Leon Furze

When ChatGPT was released – roughly 564 years ago in education years – one of the first approaches that emerged to support teachers were lists of prompts people could use with the application for things like lesson planning and resource creation. A lot of these prompts focused on techniques like getting the chatbot to role play, giving it a clearly defined role and so on.

I know this because one of my early 2023 posts took exactly this approach. The original Practical Strategies for ChatGPT article suggested six areas where educators could use different types of prompts for planning, updating curriculum documents for using communications and other day-to-day tasks.

Practical Strategies for ChatGPT in education

As the technologies have developed, however, it has become clear that it is also necessary to understand the process of working with generative AI.

Process > Prompts

If I reflect on how I use ChatGPT, Claude and so on, it’s very rare that I use a pre-planned prompt. Instead, my methods of interacting with them has changed as the technologies have developed.

There are a few key changes that have happened since 2022 which have changed my process of interacting with GenAI, and which mean I’m not that concerned about the quality of my prompts:

• The ability to attach files to prompts, including documents, PDFs and spreadsheets
• The incorporation of image recognition
• Connecting the models to the internet via search engines like Bing and Google
• The ability to write and execute code
• More recent improvements to models for research purposes and more complex queries

There have been other developments in GenAI since 2023, but for me, these are the important parts that have shaped the way that I interact with the technology.

In this post, I’ll explain each one in more detail and how it has impacted the process and lessened the need for specific prompts.

If you’re interested in AI images, audio, and video, make sure to grab the free 20+ page resource How to Spot a Deepfake by signing up here:

Uploading files and documents

Most large language model based applications now have the capability for a user to upload documents in various file formats. The exact type of file varies from one platform to another that includes Word documents, PDFs, text files and markdown files and spreadsheets, code such as .py Python files, images and so on.

As well as adding the feature of document uploads, many of the models now have a much larger “token limit”, meaning for practical purposes that they can incorporate more text in an individual chat.

In terms of the process, this has had two big implications:

The first is that it is possible to provide much more context and therefore reduce the amount of explaining necessary. In 2022/2023, a user might have written something like:

I am a secondary English teacher in Australia using the Victorian curriculum and the following achievement standards (copy and paste achievement standards). Help me to understand which of these achievement standards would work best for a unit of work on media literacy.

Now the same user can upload a .docx or PDF of the curriculum and simply type:

Identify best standards for media literacy.

The second implication is that it’s now much easier to work with multiple documents, compare and synthesise ideas such as updating curriculum materials from an old syllabus to a new one, or review and edit longer texts.

The ability to upload files is also necessary for some ways of interacting with generative AI, such as data analysis using spreadsheets or image recognition, which I’ll talk about separately.

Probably two thirds or more of my prompts include some kind of file upload as part of the process. This is indicative of the fact that I don’t use GenAI to create much content, but I’m typically working with content that already exists. For me personally, the most common form of content that I upload is a .txt file of audio transcripts which provides the raw materials for making drafts of blog posts like this one.

Image recognition

As well as being able to upload images, many of these models now have improved capabilities at recognising what’s actually in them. This can be useful in many ways. I have found, for example, that image recognition in ChatGPT is comparable to and sometimes more accurate than Google Lens for identifying things such as plants, insects, birds and types of rock…

The image recognition features of these platforms are now also pretty good at recognising handwriting, which can be useful when scanning in student work or capturing handwritten notes, whiteboard brainstorms, information from post-it notes and so on. Image recognition can also combine with other types of file upload, meaning that it’s possible to upload multimodal texts like PDFs, which include graphs, maps, charts, etc.

In terms of process, I now no longer have to type lengthy prompts like: These are the notes from a recent faculty meeting. Organise them by theme and who said what (type lengthy list of notes). Instead, I can take a photo from a whiteboard and just write:

Transcribe and organise these notes.

Connecting models to the internet

This is one of the most important changes to large language model based applications like ChatGPT, and surprisingly, one that a lot of people are still not aware of. Every time I speak with educators and students about generative AI, there are questions about how up to date the training data is and where the models can access recent information.

