You open ChatGPT, type a question, copy the answer, paste it into an email or a Slack message, then close the tab. If you're anything like most people, you repeat that a dozen times a day.
If that's how you use ChatGPT, you're only touching a fraction of what it can do. Over the past few months, OpenAI pushed "ChatGPT Work" into the spotlight — a real upgrade, but the problem is almost nobody has explained clearly what you're supposed to do differently once you have it. Even OpenAI's own materials don't spell it out well. Most people open ChatGPT Work and keep the same old habit: type, get an answer, close the window — running a new engine in the old gear.
The real difference is easy to miss: ChatGPT Work doesn't just answer questions, it can plug into the tools you already use and actually finish tasks for you. And whether it feels like it "gets you" has nothing to do with how clever your prompts are — it comes down to whether you've taught it to remember your rules.
A Simple Test: Does It Actually Remember What You Told It
Start with the simplest possible example. You ask ChatGPT Work to set a Google Calendar reminder for your brother's birthday — nothing special, regular ChatGPT can do that too.
The next part is where it gets interesting. You add one line: "From now on, every birthday reminder needs a cake emoji in the title, 'remember to buy a gift' in the notes, and this rule should apply every time going forward."
Still not that impressive on its own — plenty of chat tools can follow an instruction you just gave them. The real turn comes in the next step: you open a brand-new conversation and simply say, "Set a reminder — April 22nd is Sam Altman's birthday." You never mention the cake emoji or the gift note again. It applies both rules anyway.
That's the one difference between ChatGPT Work and regular ChatGPT that actually matters: it remembers what you've taught it and carries that into your next conversation — even a completely unrelated one. The more you teach it, the closer it gets to how you actually work. Most people never notice this mechanism exists, so they never get the chance to use it.
Know Your Lineup First, So You Don't Reach for the Wrong Tool
Before you go any further, there's something worth sorting out: OpenAI currently has several products living under similar-sounding names, and it's easy to mix them up.
At one end there's plain ChatGPT — no tools connected, just question and answer. At the other end is Codex, built for engineers writing code, which most office workflows never touch. The confusing middle ground is ChatGPT Work itself, which actually comes in two forms: web and desktop. The web version lives in your browser and has its own strengths, but its capabilities are much narrower than the desktop version. The desktop version is the one that actually ties together connecting tools, remembering rules, and getting things done automatically. If you're not an engineer, the desktop version is the one worth your time.
Once that's sorted, there's only one more thing standing between you and its real power — a rulebook written in plain language.
agents.md: A Rulebook Written for the AI to Read
OpenAI calls this rulebook `agents.md` — at its core, it's just a Markdown text file broken into a few sections: a "purpose" block explaining how the file works, "global rules" that apply to every task, a "voice principles" section that controls tone, and a "maintenance" block that teaches the AI how to keep the file tidy as rules pile up.
That birthday reminder rule from earlier? It got written automatically into the "global rules" section — you never edit the file by hand. Say the rule once in conversation, and ChatGPT Work writes it into the document itself.
Getting the file to actually work takes a few steps, and the order matters: download the template and save it as Markdown, rename it precisely to `agents.md` (not "agent" or "agent1" — variations like that won't be recognized), create a new folder on your desktop and drop the file in, then go back to ChatGPT Work desktop, turn on work mode, and choose "choose project" pointing at that folder. From that moment on, the folder is "mounted," and ChatGPT Work reads this rules file before every task it runs.
Here's the one conclusion everyone should remember: if you open ChatGPT Work without mounting a folder, you're effectively still using regular ChatGPT. Once you mount a folder containing `agents.md`, it reads the file before every task and follows your preferences, giving you consistent results. Without a rules file, the same task can come out differently every time — renaming the same batch of photos, for instance, might follow a different naming logic each time you ask.
Plugins Connect Your Tools, But Teaching It How to Use Them Is What Counts
Connecting ChatGPT Work to Gmail, Notion, Slack, or Google Calendar (Claude calls these "connectors," Gemini calls them "connected apps") isn't itself anything new — it's a settings toggle and an authorization click, done in a few minutes.
What actually determines whether this is worth anything is the next step: have you taught it how you like to use that tool? You might say: "Default meetings to 45 minutes from now on, and write the Google Calendar description as 'conclusion first, then three bullet points.'" That line gets saved automatically under the matching rule category in `agents.md`. The next time you upload a lease and ask it to set a reminder to call the realtor, it defaults to a 45-minute block and writes the description "conclusion first" — even though you never repeated either rule this time.
There's a hidden cost here too: the more tools you connect — calendar, inbox, notes, chat — the more rules pile into the same `agents.md` file, and it keeps getting heavier. Since ChatGPT Work re-reads this file before every task, a bloated file doesn't just slow things down — it can degrade answer quality and burn resources unnecessarily. The fix is to split tool-specific rules into their own separate rules file, keeping the core `agents.md` lean. That's the real thing to watch for as you scale up usage, not "the more rules, the better."
