So chances are, you’ve landed on this post from another article I wrote – a review of the indie video game Cocoon. Except, I say ‘wrote’… I had a lot of help from ChatGPT.
Let’s state at the top: it was a mixed success, at least in terms of how happy I was with the output. The prose is virtually ‘perfect’, but – in that uncanny valley way – it seems less ‘me’ and a bit too clean and deliberate. There are, naturally, some huge benefits too, but more on that later.
I’ve been meaning to try an experiment to write an entire article using next-gen tools for a while, but want to test it on something fairly generic and superfluous, rather than a heartfelt opinion piece – and having played Cocoon over the Christmas break, it seemed a good fit to review it, but let AI do the hard work.
The writing process is an interesting one to detail. At first it was incredibly similar to my usual approach. That said, I have introduced an additional step to my article outlining over the last few months – the use of voice notes.
As an aside, this is something that has been on my radar for about that last 15 years ago (I’m fairly sure I remember Paul Boag getting excited about it at the time) when Apple first launched Siri-to-text. At the time, I had this dream of narrating multiple articles during my long car commute, not that it lasted long.
Clearly, neither Siri, nor the third party tools of the era, were good enough to interpret dictation successfully enough for that, and in fact it’s only in the last few years that iOS’s native voice input has passed the threshold of being genuinely useful.
So now I tend to first dictate the majority of my thoughts when starting an article’s ideation, then later reorder and remove duplicated thoughts manually in the output text. That allows me to get ideas out quick enough, whereas traditional written transcriptions may lose trains of thought in the midst of typing, and necessitate more outlining or earlier brainstorming to avoid this. This itself can incur very heavy-handed sub editing. Of course, this is where AI comes into its own.
ChatGPT is, of course, still in its relative infancy, and does need manipulation to get results. And slightly more annoyingly, this has to happen in its 1:1 conversational pattern. This was my process for this case:
1. Told ChatGPT I was writing an article, and that I would like to mimic the style of richpaul.co.uk
[ChatGPT responded with a synopsis of the website and style]
2. Gave it a list of specific articles to concentrate on replicating. I also noted that despite using other (not me) ‘Rich Paul’ materials in its sources, I wanted to just use this website.
[ChatGPT responded with a synopsis of each article]
3. Entered the main prompt:
*“I would like you to write an article about the game Cocoon, on Xbox Series X. I would like the article to be around 600 words. It should open with a brief description of the game, explain how the mechanics and learning process works, examine some positives and negatives, and finally conclude with a nice reference to something mentioned earlier in the article, while summarising the main points. I would like you to use the following notes and rewrite them (using the style mentioned above):”*
At this point ChatGPT provided a near-perfect article. As is its usual format, it presented it in headed sections, so I had a couple of follow-ups to change this styling, as well as its use of US English, and finally asked it to be one step more ‘personal’, as per the sources.
And that was it. Done. In under a minute, with the main prompt respond taking around 10 seconds. It was so good I even said “thank you” as if I was having a natural conversation.
When you step back and think about it, it’s pretty incredible.
This speed and effort saving does come with a certain trade-offs, however. For me, the piece feels like something you would write for a formulaic academic task, rather than natural prose. And while the ending is perfect in tone (as requested by my prompt), it’s a bit too literal about the video game itself.
Overall, the balance of sentiment, across positive and negative critique is also slightly off – this is, of course, a real nuance of writing.
Importantly, LLM’s also have one other major flaw – without fine prompt detailing to the contrary, most tweaks will cause the AI to rewrite the entire thing, which can be frustrating, and even working around this, you still have to recheck it all each time, just in case of errors.
A possible solution would be to use a smaller agent such as Fixkey to just edit certain sections, but even this is time consuming.
There were some other notes of interest from the whole process:
- While obvious, it still surprised me that it wrote in the first person, as me
- I was really impressed to see that, stylistically, the LLM copied my use of em-dashes – which are incredibly overused in my writing.
- When requesting it rewrite the whole piece in a more personal tone, it went as far as fabricating things that hadn’t actually happened in my gameplay, nor had included in my notes.
- More comprehensive notes and relevant instructions would have made a huge difference, but are a massive trade-off between writing a longer prompt including all of that, and the speed of the process.
- Equally, there were elements of my notes I would have 100% wanted included in the final piece, but would need to have made that clear in the prompt. I also forgot a couple of key factual points about the game, which I had to manually add later.
- I was tempted to go even deeper with AI and use NotebookLM to summarise what ChatGPT had written, then form notes based on that. There’s real potential for both focusing and editing longer pieces that way, but more than doubles the workload.
It’s quite clear that, for now, use of AI for personalised writing is still a very hybrid approach of human input and the LLM’s broader summarisation and language manipulation skills.
It goes without saying that I could’ve simply prompted “Write me a 600 word review of the video game Cocoon” and still received a workable ‘unique’ piece of content to publish.
While a terrible thought, this seems inevitable to usurp many websites and writing streams in the future. Even doing it using the much longer manual process I’ve detailed still feels ‘cheating’ in some respect. It’s something honest creatives will struggle with for many years to come.
From a process side, the biggest learning is when to stop with promoting and tweaking, and what custom instructions are worth taking the time to prompt.
With an extra ten minutes, I could probably have made the article a lot more ‘like’ my writing. And with another hour, probably got it 99% there. But would that be worth the time and effort when I could potentially manually rewrite the whole thing much quicker?
It’s no surprise – as is becoming apparent for creative industries in general – that knowing how to use AI tools is imperative. Moderation is key, as is recognising that they are just tools, rather than solutions, and should be treated as such.