Prompting Techniques

The ability to communicate effectively is the foundation of successful task completion. This is true for all systems, be it the brain communicating with the body, colleagues on a project or digital systems. Moreover, effective communication is unique to the system and to ignore these differences is to limit the potential of the system.

Interacting with LLMs is no different. The vessel for communication is Prompts and each model has its own idiosyncrasies, a byproduct of different training methods and the data that was used. The process of prompting (read: communicating with LLMs) is necessarily explorative, there is scope for automation (see a future article) but in the absence of this, we can leverage an understanding of common techniques and thought patterns for structuring communication in order to improve our efficiency.

When you wish to communicate an idea with another, you are exposed to multiple world models. Their is both your own internal model, the construct you create for the other's interpretation of your message and their world model. Similarly the receiver has their own world model and construction of the intent behind your message. Good communication we can define, in the perspective of task achievement, occurs when each parties construction of the others intent is aligned. The role of the communicator is to share their message in such a way that the likelihood of this outcome is maximised. This is influenced by their knowledge of the other. If you meet someone for the first time, you may go to speak to them in your mother tongue, a priori you assume they also speak this language. If they do not, your model was inaccurate and communication will fail.

Consider a task routinely performed within your organisation. The process for this task being so well understood by long standing employees that their is no confusion on the data needed nor how the output is formatted. You may ask a colleague at any time to conduct this task and no further communication is needed. Contrast this with a new starter. If you requested the same task from them, with no additional guidance, you may receive a vastly 'abnormal' output. Their world model for this task is not aligned with yours and you did not, in this case, provide enough detail for them to bridge the gap. This is a failure of communication, characterised by a lack of context.

In the case of effectively communicating with LLMs, fundamentally you should aim to have a basic understanding of how they work and what data they have had access to. Many a complaint of hallucinations or poor performance can be traced back to this. When communicating with LLMs you are navigating a vast latent space, your words, tokenized, offering a trajectory through multiple probabilistic pseudo world models. Your task as the communicator is to maximise the likelihood of your intent being understood. Analogous to the case of the new starter, you must filter and constrain information such that world models are aligned, such that the receiver has sufficient context, else you should not expect good performance.

In the below slides and video walkthrough we provide detail on foundational techniques that have been used for structuring communication with LLMs.