If you’re going to use generative AI models, you’ll need to think about your prompts. This has brought forth a new career opportunity; 'prompt engineer' is now a job.
Why? Because the better the prompt, the better the output.
6 components of a good prompt
Jeff Su, a Product Manager at Google, recommends using these 6 components for your prompts.
You don't need all six components in every prompt, but in order of importance:
| Mandatory: Always start with an action (verb) and include what output you want. Verbs - brainstorm, change, cite, draft, edit, optimise for SEO, proofread, shorten, |
| The more context, the better. Think about the implicit context you know. Make it explicit for the AI tool. Who is the audience? What’s the purpose? Use these 3 questions to consider how much context is needed:
|
| Including examples within the prompt drastically improves the result |
| Who do you want the AI to be answering as (only use real people if they are very famous) |
| Define the output. Close your eyes and visualise what you want the end result to be (email, blog, report, code, bullet points, paragraphs, markdowns) |
| How you want the other person to feel, tone of voice, how casual vs formal |
Other tips and tricks
🙏 Be polite: Idk why this works but apparently it does. Some clever sounding research here. Be polite, but not too wishy-washy. Please, thank you, good response - all improve the results you receive!
**Worst case, when AI becomes our overlords it'll remember you were one of the nice ones!
📰 Ask for sources: You'll get a lot of information back, but how do you know it's true? Ask your GenAI to provide sources and links for all the facts it responds with. Then check those sources!
🟥Give constraints: Is there a word limit, a tone you want or specific style guidelines the response needs to follow?
🔁 Iterate, iterate, iterate: If at first you don’t succeed, prompt, prompt again! Refine the prompt by giving more detail or revising specific areas.
🤔 Ask for multiple options: You can ask for multiple versions or options of your product, which gives you choice.
🤏 Break complex requests down into parts: This is referred to as ‘chain-of-thought prompting.’ Divide a larger task into more manageable steps that you can review and refine before it moves to the next step.
🧠 Stay open-ended for brainstorming: To get diverse ideas, use open-ended prompts without much guidance (AKA ignore much of the above).
Now what?
Two things you can take forward:
Review your previous prompts.
What components did you cover? What could you do more of?
Create a library of prompt templates you and your team could use.
This means you don't constantly need to retype the same information. It also means you can prompt like a pro - and craft a great one. If you have recurring tasks or events, having a template is beyond helpful!
role/work context: I'm working as a [role] at [organisation]. In this [role/team] my responsibilities are [roles/responsibilities].
Piece of work context: As a [role] I've been asked to [work you want AI support with]. Could you please help me [verb] [output]. Here's the context. The goal of this is to [goal]. The audience is [audience - who, internal/external, knowledge levels]. The language tone should be [tone] and the format should be [format].
Here's an example of one from The Training Practice:
PS - I didn't use this prompt to write the March event blurb back in December. But I will use it for future events!