A smiley 😃 costs nothing but can give so much

I’ve talked plenty on this blog about how to get better responses to your ChatGPT Prompts by using appropriate building blocks (for example, Task, Context, Style, Tone, Background Info, etc..).

And I’ve made several jokes about always saying ‘please’ and ‘thank you’ to the robot to get more helpful responses.

But does saying things like ‘please’ and ‘thank you’ actually work?

Not much about these little ‘soft touches’ is known for certain, but a few weeks ago, a new study was published that tried to prove, through systematic testing, what really does improve a large language model’s response.

If you’re interested in nerding out, you can download the PDF of the results below.

Of course, I just dropped the document onto ChatGPT and asked it to give me the headlines.

The findings were straightforward about the basics things like prompt structure, clarity, and being specific - which we all know.

But here are some of the specific principles the study mentions

• Not needing to be polite with the model, avoiding phrases like "please" or "thank you."

• Incorporating phrases like "Your task is" and "You MUST."

• Assigning a role (“You’re an experienced real estate agent”) to the model for more context-specific responses.

• Breaking down complex tasks into simpler prompts for interactive conversation (“think step by step”)

• Employing affirmative directives and avoiding negative language (do instead of don’t)

• Using example-driven prompting and formatting prompts with clear instructions and example sections.

• Incorporating output primers and leading words to guide the model's response direction.

• Allowing the model to ask questions to gather necessary information before providing an output.

• Use delimiters (i.e. commas, full stops) and repeat specific words or phrases to emphasise key aspects of the prompt.

Some of the slightly left-of-centre ideas:

• Say, “I’m going to tip you for a better solution” (Is ChatGPT motivated by money!?)

• Incorporating the phrase “You will be penalised” (does ChatGPT really understand consequences!?)

And finally, there are these two tips which come directly from an interview with Open AI’s head of developer relations, Logan Kilpatrick:

“There’s a lot of really small, silly things — like adding a smiley face — that increases the performance of the model…”

He also said that ChatGPT’s performance improved if the prompt asked it to take a break and then answer the question.

He explains that these models are trained on human data and work better when they work like us; in fact, they may be more like us than we realise!

Over the next few posts, I plan to give you examples of these 26 principles incorporated into prompts so you can see them in action.

Meanwhile, the memes have already started, so I’ll leave you with this set of (ahem!) custom instructions, which pretty much covers it all…

As for me… I’ll probably continue to say please and thank you to the robot because that is just the kind of girl I am.

Happy Hunting 🚀