Master the Art of Prompt Engineering

by

Prompt engineering is one of the most undervalued skills on the planet right now. Most people use large language models like Google search— they try to find something.

LLM’s are about generation and creativity: converting data into new ideas, not finding old ones.

Effective prompt engineering is about clarity. It’s about defining concepts and enabling the model to achieve its creative potential. When we use the right techniques, we get the right results.

The following tips, placed on a sliding scale from easy to advanced, will help you level up.

1. Provide context

  • Tell the model why you’re asking your question. Tell it about the overarching task you are trying to achieve.
  • For example (context is in italics): “I am very interested in helping make the world a more peaceful place. I am writing a proposal for a town hall to approve the construction of a zen garden at the park. Help me write it using the principles of clear and logical writing.”

2. Provide it with your data

  • Many models today have a memory of your interactions with them.
  • Use this to your advantage. Ask the model questions like “based on your memory of our interactions, what am I focused on the most right now?” or “what are some patterns that you can identify in the ways I ask questions?
  • You can also upload your own documents!

3. Be clear

  • Keep your prompts simple yet equip them with detail. Be very clear. Avoid ambiguity.

4. Break down complex requests

  • Ask your model one question at a time. It has been proven that the more time your model is allowed to think, the better it performs.
  • For example, break down a task such as “write me an essay about elephants”.
    • Find me information about what elephants eat.
    • Find out what animals, if any, like to eat elephants.
    • Find information about how elephants contribute to the balance of their ecosystems.
    • Put this information together in an essay with 1 introductory paragraph, 3 body paragraphs, and 1 conclusion paragraph.

5. Specify your desired output format

  • When you specify an output format, this reduces ambiguity and gives the model one less thing to think about. Plus, this capability is pretty cool.
  • Ask for a list of pro’s and con’s or a table… instead of just seeing what it creates.

Examples:

  • “Create a table with two columns that lists all the ways in which elephants support their ecosystems. In the first column, name the way. In the second column, specify the way.”
  • “Create a list of pro’s and con’s for me focusing my next 10 weekly blog posts on AI”
  • “Create a table of the most essential Python functions in the LangChain library”

6. Ask the model more questions: create a feedback loop

  • Don’t expect a perfect answer on the first try. The idea here is to realize that the model is interactive and dynamic.
  • Keep asking the model questions that bring you closer to your desired outcome. This is so important.
  • If you see something you don’t like, tell the model and ask it to make corrections.

7. Provide the model with examples

  • This will allow the model to be guided by your expectations.
  • It is often referred to as one-shot or few-shot learning (one or few being the number of examples you give it).
  • For example, if you’re writing an essay about a lion, and you already created an essay that you liked about an elephant, give the model the essay you created about the elephant as an example and it could be very nicely mirrored.

8. Write prompts like code: a systematic approach

Think of prompts as carefully structured programs – each part builds on what came before. Here’s how to do it:

  1. Define Your Terms
    Start by clearly stating what key words mean in your context. Just as a programmer declares variables, you should declare your definitions.
  2. Build Your Logic
    Layer your concepts one at time. Each new idea should follow naturally from the previous ones, like functions in a program.
  3. Be Explicit About Context
    Tell the AI exactly how you want words interpreted. This avoids the ambiguity that plagues everyday language.
  4. Use Progressive Disclosure
    Start simple, then add complexity gradually – like good documentation. This helps the AI follow your reasoning path.

Example:

I'd like to understand computer memory. Let me define my terms:

- By "computer" I mean modern electronic devices that process information
- By "bit" I mean a single binary value (0 or 1) stored in computer memory
- By "memory" I mean the physical components that store these bits

Given these definitions, please explain:
1. How is a single bit physically stored in computer memory?
2. How does the computer move bits between different memory components?

9. Tell the model how to reason

  • Tell the model to think step by step. Tell it to be logical, precise, and clear.
  • Think about how a model would be going about this.
  • Is it grabbing your question and breaking it down into multiple prompts?
  • Be more specific so that you can make it better.

Examples:

Think step by step. Use logic along the way. Take the following instructions and break them down into multiple steps, and complete each step at one time. 
1. First, restate the problem in your own words to ensure you understand it
2. Break down the key components or variables involved
3. State any assumptions you're making
4. Think through your solution step-by-step, explaining your reasoning at each point
5. If you reach a conclusion, explain how you got there
6. If you spot any potential weaknesses in your reasoning, point them out
7. Suggest what additional information might help improve your analysis

10. Metaprompts: have models generate your prompts.

  • You can always ask your model to create a prompt for you. It has access to data that tells it what a good prompt is. Use that to your advantage.
  • Really clearly understand what you want from the model. Do introspection. Find silence.
  • Then, execute your Metaprompt.

Example:

Create a prompt for me that will help me understand component [x] of a computer. I want to understand the form of component [x] and its core functions.

PUSH THE BOUNDARIES OF WHAT YOU THINK IS POSSIBLE

Remember, Generative AI is fundamentally probabilistic… yet it has access to so much contextually rich data… and it can put it together in brilliant ways. Do extraordinary work, and if you have any additional tips, drop them in the comments below.