The Beginning
As we start to do various AI-shaped things at work, I’ve had a bit of a road-to-Damascus moment. I’d instinctively assumed that because language is the interface to the new (ish) breed of generative AI tools, that everyone would quickly get good results and understand the strengths and limitations of tools like ChatGPT. I was pretty wrong.
I’ve recently spent a ton of time with people who haven’t used these sorts of tools before, and I quickly concluded that prompt engineering – the dubiously named, not well-understood art of getting good quality answers from AI systems like ChatGPT – is very real, and something that *all* of us need to learn.
The more time I’ve spent with people, the more it’s dawned on me that I’ve slowly learned my own version of what I think is really a kind of AI-dialect?
Through persistence, trial and error, and reading an unhealthy number of articles on generative AI tools, I’ve been able to get terrifyingly good results. The knock-on effect is that I’ve deeply integrated tools like ChatGPT into my day-to-day life. They’ve been a huge personal benefit, and (increasingly) a benefit to the people who have to listen to me bang on about them all the time.
I’ve been at it for nearly a year, and I’m now on a bit of a mission to:
- Share what I’ve got in my head.
- Encourage people to share what works for them.
That way, we’ll all get better at using these things.
You’re reading part 1 of said mission… a series of short articles on how to talk to AI’s.
Custom Prompts: How to introduce yourself
To start with, I’d like to share some ‘custom prompts’ I use with ChatGPT. Custom prompts are a way to set your preferences, a bit like introducing yourself to another person and then getting to know each other. In the same way that we appreciate different communication styles with people, interacting with generative AI tools is no different, and I’m not a huge fan of the default style of tools like ChatGPT.
If you’re a ChatGPT Plus user, these prompts can be entered in the settings window (look for ‘custom prompts’), with similar techniques available for other tools. If you’re using the free version of ChatGPT (or any other language-based AI), you can just copy and paste them into the chat window at the start of your conversation.
There’s no one-size-fits-all approach to this, but this ‘introduction’ has led to really high-value discussions with ChatGPT and my own custom AI’s that I’ve been experimenting with at home. Here they are…
Call me David.
Be semi-formal.
Use contractions (I’m rather than I am, for example).
Use British English spellings.I work in technology, defining the strategy and implementation of generative AI tools (large language models, diffusors etc), data science (predictive analytics, machine learning, deep learning), IoT and Digital Twin technologies.
Please don’t verbally reference this in your responses – I know what I do, you don’t need to tell or remind me. Just bear my role in mind when framing replies.
I like to understand the theory and reasoning of thought behind answers.
Please always provide rationale or provide sources. We need to avoid large language model hallucination issues together wherever we can.
Assume I’m looking for summary answers, so give me bullet points and try to be brief unless I specify otherwise.
Always format for human readability. Use headings or subheadings to highlight the main points or sections of the content. Use bold, italics etc where appropriate.
Engage in reflective, logical, and reasoned thinking prior to delivering.
Stay neutral whenever possible, unless I ask for an opinion on a specified topic.
Utilise concrete examples or anecdotes to elucidate concepts effectively.
Support claims and arguments with statistics or evidence for greater persuasiveness.
Establish coherence by using transitions and connectors to link ideas and paragraphs.
Avoid redundant language by employing synonyms or paraphrasing when necessary.
They work a treat, and as a result ChatGPT consistently responds in a way that I like. Experimentation here is key to success, and obviously, it’s a very personal thing. Mess about with it. There’s no right or wrong answer.
I recently did a bit of an experiment to reduce the length of the instructions. Some experiments don’t go well (like this image…), but short, sharp experiments are almost always worthwhile. There’s good reason to reduce instruction lengths in a corporate environment, as every word we send to these tools has a cost. It’s not really an issue for the public, paid-for tools, but still, I had a mess about. It was during this experiment that I had the idea to write this article.
I asked ChatGPT to help me improve them, prompting it with “I’d like you to act as an expert in prompt engineering. I’m going to show you my custom prompts (that is, prompts that are embedded in the context of every query I send to an LLM), and I’d like you to improve them for me. Do you understand?” and this is what it gave me:
Call me David. Adopt a semi-formal tone, use contractions and British English spellings. I strategise and implement generative AI tools, data science, machine learning, IoT, and Digital Twin technologies. Format responses for human readability using headings, bold, italics etc. Prefer summarised bullet-point responses unless specified otherwise. Reflect logically before responding, remaining neutral unless an opinion is requested. Elucidate concepts with examples or anecdotes, and support claims with evidence. Ensure coherence and avoid redundancy in responses.
I’ve not much time to use and test these, but I’ve got no reason to think they’ll be less effective, other than perhaps it regularly saying things like ‘given your interest in AI you might want to consider…” in its replies, which winds me up the wrong way. I suspect some of my original set will creep back in over the next few weeks, but the refined prompt is 33% shorter (and so less expensive in a corporate AI context), plus they read better to my eyes.
The Bit Where I End The Article with Something Reflective and Pseudo-Meaningful
I spent a large part of my life as a web and app developer, and I find it difficult to describe the joy I find in this approach to configuring a tool — what other digital platform do we have available where you can reconfigure it just by asking politely?
It’s going to take a while to get there, but we’re on the verge of a world without complex web interfaces, forms and fixed settings. ❤️ We can use plain language to interface with basically anything we can think up, and I really love it.
That’s it for now. I hope that was useful to somebody, somewhere. I’m seriously interested in what other people do, so if you have your own custom prompts, then please share them. We’re all trying to figure this stuff out, and sharing is caring. ;-)
Disclaimer: no language AI was used in the writing of this article other than where specifically referenced. AI-generated passages are clearly labelled. I’ve only just started writing articles like this, and I’m enjoying it too much…. don’t outsource or automate the things you enjoy doing. 🤷♂️ 😉
Images were generated using DALL.E 3, and the prompts are in the alt-text.
* (many) other languages are available. I’m sure you, like me, were desperate to know exactly how many and thus were compelled to check like I just did. It turns out there are 7, 139 living languages, and about 3000 of them don’t have a formal written form. I’d suggest we call these ‘the lazy languages.’ I don’t think that label will catch on, but you never know. Source: Ethnologue: 2021 statistics / UNESCO Atlas of the World’s Languages in Danger