Does the role, context, task method still work with GPT-4o?

A Beginner's Guide to ChatGPT teaches a method built for GPT-3.5 and GPT-4 in early 2023: give ChatGPT a role, explain your context, state the task, say what format you want, then refine the answer and verify it. A year later, OpenAI's newest model, GPT-4o, is fast, reads photos and documents, and is available to free users. It is fair to ask whether all that structure is still necessary.

It is, for one simple reason. A better model is better at doing what you ask, but it still cannot know what you didn't say. Your audience, your numbers and your deadline live in your head, not in the model. The method exists to move them from one to the other.

What changed with GPT-4o

OpenAI says GPT-4o matches GPT-4 Turbo on English text and code, does noticeably better in other languages, and is much stronger at understanding images and audio. For free ChatGPT users, it also unlocks tools that were paid-only: uploading files, chatting about photos, getting answers that draw on the web, analyzing data, and Memory, which keeps facts you share across conversations.

Each of these touches one step of the method.

A side-by-side example

The book shows prompts at three levels for different jobs. One of its simple prompts for an online store is just "Create a financial report." The improved version names the period and the figures wanted (monthly revenue, expenses and net profit). The refined version asks for six months of data to be analyzed, with trends, anomalies and recommendations.

Here is how the refined level might look written for GPT-4o, with a role added and the context supplied as a file:

You are a bookkeeper who explains finances to small business owners with no accounting background.
Context: the attached spreadsheet holds six months of sales and expenses for my online shop, which sells handmade candles. I want to decide whether to keep paying for advertising.
Task: work out monthly revenue, expenses and net profit, point out any month that looks unusual, and tell me what the numbers suggest about advertising.
Format: a short table by month, a line chart of net profit, then no more than five bullet points in plain language. List any assumption you made.

What changes in practice, in an illustrative run: the book's prompts assume you type or paste your information into the message. With GPT-4o, even a free user can attach the spreadsheet, get the table and chart in the same reply, and get it faster. What does not change: without the role, the answer tends to drift into accounting jargon. Without the context about advertising, the model has no idea which question matters. Without the format, you get several paragraphs where a table would do.

Which parts can you now skip?

Fewer than you might hope, but a few things are lighter.

You can skip retyping standing context. If you turn on Memory and tell ChatGPT once that you run a candle shop or teach grade 5, it can carry that into later chats. Check what it has remembered from time to time, and restate anything important when the stakes are high.

You can often skip transcribing. Instead of describing a chart, a handwritten note or a receipt, take a photo and ask about it. Instead of pasting a long document, attach it. The context step is still there; it just takes less effort.

You can keep the role short. A single line such as "You are a bookkeeper who explains things simply" does most of the work. Long, elaborate role descriptions add little.

You cannot skip the task, the format or the check. A more capable model can do more with a clear request, so clear instructions are worth more, not less. And the book's warning still holds: any model can state an error with confidence, so names, numbers, dates and contact details need your review before you use them.