Let me start with why I changed my mind about Codex
At the beginning of the year, when people recommended Codex to me, I just waved my hand and said, "I don't write code." In my mind, it was just a tool to help programmers type out code – nothing to do with someone like me who spends every day writing PPTs, replying to emails, and making spreadsheets.
But I was wrong. Pretty damn wrong.
From April to June this year, Codex was updating almost every week. By the time I realised it, it was no longer that "developer‑only" AI. It had become something that could read documents, make charts, query data, write reports, and even turn your ideas straight into a webpage.
In plain words, it evolved from "knowing how to write code" to "knowing how to get work done." And those two things are worlds apart.
Let me give you a recent example
Last week, my boss suddenly said he needed a Q2 sales review report by Friday, broken down by region, with data analysis and recommendations for next quarter. In the old days, I'd have to ask the data team for numbers, pull pivot tables in Excel, make slides, write the document – a solid two days of work, with lots of chasing, waiting, and back‑and‑forth revisions.
This time I tried a different route.
I opened Codex, installed the data analysis plugin, and said: "Connect to the sales database, pull Q2 numbers by region, calculate both year‑on‑year and quarter‑on‑quarter changes, sort by decline, and generate charts."
Two minutes later – data was there, charts were ready, and it even added a short analysis: "The East China region saw the biggest drop, mainly due to inventory shortages in Category A…"
Then I tossed all that back to Codex: "Generate a full review report in our company's format, covering performance, root causes, and next‑quarter recommendations."
Fifteen minutes – first draft done. I spent another hour polishing the wording and details, and I was finished.
I submitted it on Thursday afternoon. My boss replied: "That fast?"
That meal money was well spent.
What impressed me most after the updates
The biggest change: it finally understands plain English
The most frustrating thing about using AI before was the gap between what you meant and what it understood. You'd say "help me analyse this" and it'd spit out a bunch of code. You'd say "make a nice chart" and it'd give you an ugly bar graph. You'd go back and forth for ages, and still not be happy.
Now Codex is different. You just talk to it the way you normally would, and it gets it. Say "I want a warm‑feeling family photo album webpage" and it won't give you a developer‑style black‑and‑white interface. Say "this report should sound professional but not stiff" and the tone of the text really matches that.
This might sound like a fluffy improvement, but in practice it's huge. Because I no longer have to "translate" my needs – I just say what I want.
Annotations finally let me stop starting over
This is a tiny feature, but once you use it, you can't go back.
Before, if the AI generated something that was 90% right but had one wrong sentence, what could you do? Either regenerate and hope for the best, or copy everything out and edit manually.
Now you just highlight that one sentence, add an annotation like "make this sound more casual", and it changes only that part. It's like having an intern who writes a draft, and you point to a paragraph and say "rewrite this" – they only change that bit, not tear up the whole thing.
Sites made "turning an idea into something" ridiculously easy
This feature is currently only available for enterprise plans, but I was lucky enough to try it.
I dropped in a product requirements document and said: "Turn this into an internal project dashboard." A few minutes later, a working webpage appeared with task lists, status labels, timelines, and document links.
The whole time, I didn't need to know what a server is, where code goes, or how to deploy. All I had to do was describe what I wanted.
My biggest takeaway: before, there was a mountain between me and "making something". Now there's just a thin sheet of paper.
/goal mode let it work while I slept
This might be the most "intern‑like" feature of all.
You don't have to chat back and forth. You just throw in a complete task: "Refactor the login module, get unit test coverage above 80%, and update the relevant documentation."
Then you go do whatever – sleep, eat, whatever. It works on its own, checks each step, and figures out problems as they come.
The first time I used it, I was a bit uneasy. I woke up in the middle of the night, took a peek, and saw it had already submitted a PR. Next morning, all tests passed.
Honestly, it's a weird feeling – it works while you're awake, and it keeps working while you're asleep.
Let's be real – it's not perfect
First, there's still a learning curve for complete beginners. They've worked hard to lower the bar, but things like installing plugins, configuring data sources, and setting permissions are still confusing for someone with zero technical background. My neighbour could build a static photo album, but if he had to connect to a database and do analysis, he'd be lost.
Second, some features are out of reach for regular users. Sites and the enterprise plugins are only available to business customers – individual ChatGPT Plus members can only look on with envy.
Third, it still makes dumb mistakes sometimes. Once I asked it to change a feature, and it updated all the related parts but missed one linked reference, causing an error. I had to dig around and patch it myself.
Fourth, Mac users beware. The latest version can spike CPU usage to 125%‑200% on some Macs, eating over 8GB of RAM. Worst of all, even after quitting the app, it doesn't release the resources – you have to restart your computer. They're still working on a fix, but it's not fully resolved yet.
So who should buy it, and who shouldn't?
Try it if:
You spend hours every day on repetitive work (compiling data, writing routine reports, digging up information)
You often need to "make something" but don't know how to code (internal tool pages, event pages, dashboards)
You're comfortable leaving the "how" to the AI, and you just focus on saying "what" you want
Hold off if:
You only need auto‑complete while writing code (Copilot does that better)
Your daily tasks are very simple (like replying to emails or renaming files)
You're using an older Mac and are worried about lag and performance
One last honest thought
I spend 20 bucks a month on this "intern". It never sleeps, never asks for time off, and never complains about any task. But I'm also fully aware that it's just a workhorse – it won't make decisions for me, won't take responsibility for me, and won't truly understand what my boss actually wants.
Give it the repetitive, process‑driven, time‑consuming but brain‑dead stuff, so I can save my energy for the things that really need judgment and creativity.
That meal money? Worth it. But don't expect it to eat the meal for you.