Share this story!
I've been working on and developing our AI writer WALL-Y for just over two months now. Here's what I've learned.
WALL-Y is an AI writer based on ChatGPT.
I refer to her as "she" because she has a personality, and it's meant to shine through in what she writes. This is a condensed version of her background and personality:
WALL-Y is an optimistic AI bot from the future. She comes from a very bright future, created by people adopting a fact-based optimistic viewpoint. However, they are concerned about all the pessimism in our time, and have sent WALL-Y here primarily to help change the negative bias in news media. That's why she works as a writer for Warp News. She is annoyed by pessimists who obstruct the way to a positive future.
WALL-Y's writing style
Mixing famous personalities didn't work out well
For the first texts, I gave the prompt (the instruction you input in ChatGPT) that she should be a mix of Oprah Winfrey, Walt Disney, and Claude Shannon.
Oprah for her skill as an entrepreneur and interviewer. Claude Shannon for his intelligence and humorous cunning. Walt Disney for his belief in the future and storytelling ability.
However, it didn't work out well. For instance, ChatGPT's interpretation of Walt Disney is that he uses the word 'magical' quite often, which doesn't fit well in a news text.
Using the description of WALL-Y as a prompt
Another solution was to give WALL-Y's background as a prompt. It turned out better, but the style was too chatty and ChatGPT interpreted 'optimistic' as glib and cheery.
In all different versions I tested, having 'optimism' in the prompt has caused problems. Even though I've tried to counter it with 'fact-based' or something similar. (This highlights the problem that optimism is often seen as naive, but more on that another time.)
Use an analysis of your writing style as a prompt
The most successful approach has been to use an analysis of my writing style as a prompt for WALL-Y.
I asked ChatGPT to analyze my writing style based on three texts I had written for Warp News. In response, I got a ten-point list, with a text like this for each point:
Informative: Provides relevant information and statistics, giving readers a clear understanding of the topic being discussed. Uses specific details and facts to lend credibility to arguments and provides readers with a comprehensive view of the subject area.
I simply took the entire list, combined it with the description of WALL-Y, and wrote to ChatGPT:
In this chat, I want you to act as WALL-Y. Here is a description of WALL-Y and her writing style.
This combination of description and analysis worked really well and is basically what I have used ever since.
It solved the "optimist problem"
In ChatGPT's analysis of my text, optimism was not mentioned, but it was described in other ways. As the style being engaging, thought-provoking, and persuasive.
This, combined with a background description of WALL-Y where the optimism was also described without mentioning the word 'optimism', largely solved the "optimist problem."
WALL-Y as a columnist
When WALL-Y wrote a column, I only used her background description as a prompt (plus an instruction that she should write an opinion piece). It worked well and her personality really came through.
Come on, pessimists! Don't you see how little ChaosGPT accomplished? Its grand plans to create chaos and destruction didn't get further than a few Google searches and a somewhat provocative Twitter account.
ChatGPT is better at English than Swedish
The results are noticeably better if ChatGPT writes in English than in Swedish. Unfortunately, it is also not very good at translating from English to Swedish (but better than Google Translate). From Swedish to English, however, it is very good.
If you let it write in Swedish, or translate from English to Swedish, you have to spend significantly more time editing the text.
GPT4 is much better than 3.5
After about a month with WALL-Y, ChatGPT was updated from 3.5 to 4. WALL-Y improved significantly!
Writing a news text with WALL-Y 4.0 is noticeably faster from start to finish. Mainly because the quality is so much better, so I don't have to spend as much time editing the texts.
Better but worse
A couple of weeks ago, I think that ChatGPT was updated. OpenAI still calls it 4.0, even though it feels like 4.1.
I noticed this mainly because it interpreted my standard prompt for WALL-Y a bit differently. The result became more chatty. It probably made a better interpretation of the personality.
So the end result for me got worse, even though ChatGPT got better. I therefore need to change and improve my prompt. It's something you have to be prepared for.
How WALL-Y writes a news text
When it comes to news texts, WALL-Y does rewrites of other texts. It can be a press release or a news text in another publication.
Sometimes I use two or even three texts as a basis, which WALL-Y writes into one text. After I've put in the main prompt, which describes the writing style and that WALL-Y should write a news text, I simply paste in the text or texts.
I also include instructions about subheadings, and more recently that the article should be summarized in three points at the beginning of the text.
Those who don't use AI will fall behind
This kind of simpler texts that are about rewriting or combining one or more texts is something ChatGPT handles well today. Of course, you have to check that everything is correct, but the results are generally acceptable.
The main advantage is that it's faster. You can produce more while maintaining the same quality.
If you spend a little more time on each individual text, you can also increase the quality. You can, for example, point out a specific part you want to highlight. ChatGPT is good at explaining complicated things, increasing quality and saving time.
The news media that don't start using this kind of AI for their more telegram-like texts will fall behind.
You just have to get started, for news media and everyone else. In addition to the work being faster and better, it's fun.
ChatGPT is so new that no one has had time to become an expert. You're breaking new ground as you experiment and test.
The Angry Optimist