Using Prompts
How to approach Comment Cat Engine prompting
Prompt Hierarchy
The Comment Cat Engine (CCE) uses a 3-prompt approach. The prompts are System, Channel, and Job, and they are explained next.
System Prompt
The system prompt is universal and sent with every request. The CCE team manages this prompt, and the intent is 2 fold:
Inform the AI about the role (e.g., you are an assistant performing sentiment analysis of social media comments).
Enforce uniform output via tooling functions and prompt instructions to allow the system to reliably integrate with the CCE API and process the returned result sets.
This prompt is quite generic and is intended to establish a baseline for sentiment analysis on the data.
Channel Prompt
The channel prompt is only part of subscription services.
The channel prompt is controlled by the user and is required. It explains “what your channel is about” and provides the analysis with context.
A prototype example prompt for Joe Rogan would be (238 characters in length, FYI):
“The Joe Rogan Experience is a long-form podcast with a wide range of guests from politics, science, culture, and sports. Joe is a martial arts expert, heavily engaged in Brazilian Jujitsu, and has a history in mainstream television shows.”
Job Prompt
The job prompt is used for both subscription and "one-off" processing.
The Job Prompt is optional for subscription plans but required for one-off processing. This prompt is available to provide per-video context. It is helpful in cases where context, bias, or experience is important, and a lack of this additional prompt information could skew the sentiment analysis.
A rule of thumb for Job prompt use:
Does the guest, topic, or host have a unique or nonconforming view?
Is there a polarizing topic where the sentiment should be aware of the views (politics, religion, controversial issues, conspiracy theories, nonconforming opinions)?
Are there elements of the content that might be misleading, confusing, or ambiguous (e.g., parody, sarcasm, creative critique) that may need guidance when performing sentiment analysis?
An example template for a Joe Rogan episode follows:
“The guest is a [description] who is an expert on [topic] and explores how [controversial topic] is misunderstood and posits how the [alternate view] of [short explanation] is not considered despite evidence and research”.
Prompt Truth
It is well known that AI and LLMs are trained on fixed, biased datasets. Commentary that goes against the grain of the training exists and is not always on the fringe.
People may try to tackle sensitive issues in unique ways, or involve characters that may be misrepresented in a standard training set, and will benefit from the Job Prompt.
If the channel risks interference from the biased nature of the LLMs and training data, the channel prompt may need to be customized to deliver the correct angle on the channel videos.
The following table outlines when we use these prompt types in subscription and one-off processing.
Prompt Type
Subscription
One-Off Processing
System
Always sent
Always sent
Channel
Required
Not Available
Job
Required
Optional
Tips Prompt Writing
The prompts necessary for the CCE are similar to your good'ole GPT prompt; however, the CCE prompts are much more focused. Here are some tips, tricks, and guidelines for prompting.
Keep the channel and/or job prompt to 300 characters or fewer. Prompts should be short.
Do not "set up" the prompt to be an assistant doing a job; we have that covered.
Do not try to jailbreak the prompts; that will just return bad results
Keep the prompt very focused on any bias or unique view that should be known when processing comments for sentiment analysis.
For example, conservative political issues, you will want to acknowledge the inherent bias of the program, host, guest, or channel.
Remember that the prompt resets after each use. Previous knowledge is not carried over; the LLM does not remember your other videos or the data it has processed.
Do not overthink the prompt! Keep it simple, to the point and as minimal as possible
Larger prompts distort the results; do not exceed approximately 300 characters.
Last updated