Since 2023, applications like ChatGPT, Microsoft Copilot and Google Gemini (formerly Bard) have had an internet connection, which effectively means that the training data cut off is much less relevant. Some people are not aware of this simply because they have not had the time or the inclination to use technologies like ChatGPT much since their release in 2022.

The internet connection means that many early approaches to prompting are now redundant. These models connect to the internet in a variety of ways:

• ChatGPT through OpenAI’s partnership with Microsoft uses the Microsoft Bing search engine
• Google Gemini, of course, uses Google and also has access to YouTube videos (since Google owns YouTube)
• Microsoft Copilot uses Bing
• Perplexity uses a variety of search methods in its proprietary application including Bing
• Claude, a powerful large language model from Anthropic has lacked internet connection until very recently, when it was added for paid users

Connecting a language model to the internet has a few implications, some good and some bad. On the plus side, it is now generally possible to enable search in ChatGPT or use a phrase like use internet search in the prompt to get more accurate and up to date information. Some features also enable models to conduct more complex internet searches, which I will discuss later.

One downside is that models can sometimes search the internet even when you don’t want them to, pulling in irrelevant data. And, as evidenced by Google’s terrible “AI overview” search results, sometimes large language model based systems can incorporate erroneous or irrelevant information, since they lack common sense. Large language models themselves do not make very good search tools, so the best way to use these is for the language model to “hand off” the search query to a traditional search engine.

As with any searches conducted online, it is important to check the accuracy and validity of sources and questions, since sometimes AI can return results which are not accurate or not from trustworthy sources. However, unlike the earlier versions of ChatGPT, the application now at least provides clickable citations to those original sources.

It is worth pointing out that the training data is still very important, but the primary purpose of the training data of a large language model is to teach it the rules of syntax and grammar and not to enshrine any particular knowledge in the model.

A large language model with no internet access is not a reliable source of information and should not be used like a database or search engine.

For my process, working with large language models in education, internet access has some significant benefits. For example, Perplexity can quite easily pull in even obscure curriculum documentation like the content elaborations from the Australian Curriculum website. This is a task beyond most human users since the curriculum website is a nightmare to navigate, with endless pages of dropdown menus and nested links.

Writing and executing code

One thing that I always try to impress on educators is that large language models are software developed by programmers for programmers. Out of all of the areas they are being used successfully, computer programming is the most prolific. Applications like ChatGPT started life in fields of research like machine learning, translation and code completion tools: technologies designed to apply algorithms to data which has clearly structured syntax.

Computer programming languages are much more formal and reliant on accurate syntax than human languages. If you’re writing in English and you leave out a comma, you might get an angry comment from a particularly pedantic English teacher, but it’s unlikely most people will notice. On the contrary, if you leave off a semi-colon whilst writing JavaScript, the entire section of code will stop working.

When you train language models on large data sets of programming languages, they become very capable at writing in those languages. Most people I know couldn’t care less that ChatGPT can write code, but I really want you to care. Right now, there’s a lot of talk about “vibe coding“, which can basically be summed up as using natural language to code partial or complete applications. Vibe coding is fun and can be intellectually challenging, but I want you to focus on something even more straightforward.

Large language models are very capable at writing code. Some can run the code that they write. Everything that happens on a computer happens in code. Therefore, it follows that:

a large language model can write and run a simplified version of most things that you can do on a computer.

Let that logic sink in for a moment.

• Want to create a spreadsheet with particular column headings and formulas in each cell? ChatGPT can use code to create the spreadsheet and provide you with a downloadable .CSV or .XLS file.
• Want to convert a PDF to a Word document? AI can use code to extract the text from the PDF, clean up any strange formatting and export it as a downloadable Word document.
• Having a conversation with ChatGPT, and want to turn your thoughts into a slideshow? Prompt it to generate a .pptx file.

All of these things are possible, particularly in ChatGPT, because all of these things are possible with code. I encourage educators, once they’ve got familiar with the basics of these applications, to really dig into the implications of this.

If you are using a computer for something that involves converting files, creating files, moving data from one type of file to another, analysing the type of data in certain platforms, creating websites or parts of websites, conducting data analysis and with mathematical code libraries, any of these tasks can be replicated with AI.

