Ok, back to today’s topic.

Spoiler Alert: This issue is tailored for ChatGPT, but you can use it for more.

OpenAI recently released six prompting strategies and tactics to get better results from ChatGPT. These strategies will help you make the most of ChatGPT and ensure you get clear, accurate, and valuable responses.

By the way, although they are recommended for ChatGPT, you can also use these for any other conversational AI powered by LLMs.

Let’s get into it!

1. Write Clear Instructions

No Shit.

But actually - are you doing this at the moment?

If you’re not clear, AI won’t be able to provide you with a specific outcomes. It can’t read your mind!

• Add more details to your questions for better answers. If you are clear about what you want, you are more likely to receive it.

• Ask the model to take on a specific role. Choose formal / informal language options.

• Use separators to mark different parts of your input. Don’t just word-dump.

• Outline steps to complete a task. For example, choose to answer a question using the STAR method.

• Use examples.

• Mention the desired output length (believe it or not, you can count words - it might not be 100% but at least the output will be reasonably accurate).

For example, instead of asking:

“How do I add numbers in Excel?”

Ask:

“How do I add up a row of amounts in USD, automatically, for a whole sheet of rows, with totals in a ‘TOTAL’ column to the right?”

2. Provide Reference Text

Don’t let AI invent anything. Give the right references. Especially for niche topics.

• Ask the agent to use the provided reference / reference from your chat history to formulate a response.

• Ask the agent to include citations from the given references in the text replies.

For example, ensure to include relevant company information, text references, examples, etc in your prompts.

3. Split Complex Tasks into Subtasks

Breaking down complex tasks improves accuracy.

• Find the right instructions for user questions using intent classifications

• In long chats, shorten or filter past dialogue

• Create a full summary by summarising long documents

Example, if you are building a marketing plan, break it down in sections:

• Analyse target audience

• Identify key messaging

• Suggest marketing channels

• Outline content strategy

• Propose budget allocation

4. Give the Model Time to “Think”

Let’s test your maths. What is 17×28?

It probably took you 1-2 mins to calculate the right answer in your head (if you didn’t just guess or make a mistake, that is). Don’t hate me.

Either way, the point is, the same happens to LLMs. If they try and respond immediately, they can make more mistakes.

Did you know you can ask ChatGPT to think through the answer?

Me neither. But here’s how:

• Instruct the agent to think carefully through the solution before providing an answer.

• Use a sequence of questions or an internal monologue to guide the model’s thought process.

• Ask to review previous responses to see if anything has been overlooked.

“Think through this step by step” or “Let’s approach this systematically” are good options.

5. Use External Tools

ChatGPT is great, but it also has a lot of weaknesses. Good News though, these can be supported by external tools.

If a task can be done more reliably or efficiently by a tool rather than by a language model, offload it to get the best of both.

• Use code execution for more precise calculations than APIs

• Give the model access to specific functions 

• Create images or browse the internet with different tools

For example: for complex data analysis you might ask the agent to write Python for processing the data, then use a code execution engine to run it and feed back the results.

6. Systematic Testing

This is more relevant to actual prompt engineers that require consistency and reliability of prompts and outputs - and not absolutely necessary when interacting with ChatGPT on a day-to-day basis. But let’s still cover it for the sake of completeness.

Sometimes, when making a change to a prompt it might better or worsen performance of the agent. To ensure that results better performance, it might therefore be best to establish a comprehensive test suite.

• Refer and compare outputs to gold-standard answers. These are expert-created, ideal responses to specific prompts. Like the “perfect” answer to an exam question.

For example, create a set of diverse prompts covering various scenarios in your business, and regularly test the AI’s responses to ensure quality and accuracy.

By applying these six strategies, you’ll be well on your way to mastering prompting for day-to-day use.

Remember, practice makes perfect, so don’t be afraid to experiment and refine your approach.

Happy prompting!

If you are not a member of the AI Leadership Forum, join today and learn even more on prompting and effective use of ChatGPT.

This Week in AI

• Apple’s highly anticipated AI features (Apple Intelligence) will be delayed - supposedly now rolling out in October. Although they did release a teaser…;

• AI is taking over the 2024 Paris Olympics, with AthleteGPT, AI powered tracking, talent scouting, personalised viewer updates, and more. If you want to watch something specific, just chat with this GPT here! ;

• DeepMind’s new AI won a silver medal at the (math) Olympics.