The Sidebar's Three Identities: Preview Editor, Built-in Browser, Side Chat
There's a part of ChatGPT Work desktop that often gets overlooked — the sidebar, which actually does three jobs at once.
The first is a file previewer and editor: you can select a section of a rules file right in the sidebar, ask it to expand a new section in the same format, and undo instantly if the edit goes wrong. The second is a built-in browser: when what you're discussing lives on an actual webpage, you can highlight a specific section right on the page and describe the change you want (say, "give this section a dark-mode preview"), and it'll first show you a mockup — if the site itself is already connected with admin-level permissions, it can push the change live directly. Those are two separate actions, and "seeing the mockup" and "actually shipping it" are separated by a layer of authorization. The third identity is a "side chat" window: while your main task is still running (say, it's reading through a long PDF), you can open a side conversation for a related question you don't want to interrupt the main thread with. It continues with the main task's context to answer you, but that side conversation doesn't feed back into the main task, and its contents vanish once you close it — copy anything you want to keep.
Turning Repeated Edits Into a Skill You Can Call Again
When you notice you're making the same edit over and over — say, rewriting an AI's draft into your own voice every time — ChatGPT Work offers something called "Skills": feed it a set of inputs, run them against rules you've defined, and get the output format you want.
In practice: put the AI's draft and the version you actually published side by side in the same conversation, and ask it to identify the repeatable style differences between the two. It turns that into a reusable skill, and updates the newly identified tone rules into the "voice principles" section of the rules file. The logic is simple — you're going to make the same edit again anyway, so you might as well teach it the logic behind that edit now, and explain it to yourself less and less over time.
Once you've connected enough tools and built up enough rules, the last step is building a "routing map" for all of them — say, adding a folder for travel preferences (loyalty tier, seat preference, past itineraries) and teaching it that any travel-related request should automatically route to that folder. That way, the rules file doesn't turn into one giant pile of every detail of your life — it splits into sub-folders, each managing rules for a different part of your work or your life.
FAQ
Q1: What's the difference between ChatGPT Work and regular ChatGPT? The core difference is that it connects to the tools you already use (Gmail, Notion, Slack, Google Calendar) and remembers your working habits through a rules file, carrying those rules into brand-new conversations without you having to repeat them every time.
Q2: Do I have to manually edit the agents.md file? No. You can tell ChatGPT Work the rules you want it to remember in plain conversation, and it writes them into the matching section of the file automatically. You can also view or tweak the file directly in the sidebar's preview window.
Q3: What happens if I never set up a rules file and just use ChatGPT Work directly? Based on the demo, without a mounted folder containing a rules file, ChatGPT Work behaves essentially like regular ChatGPT — the same task can come out with different details or formatting each time.
Q4: Does the rules file get harder to use as it fills up with rules? Yes. As you connect more tools and accumulate more rules, it's recommended to split tool-specific rules (for Google Calendar, Gmail, and so on) into a separate file, keeping the core rules file lean — because the system re-reads this file before every single task.
Q5: Can the built-in browser in the sidebar actually push website changes live? Based on the demo, the built-in browser can first generate a mockup of a change (like a dark-mode preview), but actually pushing that change live requires the website itself to already have a connection authorization set up with ChatGPT Work — that's a separate step, not something the browser alone can do.
Q6: If I use both ChatGPT Work and another AI assistant (like Claude), do I need to write two separate rule sets? The rules file itself can live in the same mounted folder — different assistants just read different filenames (ChatGPT reads `agents.md`, another assistant reads its own equivalent file), while the rest of the folder's content and logic can be shared. You don't need to rebuild an entire separate folder structure for each assistant.
Source disclosure: This article is adapted from a video published on 2026-08-25 by the YouTube channel Jeff Su, titled "Learn 95% of ChatGPT Work in Under 20 Minutes" (https://www.youtube.com/watch?v=KmcTu2EigTs). Descriptions of ChatGPT Work's features in this article are a reorganization and explanation of the demonstrations in the original video; for actual feature definitions and interface details, refer to OpenAI's official documentation. This article does not constitute an official user manual.
Further reading: We've also covered what OpenAI Codex actually is and how it differs from ChatGPT and Claude Code for non-technical users, why more software is shifting to token-based pricing as AI agents change how procurement and cost governance work, what it actually costs to wire your own tool straight into Seedance 2.0's API instead of paying a subscription, and what GPT Image 2.5's Sunburst/Flare release means for e-commerce, game art, and brand asset production.
This article is about teaching AI to remember your rules; the next question worth asking is — once those rules are scattered across ChatGPT, Claude, and Gemini's separate tools, does your team have one place to see the token spend piling up behind them every day? Try AI Token King free → (https://aitokenking.com.tw/home), and bring the usage and costs scattered across different tools together into a single ledger.
