Beyond Hot AI Prompts

I’ve been working with Max Tokens, Temperature, Top-p, Frequency Penalty, Presence Penalty parameters for ChatGPT to take my content creation prompts to the next level.

However, despite my online research, the best suggestions come from asking ChatGPT about my use cases.

For example, I remove similar results, leaving only the most interesting.

Provide in plain text English suggested ChatGPT settings for `{options}` that would be relevant for analyzing content quality using heuristics to eliminate the bottom 50% as an unbiased content analyst.

Briefly explain why those suggestions.

{options} = "Max Tokens, Temperature, Top-p, Frequency Penalty, Presence Penalty"

The response:

1. **Max Tokens**: 512

   - **Why**: By setting a limit on the maximum tokens, you ensure that the analysis stays within a reasonable length, focusing on key quality aspects without getting bogged down in unnecessary details.

2. **Temperature**: 0.7

   - **Why**: A temperature setting of 0.7 will allow some creative variation in the analysis without drifting too far from the most probable path. It offers a good balance between structure and exploration, allowing the model to identify unique patterns in content.

3. **Top-p**: 0.8

   - **Why**: This setting filters the token selection to the most probable tokens, ensuring that the analysis stays relevant to the topic. A value of 0.8 strikes a good balance between keeping the analysis coherent and allowing some flexibility in expression.

4. **Frequency Penalty**: -0.5

   - **Why**: A negative frequency penalty discourages the model from using overly common and repetitive phrases. This can aid in analyzing content quality by focusing on more specific and nuanced aspects rather than relying on broad generalizations.

5. **Presence Penalty**: 0.2

   - **Why**: A slight presence penalty encourages the model to utilize information that is present in the context more effectively. It helps in closely aligning the analysis with the specific content under examination, enhancing the precision of quality evaluation.

From above, I update my prompt as follows.

// ensure that the analysis stays within a reasonable length, focusing on key quality aspects without getting bogged down in unnecessary details
Max Tokens: 512

// good balance between structure and exploration
Temperature: 0.7

// good balance between keeping the analysis coherent and allowing some flexibility in expression
Top-p: 0.8

// focusing on more specific and nuanced aspects rather than relying on broad generalizations
Frequency Penalty: -0.5

// closely aligning the analysis with the specific content under examination
Presence Penalty: 0.2

When I want to write a few hundred words as a personal essay, I use the following ChatGPT parameter settings.

// Detailed, personal, and engaging essay
Max Tokens: 400

// Creative and interesting but still on-topic and coherent
Temperature: 0.8

// More diverse and less predictable text
Top-p: 0.875

// More unique and tailored to me
Frequency Penalty: -0.25

// Key themes are present without being repetitive
Presence Penalty: 0.1

And for creating shorter social media posts, these ChatGPT settings help greatly.

// Short and concise
Max Tokens: 80

// Creative and engaging
Temperature: 0.7

// Balance of focus and diversity
Top-p: 0.8

// Unique expressions
Frequency Penalty: 0.4

// Maintaining focus with some new ideas
Presence Penalty: 0.5

What’s the Temperature?

For the Temperature parameter, setting it around 0.7 to 0.8 serves me well in creating variety for a topic. And closer to 0.3 to 0.4 keeps the variations tighter for fine-tuning the content.

For example, a 0.7 temperature hashtag prompt for this article suggests.

#ChatGPT #AIWriting #DigitalMarketing #ContentCreation #SocialMediaPosts #CreativeWork #MaxTokens #FrequencyPenalty

And a 0.4 temperature hashtag prompt for this article suggests.

#ChatGPT #AIParameters #DigitalMarketing #CreativeWriting #SocialMediaContent #UniqueExpressions #EngagingContent #AIInnovation

That 0.7 temperature offers several directions, yet the 0.4 temperature provides a tighter association with this article.

So for content variety, set a higher temperature. Yet, once you have a suitable temperature, we must work with Frequency Penalty and Presence Penalty.

What’s a Penalty?

A penalty parameter controls how creative or repetitive a response is. In effect, how often are there common, rare, or unique words and phrases in a result?

Frequency Penalty is about how often common or rare words are used, and Presence Penalty is about encouraging or discouraging the use of unique words throughout the text.

It means that when you look to AI prompts that speak in your voice, trend the Frequency Penalty into the negatives and keep Presence Penalty positive, yet close to zero.

And when you want to be expressive and try new things out, raise those parameters above zero.

Refer to the personal essay and social media post settings above as examples.

Why Care about ChatGPT Parameters?

ChatGPT parameters are how the prompt of “Rewrite the following content.” remains in your voice versus that of ChatGPT.

More importantly, there’s general growth from meh to good and even wowza.

So what has been your experience working with ChatGPT parameters to create impressive results?

ChatGPT Parameter Defaults

// This setting allows for long and complex conversations
Max Tokens: 4096

// balance between randomness and determinism in the generated text
Temperature: 1.0

// all possibilities are considered, leading to more diverse outputs
Top-p: 1.0

// no bias towards or against frequently or infrequently used words
Frequency Penalty: 0.0

// no penalty or reward for using new words
Presence Penalty: 0.0