Plate or Bowl: A Simple Case Study in Prompt Refinement

At Advancing Analytics, we have a question we like to ask new hires: Are you a plate or a bowl?

This is just a silly icebreaker really but often leads to some interesting justifications as to people’s responses.

In a chat with colleagues someone made an offhand comment about Morpheus from the Matrix offering a plate or bowl instead of the red or blue pill from the movie. This comment made me want to try and generate this image using AI and was also an interesting example of prompt refinement.

Prompt refinement is the process of changing your prompts in different ways to lead to different results. During this process I started with an initial prompt and made small changes until I got the desired result.

I used the Bing image creator AI for this. It is completely free if you have a Microsoft account and is super easy to use. You just navigate to the page enter your prompt into a text box and hit create. The AI after a brief pause will usually offer you four options based on your prompt. If you can use a search engine, you can use this without any additional technical skills. Try it out for yourself!

My first prompt was “Morpheus offering the choice of a plate or bowl”.

Some interesting images but I didn’t think anyone would understand what I was trying to reference.

Often a more detailed prompt will get you where you want to go. I noticed in the pictures generated that three of them had something in just one hand and the other image had two plates, so I wanted to try and be more specific.

My next prompt was “Morpheus offering the choice of a plate in one hand or a bowl in the other”.

Only three images back this time. One of them slightly better in that there was a plate and a bowl but none of them really looking that much like Morpheus.

I decided to try and be even more specific in the next prompt and reference some of Morpheus’ clothing.

The next prompt was “Morpheus in a trench coat and sunglasses offering the choice of a plate in one hand or a bowl in the other”.

This response was a lot closer to what I had in my head. The AI seemed to pick up on the matrix theme as it made the bowls in one of the images Red and Blue all by itself. I thought I was a lot closer now but thought I could refine the prompt a bit more. The next thing I wanted to do thought was steal the red and blue idea. I thought having more description of the plate and bowl might get the AI to create a plate in one hand and bowl in the other rather than two bowls or only one.

The next prompt used was “Morpheus in a trench coat and sunglasses offering a red plate in one palm and blue bowl in the other palm”.

I still couldn’t seem to produce what I was looking for. The image generator didn’t seem like it would create a plate and a bowl. I thought I would do a last attempt and really try and describe the plate a bit more.

The prompt used was “Morpheus in a trench coat and sunglasses offering a flat wide red plate in one palm and blue bowl in the other palm”.

I got some ok results with prompt 5 but a clear winner also emerged.

There we are Morpheus offering a red plate or a blue bowl! This is just a silly example, however I hope it shows that by changing your prompts in small ways you can get really clear differences in the outcome when using AI. This also applies to LLMs as well so if you don’t get the response you are looking for straight away try changing your prompt, slowly adding more detail and context to get closer to your intended result.

Get in touch today if you want to learn how Generative AI can revolutionise your business processes..

One last question, are you a plate or bowl?