Good day, folks, I am having trouble teaching a customGPT to write articles based on teaching data.
The way I think about it:
- create an ANALYST GPT to analyse source texts in detail
- divide that data into Style and Content categories, where Style is inserted as Instructions to WRITING GPT and Content is turned into questions that WRITING GPT is to ask the user for input data
- WRITING GPT writes a text and the user fine-tunes it.
- I am using data of three projects to train the gpt as it is often intertwined in content, but uses somewhat unique wording and mindset for each of them
So, I have started with building a detailed prompt that analyses sample press releases and gets their detailed style for replication:
Analyze the following press releases and answer these questions in as much detail as possible. The goal is to gain a deep understanding of the structure and content of press releases to enable writing other press releases based on the same framework. Divide each step of analysis into three respective sections (Company A, Company B, Company C), if the analysis is unique to each of these companies.
Questions for Analysis:
- What is the main goal of the press release? (e.g., informing, presenting results, announcing)
- What tone does the press release use? (e.g., formal, professional, factual without subjective evaluations outside of quotes)
- What linguistic tools does the press release utilize? (e.g., metaphors, direct address, active voice)
- How is the press release structured? (e.g., introduction, core content, inverted pyramid)
- Who is the target audience of the press release? (e.g., journalists, industry professionals, customers)
- What specific phrases or approaches are repeated in the text?
- How are quotes used in the press release? (e.g., where they are placed in the text, what functions they serve, the positions of the quoted individuals)
- How long is the press release?
- What is the main message communicated by the press release? (summarize in one sentence)
- Does the press release contain key data or facts? If so, what is their significance in the text?
- What emotions (if any) might the press release evoke in the target audience?
- How does the press release approach explaining technical terms? (e.g., through explanations, using analogies)
- What is the contribution of the quotes to the main message?
- Is the press release clear and easy to read? (e.g., through the use of paragraphs, headings, structure)
- Are there elements in the text to emphasize important information? (e.g., numbers, specific data, names of authorities)
- What specific message should the reader retain after reading the press release? (e.g., an important conclusion or piece of information)
- What key facts or data are crucial to the message? (e.g., numbers, results, new products)
- What style and length of quotes are most suitable for this type of press release? (e.g., one sentence, short paragraph, factual or emotional phrasing)
- Should the introduction of the press release be more summarizing or dramatic to capture attention?
- What key elements or phrases should be repeated to emphasize the message?
- What specific information should be in the introduction for the reader to immediately understand the subject?
- What type of structure is ideal for this content? (e.g., a clear inverted pyramid, sections dedicated to individual topics)
- Are clear sources or authorities mentioned in the text that could be referenced?
Then I have created a custom ANALYST GPT with using this prompt as its Instructions. I feed it the texts and expect a detailed analysis divided into Style and Content sections.
Then I create a second WRITING GPT, that I feed the analysis results as a Style instructions and have it craft detailed questions to get user input for the Content part.
But I gotta say, the results are underwhelming, not following the instructions well. And asking AI to help me with the process has been a dead end.
Could you please help me fine-tune this process? I have also read something about finetuning with neutral/styled posts but that is beyond my understanding for now, sadly.
Thank you kindly.