This article will explain the basics of how Quil's AI works, and how to write out a prompt
How does Quil get it's information?
Before generating notes or follow-up material, Quil must have the context or information given to it beforehand. This can be done by having Quil's notetaker listen in to your calls, and by feeding the notetaker information through manual notes during the meeting.
How does Quil generate notes?
After getting context and information from it's notetaker, Quil's AI will then run a series of prompts though the information that was given to it (the transcript, and the manual notes). These prompts are selected by choosing a meeting type, each meeting type has it's own set of prompts that are specific to that type of meeting (e.g. interview, debrief, reference call, client intake)
Tip: Think about Quil's AI like a student that must a read through the transcript and the manual notes to then complete the instructions on a test, with the 'test' being the prompt's that were written out. If there is not enough information given to the AI, then the AI can not answer all the questions, or complete all the instructions given to it.
Prompt Engineering
CAUTION: Quil's notetaker utilizes AI technology, and users should be aware of certain limitations:
- Quil generates drafts that are approximately 85-90% complete and should be treated as preliminary.
- Despite the capabilities of the AI, there are inherent limitations that may not be overcome by prompt refinement alone. Users should verify accuracy and completeness independently.
- The quality of Quil's responses depends on the prompt, and on the content of the call
Now that we know how Quil works we will learn how to write out a very basic prompt.
Very basic prompts can be divided into 3 requirements to be successful.
Instruction/Question
Target
Details/Rules
We will use a simple question that the user would normally ask in call as an example:
"What is the candidates company background?"
Instruction/Question
The first thing that must be established in a prompt is a clear question, or instruction.
For our example question we could use several different words that would describe what we want from the AI
"What is the candidates company background?"
- What is...
- Define...
- Summarize...
- Explain...
All of these instructions/questions will give the AI a clear task to do/answer.
Note: Refrain from asking questions that the AI can answer with "yes" or "no"
(e.g. "is...", "does...", "can..."
Target
The 'target' in a prompt is the person who's going to be the focus for the AI's response.
In our example question "What is the candidates company background?" the 'target' specified is "the candidate" so the AI will use the information provided by "the candidate" to complete it's task.
The problem with the wording of this question is: the AI doesn't have enough context in the transcript, or the manual notes to know who the candidate is (when in call, the user won't normally mention "the candidate is XXXXX person" so the AI won't know who the candidate is. Also, not all calls have a candidate, sometimes it might me a client, a reference) so using the wording "the candidate" is incorrect.
The context the AI will always have, independent of the meeting type, or the type of call is who hosted the call AKA "the host" and who is participating in the call AKA "the participant".
Therefore the correct wording for our example question would be:
"What is the participant's company background?"
The example question now has clear instructions and a clear target.
Note: Depending on the call, sometimes the conversation will mention someone who is not in call, a third person, e.g. In a reference call there will be: the host, and the reference (the participant). In this call the host and the participant would speak about a third person who is "the candidate". In this case it is okay to use "the candidate", or any other title as the target, since the AI will have enough context in the transcript to know who the third target is.
Details/Rules
This aspect is the most important since it will define how the AI gives their response. To explain this we will look at some examples of the prompt and the AI's response
Prompt:
"What is the participant's company background?"
Response:
"XXXXXX worked previously at X, Y, and Z."
This response that we got from the AI is not the response the user would typically want due to the lack of details. With this in mind we will give the AI very specific details, and rules to follow for it's response.
Prompt:
"What is the participant's company background? For each company mentioned, specify:
Company name:
Compensation:
Title:
Reporting Structure:
Additional details:
Any information that was not mentioned, remove it from your response."
Response:
"Company name: XYZ
Compensation: 100
Title: worker
Company name: YZX
Compensation: 110
Title: staff
Company name: ZXY
Title: boss
Reporting Structure: xxxxxx
This response that we got from the AI has significantly more details, it follows a structure, and has additional rules to clean up the response.
These details/rules is how we get the most personalized and customized notes from the AI using the prompts. There are few limits to what rules you can establish in a prompt, if you find that the response from the AI does not fit your liking, try to change the wording of your prompt, or just adding additional rules to the prompt.
Tip: Details/rules work best when at the start of the prompt.
Tip: The less characters used in your prompt, the better the AI will understand your request
Tip: To get the best answers, make sure you ask all the appropriate questions in call. Use the script to help guide your questions in call.
To understand the best practices and specific rules for Follow-up Material, or Database Notes, please access their specific pages.
If you require additional help, reach out to your main point of contact at Quil or email support at support@quil.ai.