Prompt with GPT
How to prompt with GPT
Last updated
Was this helpful?
How to prompt with GPT
Last updated
Was this helpful?
The Generate with AI feature allows users to scaffold a prompt with natural language.
Simply click Generate with AI, choose the type of your task, and describe what you're looking to accomplish. Check out this guide to get started:
The system define the AI's persona, objective, specific tasks or rules.
Note: this definition varies from user-to-user and prompt-to-prompt. Too many System instructions, and the output may degrade - so feel free to experiment by adding the instructions in the User prompt!
There are no hard rules about what to include in the system prompt. However, we generally recommend including specific, well-defined tasks while avoiding extremely complex logic.
The user and assistant define the conversation between you (the user) and the AI (the assistant).
For example, in a conversation between you and chatGPT: your message is the "User" message and chatGPT's response is the "Assistant" response.
Using both the User and Assistant allows you to provide examples that help the model "learn" your desired output format (JSON, HTML, etc) and language
User-Assistant pairs (also known as one-shot or few-shot prompting) are examples we provide to the model to help it "learn" our desired response:
Add an example User message
Add the expected Assistant response
(optionally) Add more user-assistant pairs
Add another User message with your input
For a detailed guide, check out our tutorial on K-Shot Prompting:
System: Your task is to extract the unique transcript ID from the end of the URL.
User: https://app.fireflies.ai/view/AirOps-Intro::DkMo5PetdLeCDglp
Assistant: DkMo5PetdLeCDglp
User: https://app.fireflies.ai/view/Mdei9OrqpQlWIijd
Assistant: Mdei9OrqpQlWIijd
User: {{link_to_sales_call}}
Using more than one example is helpful for nuanced logic.
In our text extraction example, we provided two URL's with different formats to help the model understand how to extract text under multiple scenarios.
Use a real-life example: The best way to generate high-quality outputs is by providing a high-quality example to the model. Use the most realistic example when possible, whether from the internet or from your own data.
A user-assistant pair is highly recommended for GPT-3.5.
However, there are cases when you would not need to use a user-assistant pair.
To output in a consistent format when using GPT-3.5 (e.g. JSON)
To generate a specific language and style
For examples that can't fit into the context window (e.g. long transcripts)
For tasks that can be prompted in one-shot with GPT-4
You can provide examples to GPT-3.5 Instruct by providing "Input" and "Output" pairs instead of User and Assistant pairs.
System:
Your task is to extract the unique transcript ID from the end of the URL.
Input: https://app.fireflies.ai/view/AirOps-Intro::DkMo5PetdLeCDglp Output: DkMo5PetdLeCDglp
Input: https://app.fireflies.ai/view/Mdei9OrqpQlWIijd Output: Mdei9OrqpQlWIijd
Input: {{link_to_sales_call}} Output:
The last line is Output: because GPT-3.5 Instruct will then "fill in the blank".
GPT-4 Vision is a model that can analyze images. To use GPT-4 Vision in AirOps, you need to provide a link of a downloadable image to the model:
Add an LLM step to your workflow
Change the model to GPT4 Turbo Vision
Add an Image URL to the User message where it says "Add an Image URL"
To upload a single image from your desktop, change the input type to File Media which accepts a .jpg, .png, or .webp file
AirOps will return a downloadable URL to this image
Pass the input as liquid syntax to GPT-4 Vision where it says "Add an Image URL"
To upload an image from Google Drive, choose Short Text as an input type
Change the Google Drive image so anyone with the link can view it
Add a code step to transform the Google Drive link into a downloadable URL replacing short_text_input with the name of your input:
Pass the output of the previous step into GPT-4 Vision where it says "Add an Image URL"
For a more detailed guide on GPT-4 Vision, check out our tutorial here: