Lengthening the sections of a marketing framework such as AIDA, PAS, Feature-to-Benefit, or Before-After-Bridge

API
1

I am using DaVinci Instruct Beta to create ads using the AIDA marketing framework: Attention, Interest, Desire, Action.

I start out with my initial request for the Awareness portion:

INPUT

{
  prompt: 'Write a witty advertisement using the AIDA method for: Hush Blankets — Weighted blankets.\n' +
    'Awareness:',
  max_tokens: 19,
  temperature: 0.5,
  top_p: 0.2,
  frequency_penalty: 0,
  presence_penalty: 0,
  best_of: 1,
  stop: [ 'Awareness:', 'Interest:', 'Desire:', 'Action:' ]
}

Then I take the output from that and pass the output back in to get the next section.

{
  prompt: 'Write a witty advertisement using the AIDA method for: Hush Blankets — Weighted blankets.\n' +
    'Awareness: Hush Blankets are the perfect solution for those who suffer from anxiety, insomnia, and restless\n' +
    'Interest:',
  max_tokens: 99,
  temperature: 0.8,
  top_p: 0.8,
  frequency_penalty: 0.75,
  presence_penalty: 0.75,
  best_of: 1,
  stop: [ 'Awareness:', 'Interest:', 'Desire:', 'Action:' ]
}

First off, let me just say that I’m aware that it’s returning an incomplete sentence for the “Awareness” portion. I plan on fixing that by making another call to GPT3 until I get a finish_reason of stop.

I think that this is the best way to go about getting GPT3 to create an AIDA framework ad, but I’m having trouble whilst trying to lengthen the next portion:

OUTPUT

{
  id: 'cmpl-3Cie5Xrf1TsuAtxclBcoffEvhnAkc',
  object: 'text_completion',
  created: 1624160565,
  model: 'if-davinci-v2',
  choices: [
    {
      text: ' nights. Made of high-quality materials, the blankets are adjustable to your needs and fit all sizes\n',
      index: 0,
      logprobs: null,
      finish_reason: 'stop'
    }
  ]
}

The problem I’m facing here is that this text is just too short, and if I run it again, it’s going to give me the next portion: “Desire”.

So, I’m offering up my prompt in hopes that someone can help me lengthen the “Interest” and “Desire” sections of my AIDA framework.

Currently, I’m doing this by passing the output into DaVinci to create a “lengthener” as I call it.

LENGTHENER

{
  id: 'cmpl-3Cie7uVTQIjg6yUUP5i05sO8UvBIy',
  object: 'text_completion',
  created: 1624160567,
  model: 'davinci:2020-05-03',
  choices: [
    {
      text: '!!!\n' +
        '\n' +
        'The blankets are made of high-quality materials and fit all sizes. The blanket is',
      index: 0,
      logprobs: null,
      finish_reason: 'length'
    }
  ]
}

Now, the problem that I’m facing is that DaVinci always returns length as its finish_reason. Therefore, I’m assuming that I need to take that output and now pass it back into GPT3 to get a complete sentence with a finish_reason of stop.

Am I going about this correctly or am I just completely missing something? How should I go about lengthening the output of a particular section in a marketing framework?

1 Like
3

That totally makes sense! I do get some responses that contain other “instructions” with the colon at the end, such as “Arousal:”, so maybe if I taught GPT3 a little about the AIDA framework before asking it to create one for me, maybe that would help? I’m very new here, so any information I can get helps!

Thank you! :sweat_smile:

1 Like
5

Why are you breaking it up like that? Seems like a single prompt works just fine.

image
image1898×574 44.4 KB

1 Like
6

It works sometimes when you do it that way. It says “Decision” instead of “Desire”. The problem that I’m having is getting longer output for the Interest and Desire (not Decision) sections.

7

Ah, add “very detailed” to the instructions. Also consider a few-shot example.

image
image1664×354 47.4 KB

3 Likes
8

That’s pretty good. I can probably take part of that and pass it to Ada to get it to elaborate a bit. What do you mean by “few-shot”? As in passing the output back into the prompt?

Thanks! :innocent:

1 Like
9

That actually worked out really well! Thanks!

image
image2031×1169 264 KB

image
image2031×1169 299 KB

2 Likes
10

“Few shot” in this case is to just give it one or two examples to model off of. You will get more consistent performance but less creativity. Glad to help!

1 Like