, inspect the live website, use the Shopify integration, switch to Chrome when necessary, help create an image, update a document, and return to the store to verify the result.

I do not have to repeatedly explain the business or reconstruct the objective every time the tool changes.

The plugin controls, visible browser activity, permission requests, progress updates, and ability to intervene make the work feel inspectable. I can see what it is doing, stop it when something looks wrong, and redirect it without abandoning the task.

That is a better interface for AI than a chat window that only returns instructions.

It feels less like asking a chatbot for advice and more like supervising someone who can research, operate, and check their own work.

My trust still has boundaries

This does not mean I blindly trust an AI with everything.

I do not give ChatGPT my passwords.

I do not paste credentials into a conversation.

I also do not sign sensitive accounts into a remote cloud browser. Even if the product includes safeguards, I am personally not comfortable handing a remotely operated browser access to important business accounts.

There are several browser experiences inside ChatGPT, and the distinction matters.

The cloud browser runs remotely and separately from the browser on your computer. OpenAI currently positions it primarily for supported public, signed-out websites. It cannot access your existing tabs, local files, browser extensions, saved passwords, or local signed-in sessions. I find it useful for public research, but I do not use it for administering Divino.

The built-in browser uses a separate browser profile and can support its own sign-in. I still prefer not to sign sensitive business accounts into it unless there is a specific reason.

For Shopify administration, I use ChatGPT’s Chrome integration with my existing local browser. I sign in myself, then allow ChatGPT to work inside that already authenticated session. The AI never needs to know or receive the password.

OpenAI’s Chrome documentation confirms that ChatGPT can work with websites where the user is already signed in and lets the user approve or block access at the website level.

That is a meaningful difference for me.

I am giving it controlled access to a session, not giving it my credentials.

It is still not risk-free. ChatGPT can process page content, screenshots, tool activity, and anything else required for the task. OpenAI itself advises users to avoid exposing secrets or highly sensitive information and to remain present when such information is necessary.

My approach is therefore simple:

I sign in myself. I keep the task narrow. I grant access only to the website required. I keep customer and payment information out of context. I manually review consequential changes involving money, permissions, customer communication, or publication.

I trust it to operate within boundaries. I do not outsource those boundaries to the AI.

This does not replace Claude Code for me

For deep, terminal-centered software work, I still use Claude Code extensively.

It is fast, direct, and extremely comfortable when the task lives primarily inside a repository.

ChatGPT Codex becomes more compelling when the work crosses boundaries.

A real task may begin with code, move into Shopify, require research, involve a Google Drive folder, need a new product image, trigger an email workflow, and finish with someone verifying the live customer experience.

That is where the difference becomes obvious.

I am no longer choosing an AI tool purely by asking which model writes the best code. I am asking which product can follow the complete workflow and help me reach a verified result.

For a growing amount of my work, that answer is ChatGPT.

From telling me what to do to doing it with me

For years, AI assistants were excellent at explaining tasks.

They could tell you where a setting probably existed, provide steps for creating an automation, suggest improvements to a product page, or generate a piece of code.

You still had to carry that advice across every application yourself.

Computer use closes part of that gap.

ChatGPT can now research the problem, recommend a change, apply it through the appropriate integration, fall back to the interface when necessary, and visually verify what happened.

It still requires supervision. It still makes mistakes. There are still areas where I will not grant it access.

But the overall workflow has changed.

I spend less time translating AI advice into manual action. I spend more time reviewing outcomes, correcting decisions, and moving to the next problem.

That is the part of ChatGPT Codex I did not expect to value this much.

The code is excellent.

The ability to follow the rest of the work is what changed how I operate.