It’s super geeky. You feel super geeky doing it as you watch ChatGPT open up its code interpreter window and produce lines of multi-coloured text against the black screen. It feels like you’re slipping into the 1990s movie, Hackers (great movie, still holds up), but again, the implications for the process of how you use generative AI for things like curriculum design is huge. I treat ChatGPT not like a teaching colleague or a student, but like a versatile piece of software, a Swiss Army knife for generally passable code. It’s not better than an expert human software developer, but it’s certainly better than me.

Recent advances in reasoning and research applications

I don’t buy into the hype that “reasoning” models have made applications like ChatGPT better than humans across most domains, or that this kind of approach will suddenly lead to AI becoming conscious and achieving godhood. I also don’t think that applications like OpenAI’s Deep Research are capable of replacing entire portions of whole industries. That doesn’t mean that they’re not useful technologies, and like all of these advances discussed so far, they have had an impact on the processes that I use for working with AI.

The “reasoning” models take the first response from the chatbot and pass it back to itself, allowing it to iterate through its response and improve the quality of the answer and the logic behind it. I use OpenAI’s o3 model probably as much as the standard 4o model now, particularly when undertaking tasks which are less straightforward than simple search-style queries. This includes tasks like the aforementioned writing and executing of code, since o3 is generally more accurate, but also now that o3 has an internet connection, I have found it useful for research based tasks, which might involve gathering up resources, curriculum documents and so on.

The various deep research applications from OpenAI, Google and Perplexity, which I have written about before, also use a form of “reasoning” model to improve the efficacy of their searches. And in terms of improving the process of using these technologies, one of the biggest advantages is that you can be more ambiguous with your initial prompts and spend less time crafting clear instructions.

Take this comparison of a response from 4o to the same query used in “Deep Research”:

While I hope you never have to read through a 12,000 word AI-generated report like the one produced here, the quality and breadth of the sources Deep Research has identified is much better than the 4o response. This lengthy report could be used as the basis for some genuinely helpful materials.

I use Deep Research to scan the US, UK, IB, and Australian curricula when identifying areas to teach AI ethics, for example. It is much faster than me trying to navigate half a dozen or more poorly formatted government websites.

Summary

Here’s a quick summary of all of these parts of the process:

Start with process, not pre-determined prompts.
Instead of stock prompts, begin a dialogue knowing what you need to accomplish and let the exchange develop in natural language.

Use file uploads to give context.
Upload curriculum documents, transcripts, datasets or images so the model can “see” what you are working with. This removes the need for lengthy explanations and makes tasks such as syllabus comparisons or large-scale text revisions far faster.

Use image recognition to translate visual to text.
Photographs of handwritten notes, whiteboards, student work or field observations can now be transcribed and organised with a single instruction. Multimodal PDFs containing charts or maps are equally accessible.

Enable the model’s internet connection.
ChatGPT, Gemini, Copilot and others can pull current information and cite sources. Use search when you need fresh info, but verify results and watch for irrelevant or spurious links.

Treat code execution as a digital Swiss-Army knife.
Large language models can write and run Python and other code to create spreadsheets, convert files, build slideshows and more. If a task is code-friendly, the model can probably automate it.

Use reasoning models for complex or ambiguous work.
Iterative models like OpenAI’s o3 and dedicated “deep research” tools refine their own answers, making them well suited to curriculum mapping, resource scans and other multi-step investigations where the first draft isn’t enough.

Combine these features into a single, fluid workflow.
A typical session might involve uploading a syllabus, asking the model to identify media-literacy outcomes, having it fetch recent policy documents online, and finishing with a code-generated slide deck, without any “prompt engineering”.

Download the PDF:

• LinkedIn
• Bluesky
• Facebook
• ShareCopied to clipboard
• Mastodon
• Threads

Want to learn more about GenAI professional development and advisory services, or just have questions or comments? Get in touch:

Cover Image: Hanna Barakat & Cambridge Diversity Fund / https://betterimagesofai.org / https://creativecommons.org/licenses/by/4.0/