A prompt is generally made up of up to four elements:
- Instruction - a specific task or instruction you want the model to perform.
- Context - external information or additional context that can steer the model to better responses.
You can provide additional context with a persistent instruction (though not all platforms offer this feature) also known as a custom instruction. - Input Data - the input or question that we are interested to find a response for. This might be pasted text, a link to a website, or a document, depending on the platform you are using.
- Output Indicator - the type or format of the output. Again, this can be covered generally in a persistent instruction.
Prompts: Do
Start Simple and Iterate: The best way to arrive at an effective prompt is to start with a simple instruction and then iterate according to how you want the output to look.
Be Specific: The more descriptive and detailed the prompt is, the better the results.
Be Precise: Just like any effective communication, you should be direct and avoid ambiguous language or words that can have multiple meanings. Fortunately, lawyers are already very good at doing this!
Use Positive Phrasing: Tell the model what it "should do" not what it "shouldn't do". LLMs seem to be as guilty as humans of falling into the "don't think of a pink elephant" trap. Positive wording keeps the model focused on the detail you need.
Remember its limitations: No matter how much you tinker with your prompts, the LLM may still produce factual hallucinations and biases. Always sense-check your output.
Prompts: Don't
Ask too many questions in a single prompt: It is better to ask each question as an individual prompt if possible. Some models, such as ChatGPT, will already take your previous questions (and its own response) into account.
Mix topics: For the same reason, you should avoid mixing too many topics together in the same chat window for chatbots like ChatGPT. The model may infer a connection between your prompts, when in fact there is none.
More Tips
- Chain of Thought commands: LLMs are, ironically, not very good at algorithmic tasks, and especially poor at maths and logic puzzles (though improving all the time). If you want the model to carry out an algorithmic task, it can be useful to ask the model to show its reasoning or chain of thought, or work "step by step". For example, if you wanted the model to suggest a litigation strategy.
- Persona prompts: Asking the LLM to respond as if it were a certain persona (like "High Court Judge" or "skilled negotiator") can be an effective technique to deliver a response in the style you want.