Improving Prompt Craft for ChatGPT
I’ve been an early adopter of AI tools, using earlier versions of MidJourey and ChatGPT. However, it wasn’t until I took a Prompt Engineering for ChatGPT Coursera program by Vanderbilt University that I realised how ineffective my prompts were.
In the early days, most were using it as if it were Google - asking it simple questions or information about current events, even though it’s only trained on limited data up until September 2021.
It wasn’t until I started using personas, context and tail prompts that I started seeing much more useful outputs.
I’ve shared a couple of my key takeaways in my free Prompt Craft cheat sheet too if you'd like to learn more.
Probably, not magic
On average, only 14% of the population has used ChatGPT. To those that use it for the first time, it can often seem like magic, watching it craft new stories or songs in seconds. In reality, it’s all just probability.
ChatGPT uses probability to calculate the next word in a text based on the patterns it has learned from the training data it was trained on. It relies on a statistical approach called language modelling.
When you provide a prompt to ChatGPT, it uses the context of the prompt and the preceding text to generate the next word. It does this by estimating the probability of each possible word that could follow the given context.
This is why context is critical when asking ChatGPT questions. You shouldn’t just ask a question, you be assigning personas, be concise yet specific and add tail prompts to ensure the output you get matches your request.
Context, but conceise
When I first started using ChatGPT, I would generate single-sentence prompts and then refine them throughout my interactions with the chatbot. But by adding additional context to your prompt - such as skill level, personality or time frames - you generate more relevant information a lot quicker.
Despite the fear of being impolite. You really shouldn’t use excess words for your prompt as the additional tokens could confuse the output the AI is trying to generate.
Personas
Because ChatGPT uses probability, you get more relevant results if you assign it - and sometimes yourself - a persona. This helps narrow the perspective of the AI so it’s focusing on a particular topic, task or industry.
My favourite personas for the AI to act as a marketer, creative director, business analyst and university lecturer. When assigning myself a persona, if I want the output to be simplified I may ask it to explain it to me as if I were an 11-year-old. Or if I want more detail about the theory, I may assign myself the persona of an MBA student.
Tail prompts
After you’ve assigned a persona and given the AI a task, it doesn’t mean you need to finish your prompt. Tail prompts are a great way of adding commands on top of the primary task. For example, you could ask it to remember the persona for future requests, ask you for additional information or suggest alternative solutions.
Outline Expander
Having the AI act as an outline expander has been a huge help, particularly when you assign the AI a persona relevant to your industry or task. When tasked with creating an outline, the AI will generate the headings and sub-headings, and upon request, expand on the types of information to include within each subheader.
This is a game changer if you’re frequently creating business strategies or proposals. Rather than having to recycle the headers in old documents (and potentially leave the previous client name or some random statistic, AI can coach you on the content to include from a fresh slate.
Final thoughts
It was a very insightful experience taking a more formal prompt engineering course. I probably could have got the same information for free from youtube, though it would have taken much longer to find all the right content and the fact I needed to hand in assignments helped me ‘learn by doing’ rather than by watching.
I’ve shared a couple of my key takeaways and some examples of the tips above in my free Prompt Craft cheat sheet.
If you have any questions, or feedback or would like to learn more, just reach out.
Comments