Over the past three years I’ve written dozens of articles offering advice on how to use GenAI well (I’ve also written plenty about why not to use AI… check them out elsewhere on the blog). As the technologies have developed and improved, I’ve changed the way I teach people how to use these applications.
Back in January 2023, one of my first major articles on AI was Practical Strategies for ChatGPT in Education. That article was focused on prompting in six areas – planning, refreshing, improvising, personalising, collaborating and communicating – and offered example prompts for educators.
In the intervening years, however, prompting (or “prompt engineering” if you must…) has become much less important. It’s no longer necessary to prime AI models with lengthy prompts giving extensive contextual information. In most applications it’s now possible to add extensive contextualising materials in a variety of formats, including PDFs, docs, web links, spreadsheets, and images.
There have been other improvements to the technologies which also reduce the need for perfect prompts. Improved internet access, “thinking models”, applications like Deep Research and other all make answers more accurate and useful – even if AI companies haven’t fully addressed the issues of hallucinations.
Soon, I’ll be releasing a new complete course to update Practical AI Strategies and offer detailed lessons on how to apply each of these steps. In this post, I’m outlining 5 key steps of what I’m calling The Practical AI Process.
The Practical AI Process
I’m… not incredibly imaginative when it comes to naming books and courses. So far, I’ve written Practical Reading Strategies, Practical Writing Strategies, Practical AI Strategies, and published a complete course in 2024 called… Practical AI Strategies. Based on my work with thousands of educators over the last few years, I can’t really get past the idea that what people want is practical, easy to follow advice on how to use these technologies in their day-to-day work.
So The Practical AI Process follows suit, and outlines five practical ideas which can be applied to almost any GenAI platform, including multimodal GenAI such as image, video, and audio generation.
Here are the five steps:
1. Apply your expertise
Before even cracking the lid on Copilot, ChatGPT, Gemini, or whichever AI application you decide to use, I want you to lean in to your existing expertise. I’ve written before about a major problem with GenAI: if you’re not already an expert in the area you’re using AI to work on, you’ll be much less likely to spot hallucinations and errors. I followed that original post with an article on three dimensions of expertise: domain, technological, and situated.
In that article, I argued that we need to balance the disciplinary, subject matter domain expertise with some technical knowledge of how GenAI works. I also suggested that we often neglect the situated, lived experience developed over time: that contextual expertise that is often the mark of a professional.
Step one of the process, then, is to think about the knowledge, skills, metalanguage, and context of your area of expertise, and use that to interact with GenAI to produce better outcomes.
2. Select the right model
Once you’ve decided on what you’re going to do, and how you’re going to apply your expertise to the task, you need to select the right model for the job. I’m platform agnostic: I don’t care whether you’re using ChatGPT, Claude, Gemini, or some LLM-based application you knocked up in your shed. As long as it’s the appropriate application to get the task done, it will be much more effective.
This is slightly more complex than it sounds. Most people haven’t had the time or the inclination to keep up with every change in GenAI since 2022. Many educators I speak with have never tried so-called “thinking” models, or applications like Deep Research. Many have also not yet realised that it’s possible to manually turn on those modes, including in free versions of applications like ChatGPT.
At a minimum, I think you should know which types of model are suitable for which jobs, such as:
- Fast, mini, or instant models (e.g., Gemini Flash): quick responses, idea generation, bouncing low-stakes ideas, anything that doesn’t require factual information
- Thinking or reasoning models: more complex queries, prompts requiring more factual information, prompts involving the synthesis of ideas from multiple documents or web sources
- Deep Research style applications: extensive web searches and summarising lots of information
- “Canvas” (ChatGPT, Gemini), “Artifacts” (Claude) and similar applications: generating code including for simple websites, web apps, simulations and data analysis
- Custom chatbots like GPTs, Google Gems, Claude Projects, and Copilot Agents: repetitive tasks which involve a lot of contextual information or specialised instructions
- Browser based and other “AI Agents”: semi-autonomous web browsing tasks and for some reason lots of online shopping and travel bookings…
- Local and open source AI models: private, offline, cloud-free usage.
3. Add context
Once you’ve selected which model or application to use, you can make decisions about how much context the task requires. For some tasks – such as those where the “instant” or fast models are suitable – you probably don’t need any. But for everything else, a little context goes a long way. Modern commercial platforms like ChatGPT and Gemini can handle context in a variety of files, including:
- PDF documents combining text and images
- Word docs
- Spreadsheets
- Code (e.g., .py files and .html files)
- Images
- Images of text (e.g., handwritten notes)
- Audio and audio transcripts
- Video files (Gemini advanced)
4. Use the internet
This one sounds like a no-brainer to anyone who has used GenAI much over the past few years, but many people are still unaware that most commercial models now have internet access.
In some models, like Microsoft Copilot and Google Gemini, this is enabled by default. In others, like ChatGPT and Claude, you can turn on internet access in the settings/tools option below the text entry field. Forcing the internet connection on in the free version of ChatGPT, for example, makes GPT5-mini more accurate (though still not as accurate as GPT5-thinking).
Sometimes, it is necessary to include a phrase like “search online and…” or “use the internet…” within the prompt to trigger the internet search. Most models will show you that internet search is running with a sort of loading icon, and will provide links either as clickable footnotes, or in a list at the end of the response.
5. Iterate and refine
It is highly unlikely that you’ll get exactly what you’re looking for in a single prompt (sometimes called a “one-shot”). Instead, you’ll probably have to dialogue back-and-forth with the chatbot for a while to get what you need. once you’re in the habit of tweaking the results, however, you’ll likely find that this process can still be fast and effective.
It’s also often worth bouncing between platforms, selecting the right model for individual parts of a task. For example, it’s possible to generate fairly sophisticated code for a simulation (e.g. a maths or physics simulation) in Claude, and then copy that over to the slightly more powerful ChatGPT GPT5-thinking model for refinements. Or, you might create an image in one platform, and then edit it in another.
Pulling it together
There is no magic formula for using GenAI – no instruction book or helpful acronym for magical prompting. In fact, one of the biggest issues with the technology is that it doesn’t come with a clear user guide. The chatbot interface through which we interact with many of these platforms is also woefully inadequate.
Applying the steps of this process is just one way to get more control over the outputs. Applying expertise, selecting the right model, adding context, making use of the internet connection, and iterating and refining will result in overall better quality outputs from text, image, audio, and video models.
The new course launches next week on Monday November 10th.
Join the mailing list for a launch discount until Sunday 23rd.
Leave a Reply