Most people burn their first ten minutes with any AI re-explaining who they are. Do the setup below once and you never pay that tax again.
The idea is simple and it is the same one behind every good AI workflow: the model is only as useful as the context it starts with. Type that context fresh every time and you waste it. Write it down once, in files the tool reads automatically, and every task starts from a running start. This is the ChatGPT version of that setup, updated for what actually shipped this month.
Three names to get right first, because the popular version of this guide gets them wrong. The workspace agent is called ChatGPT Work, which OpenAI launched on July 9, 2026. "Cowork" is Anthropic's competing Claude product, not a ChatGPT feature. There is no "Extended Thinking" toggle in ChatGPT either; that is Claude's term. And "plugins" no longer exist. They were retired and replaced by Apps and Connectors. The steps below use the real names.
The 30-Minute Setup
The Folder Structure
Keep it boring and predictable. The model does better when the layout never changes.
Codex-Work/
ABOUT ME/
about-me.md
my-voice.md
my-rules.md
PROJECTS/
TEMPLATES/
ChatGPT Outputs/
Four folders, one job each. ABOUT ME holds the three context files. PROJECTS is your live work. TEMPLATES holds the formats you reuse. ChatGPT Outputs is where finished work lands, so generated files never clutter your source material.
The Three Files That Do the Work
These are the whole point. Everything else is plumbing. Each one is short, plain markdown, and written so a model that has never met you gets you right on the first try. Start with these skeletons and fill them with your specifics.
about-me.md
# About Me - Who I am: [name, role, company, what you actually do all day] - What I'm working on: [current projects, the one that matters most] - My clients / audience: [who you serve, their level, what they want] - My stack and tools: [the software and formats you live in] - What a good outcome looks like: [describe a great result concretely]
my-voice.md
# My Voice - Tone: [direct, warm, dry, formal... pick words you'd actually use] - Sentence style: [short and plain, or long and layered] - Words I never use: [the cliches and filler that make you cringe] - A real sample of my writing: "[paste two or three sentences you actually wrote]" Match the sample above more than this description.
my-rules.md
# My Rules - Always: [lead with the answer, cite sources, ask before assuming...] - Never: [no bullet lists unless asked, no hype, no em-dashes...] - Formatting defaults: [headings, length, where files get saved] - When unsure: [ask one specific question rather than guessing]
The sample inside my-voice.md is the highest-leverage line in the whole setup. A model matches a real example of your writing far more reliably than it follows a description of it. Spend your time picking a good sample, not perfecting the adjectives.
Wiring It Up
With the files written, three things connect them to ChatGPT:
- ✓Add the folder to ChatGPT Work so it can read and write inside it.
- ✓Select GPT-5.6 and set reasoning to the highest level your plan allows. On eligible plans that is High or Extra High; Pro runs the strongest tier. This is the real setting behind the "Extended Thinking" advice you may have seen, which is a Claude term, not a ChatGPT one.
- ✓Add a standing instruction: "Before every task, read the files in ABOUT ME and follow them." Put it in your custom instructions so it applies to every session.
A description of your work is an opinion. A file the model reads before every task is a standing order.
Connect Only What You Need
ChatGPT Work can reach into your other apps through Connectors and Apps, the successors to the old plugin store. It is tempting to switch everything on. Do not. Connect the two or three you genuinely use, your mail, your calendar, your drive, and leave the rest off. Each connection is something with read, and increasingly write, access to your accounts. Treat that as a decision, not a default.
End With a Real Task
Do not finish on configuration. Give it something real, "draft the weekly client update from the notes in PROJECTS, in my voice, and save it to ChatGPT Outputs," and watch it use the context you just wrote. That first finished output is how you know the setup took, and it is the moment the half hour pays for itself.
The Honest Part
Two caveats, because a setup guide that oversells itself is worse than none. First, the strongest experience here leans on the Pro plan, which is not cheap. You can run this whole setup on a lower tier through the desktop app, you just get a less powerful model and shorter agent runs. Decide based on how much of your week this actually touches.
Second, and more useful: none of this is really about ChatGPT. The winning move, context in files the tool reads automatically, works in every serious AI tool. It is exactly what a CLAUDE.md file and skill files do for Claude. If you live across tools, write the context once in plain markdown and point each assistant at it. The tool changes. The files, and the hours they save you, carry over.
Want this kind of agentic setup for your own brand or product?
I help founders and brands build generative AI systems that actually ship: content engines, agentic workflows, and AI-built sites that are production-grade down to the schema layer. If you want a second pair of eyes on your setup, book a session.
Book a Growth Chat
Product names and model details reflect OpenAI's launches as of July 2026: ChatGPT Work shipped on July 9, 2026, powered by GPT-5.6, and is positioned against Anthropic's Claude Cowork. GPT-5.6 reasoning levels, the retirement of plugins in favour of Apps and Connectors, and desktop app availability were checked against OpenAI's help documentation and launch coverage. Features and plan availability change quickly, so confirm the current setup in your own account. This post is not sponsored.