Thanks Jeff for your time and further explanations. This is indeed a great example for blogpost generation.
Somehow you are still missing the point which was to achieve copy.ai results or better. As creating whole blogposts is not currently a feature of copy.ai that still remains an open exercise. You don’t seem to be familiar with copy.ai features.
Before you reply back with another “how to generate a whole blogpost example” let me give it another try using the information you provided already.
Thanks again for your patience and willingness to help.
Hi Gerard,
I assumed copy.ai was another of the generic blog generators.
I just checked their website and all of those features can be accomplished using a single prompt, and it’s a much easier job than a custom article generator.
In fact, with a few hours work you could do a much better job than all of the generic copy.ai templates.
I recently created a prompt for a client that responds to his emails in his own personal style. It took approx 16 hours to build / test the prompt and it has approx 95% success rate (he only needs to press re-generate 1 in 20 emails).
You can 100% easily replace all of those copy.ai templates with a single custom GPT3 prompt and achieve much better unique copy!
Good luck.
Jeff
1 Like
That’s amazing! Looking forward to take my prompt skills to your level.
I’m just a novice prompt engineer 
Gerard.
Hi Gerard,
GPT3 already knows just about everything.
You simply need to tease the knowledge out of it!
It already knows all of the copy.ai template structures (or similar).
If you play with GPT3 creatively and get a “feel” for it I really cant think of many things you will be unable to accomplish.
Ignore anything you read from before 6 mths ago - it’s mostly irrelevant when playing with prompts for the latest large language models.
The new models seem to behave totally differently to all of the others!
I am mostly using Davinci 2 (Temperature 0) to interactively create prompts.
Once built you can decide which parts of your prompt to designate as variables to be passed to the API.
At run time I use Davinci 2 (play with Temp/freq/presence) or legacy Davinci (seems more creative).
When you have something working on Davinci, you can simply test the same prompt on smaller/cheaper models to see if you still get the output you need.
20% of the time you will.
2 Likes
I put together a prompt that gave me some good results. This is the final version which I tried in for a few different scenarios.
You are an excellent copy editor. Your copy has won many awards and accolades. You don't require a lot of direction and will make the right decisions given the information provided.
++This is the text you receive as input:
<<INPUT>>
++Copy direction:
Instructions: <<INSTRUCTIONS>>
Tone: << Friendly | Luxury | Relaxed | Professional | Bold | Adventurous | Witty | Persuasive | Empathetic>>
++This is your proposal:
Thanks for any feedback or improvements you foresee.
Gerard.
3 Likes
Nice job!
If you want it to write in YOUR style, simply include 3 full examples below your prompt.
Then use the WHOLE prompt inc examples + inputs for new copy required.
You can tune that until you get outputs that look like they were written by you.
Another option is to create a prompt that receives your raw incoming communications as input & auto creates the reply.
You can combine that approach with a vector database containing your knowledge base if your communications are related to business. I am working on a test for the vector database idea at the moment.
1 Like
Thanks!
I’ll try that. Style transfer using prompts!
What do you mean by vectors? Do you mean collection/list of sample communications in different contexts?
Good luck with that experiment!
Gerard.
Yes, style transfer with a “few shot prompt” works very well.
Ref vector databases, it is a newish approach to storing info - which can be searched semantically and the search results used to generate AI content.
I am currently beginning testing this approach for auto generating a specific subset of business email responses by referring to the vectorized business knowledge base.
This is quite an interesting video exploring using Open AI with Pinecone vector database: Beyond Semantic Search with OpenAI and Pinecone - YouTube
1 Like
Oh I see. Got it!
That seems a lot of work to get around the limitations around openAI API restrictions.
Wouldn’t be easier if openAI increased the working memory for the models?
Gerard.
Vector databases facilitate 200ms semantic search on a trillion documents.
This opens up possibilities that would not be possible with an expanded working memory.
The improved functionality of Davinci 2 combined with the 4k token limit does enable you to do a LOT that was previously impossible though, especially for your copy.ai project which can be implemented using custom few shot prompts.
1 Like
If we had access to the model this wouldn’t be necessary so it’s bound to the restrictions imposed by openAI API one way or another.
Nevertheless I agree on your statement although it’s not a very common use case to have trillions of documents at your disposal it’s enabled by a vector database.
Let’s make the best of the current situation as it lasts.
If you can share 2 or 3 copy ai results and the input that is provided in their tool to kick off their generation, it would be fun to try to engineer a prompt that is competitive with that…
1 Like
Hi Jeff,
I hope you are doing well! 
I’m relatively new to this, and I’m on the path of learning it… but I have pretty specific needs, which I guess would take a bit of time to accomplish. I was wondering if you’d be open to some collaboration? Or even taking me as one of your new client
Let me know what you think!
Thanks,
Pete
Hi Pete,
Hope you are doing well too!
Interested to hear more about your specific needs.
Many projects can be accomplished quite quickly with the new AI models
Feel free to send me some further info or we can schedule a chat if you prefer.
Thanks
Jeff
Hey Jeff,
Sounds great! I don’t think here we can send private messages, so any chance you have discord? Or what channel would you recommend? Would be great to chat about it.
Pete
Lots of great info in this thread Jeff - thanks for that!
I’ve been playing with your prompt example to generate an entire article & had a few questions I hope you won’t mind answering.
-
Where does your prompt end & the output begin? It seemed to work best for me if the end of the prompt was “First create an 11 paragraph article outline:”
-
I could not replicate the full output you got in your example. Did you get that entire output from a single request?
For me, most of the time I will just get the outline, and then I will run it again, adding the first outline section heading that it generated to the end of the prompt, and then GPT-3 will finish the article, although it still didn’t replicate your results as robustly.
I usually either get a list of numbered point or a paragraph or two for each section, but not both.
Up until now, I didn’t think I could get long form content like this without few-shot prompts or a fine-tuned model, which are obviously more expensive. So thank you for opening my eyes to better prompt engineering
Hi Jim,
I build the prompts interactively - Davinci 2 - T= .7 Freuency/presence penalties both set to approx 1/3rd of their max value. These are starting values and then you play with the parameters as you test.
Then I type something like:
“I need an expert to demonstrate how GPT3 can be used to write an article at least as well as Copy.ai.” and hit submit. Davinci will then generate some text which I either accept, or edit to steer the prompt creation in the direction I wish. It’s an interactive process to allow Davinci 2 to partially guide the content creation & prompt evolution using it’s own knowledge & patterns.
When your prompt is complete to the point it will write a full article, then you can identify the sections of the prompt that can become variables (to be inserted at runtime to create an unlimited number of articles on different topics).
First, use interactive generation to create prompt, then use the created prompt with the variables inserted to determine article titles/content for subsequent articles.
Does this make sense?
3 Likes
