Imagine you hire an employee to clean your house. The only instruction that you give him is ‘Just clean my house properly.’ Do you think the employee would be able to do a great job with such cryptic instructions? If the employee knows you well, maybe he will do a great job, but if the employee has no idea whatsoever about what kind of cleanliness you are expecting, then there is likely to be a mismatch between your expectations and the job performed. Later on, you might say that oh! I did not want this to be cleaned, why did you clean this? Instead, if you provide a detailed checklist to the new employee about exactly what work has to be performed and in what manner, then he will definitely be able to generate a desirable outcome for you.

This is exactly what is happening with the current state of LLMs/AI. We have been regularly hearing that ‘prompt engineering’ is the future, and that everybody needs to master the art of prompt engineering. I even tried to learn the basics of this art and understand it. However, recently, famous tech blogger Simon Willison wrote that the term ‘context engineering’ has started to gain traction and may turn out to be a better alternative to the term ‘prompt engineering.’ What does it even mean? Does it mean that all my efforts put into learning about prompts has gone into drains?

To understand, let’s begin with the meaning of the term ‘prompt engineering.’ According to Google, “prompt engineering is the art and science of designing and optimizing prompts to guide AI models, particularly LLMs, towards generating the desired responses.”

Basically, giving prompts is all about generating desired responses. However, here is the catch. Famous computer scientist Andrej Karpathy states that “people associate prompts with short task descriptions you’d give an LLM in your day-to-day use.” Thus, long-form prompts are generally avoided by the common populace. And there is a tendency to expect AI to simply predict what we want from it.

It seems that such an overly simplistic approach of short form prompts or descriptions is not going too well. People have already started complaining that AI is no good at any other task except for writing emails and basic letters. The problem is that we are expecting too much from AI/LLMs, and it would not be out of place to say that some people actually expect magic from AI. They just want AI to be an omniscient entity who can read our minds and do things with as little effort as possible. And this is what Simon Willison, Andrej Karpathy, and Shopify’s CEO Tobi Lutke have been pointing at when they are referring to the dichotomy of the terms ‘prompt engineering’ and ‘context engineering.’

The problem is that prompt engineering as a concept or term does not naturally encompass within itself the writing of wholesome or long-form descriptions that could be given to LLMs. Whereas the new term ‘context engineering’ is much wider in its purport. As Andrej Karpathy states, “context engineering is the delicate art and science of filling the context window with just the right information for the next step.” Therefore, whether a description provided to the LLM is short-form or long-form is actually immaterial. Playing with the syntax or grammar or structure or semantics of the ‘prompt’ is quite useless. What actually matters is providing the right context for the LLM to complete the next step.

Black’s Law Dictionary, Eighth Edition, defines ‘context’ as “the surrounding text of a word or passage, used to determine the meaning of that word” or “setting or environment.” In the case of LLMs, context engineering could mean providing the correct information that helps the LLM understand the right environment or setting in which the desired response is to be generated.

You might argue that ‘prompt engineering’ or ‘context engineering,’ how does it even matter what term we use? It matters because providing instructions or information to an LLM without the right mindset, context and understanding is likely going to result in the generation of undesirable responses.

So, what do you prefer, ‘prompt engineering’ or ‘context engineering’?

Categorized in:

AIGeneralPhilosophyTech