The transcript was the main input coming into the process, and this was the data that needed to be manipulated into different forms. Therefore, the solution lay in manipulating that data with six different meta prompts for each ChatGPT module. Originally, making this a linear process was causing problems as the LLM could not handle such a vast amount of work and it was costing a lot of tokens. In order to achieve a 'good' output, I mainly drew on few-shot prompting to provide lots of examples of the desired output.

This meant each AI module could closely align with the brand voice and tone and other viral examples of thumbnails, so it had a strong model of reference to design its output. Additionally, it provided emails to the audience about the research in the script (the original input).

The automation takes under one minute to complete and saved them on average 60 mins, when posting long form content. That valuable time saved is reinvested into their content strategy, which ultimately helps drive revenue for their company.