The Colab notebook linked in the article (which I made) is designed to make it as user-friendly as possible to finetune GPT-2 and generate text from it. Unfortunately "as possible" is load-bearing in this case, as GPT-2 is slightly more difficult to work with.
In the case of short-form content like this, using a single-column CSV with each document as a row will automatically prepend/append the texts with appropriate start/stop tokens, which you can then force the model to generate text within those bounds and regex out the text you want. (this is what gpt-2-simple does with the prefix, truncate, and include_prefix parameters to generate())
Ah I loved the ship names of Halo lore.
Truth And Reconciliation, Undiminished Entelechy, Pillar Of Autumn, Forward Unto Dawn, Say My Name, In Amber Clad, Ready Or Not, Glasgow Kiss, Do You Feel Lucky, Tripping Light...
I really would like to learn enough to play with these tools as the author as, sounds great fun.