The tireless writing assistant

How AI can help authors to work smarter, not harder

Quite a few writers (the ones who haven’t used AI much) think that having AI help with their work simply means hitting a button and going off to make a cup of tea while it writes their novel or film script. And, yes, you could do that. There was one entire movie written (sort of) by AI. There are some bestselling novels that definitely read as if last year’s AI wrote them. But it’s not how to use the technology effectively.

AI isn’t ein Zauberlehrling for writers, a writing slave that will demonically churn out ready-to-use pages. On the other hand, it’s much more than a versatile suite of tools. Better to think of it as an assistant-cum-researcher-cum-beta reader. Is that a little too broad? Maybe some examples would help.

Research first of all. I’m not the only writer who spends a big chunk of their time sifting through web pages and books searching for that one elusive fact or historical detail. I don’t normally object to having my nose stuck in a book, in fact it’s very heaven, but it isn’t the best use of my time when deadlines are looming.

Recently I was writing a roleplaying scenario and needed examples of “moral riddles” – not logic puzzles, that is; more like Portia’s caskets. Claude whizzed through all of human knowledge and, knowing that this was for a medieval or early Renaissance setting, in minutes it guided me to some examples from English, French and Italian literature such as some late 14th century stories from John Gower's Confessio Amantis. I’m not even sure how I could have found those by old-style Google search. Most likely I’d have had to plough through dozens of JSTOR articles – and, as I said, that’s usually my idea of fun, but not when the information I’m looking for is just to fuel a throwaway point in a story.

Another time I half-remembered an incident from Jean de Joinville’s account of the Crusades. I have the book, but I’d already spent a frustrating half-hour scouring through it and wondering if my memory was faulty before I realized I could grab a public domain version of the book from Gutenberg (we’re okay; Joinville was writing in the 13^th century), drop that into NotebookLM, and locate the episode I was looking for in seconds.

Of course, hallucinations can happen (note the name of this Substack) so you always need to fact-check information the AI provides, but it gives you a quick first-pass to find that elusive needle in the bookstack. Actually, full disclosure now: I usually use Perplexity, another AI, for the fact-checking.

Then there’s managing continuity. A lot of my work in games involves building intricate worlds, whether for a sprawling fantasy series like the Vulcanverse (nearly 750,000 words) or a complex videogame narrative. Keeping track of every character, location, artifact, and historical event is a monumental task. This is where AI can be an invaluable assistant.

You just feed the AI your existing work, notes, and worldbuilding documents, then you can query it to check for inconsistencies, recall specific details about your world lore, or generate summaries of key elements. Think of it as an exceptionally diligent research assistant who has digested your entire story universe and can instantly cross-reference information, like a fan who remembers all those details about your work that you’ve forgotten yourself.

My French publisher got in touch to ask me, “Is there a name for the mountains in the south-east part of the Fabled Lands? We’re making a new map and the artist needs to know.” Not long ago that would have meant an hour or two searching through the texts of all the books (looking for something that may not exist is harder than looking for something that does) but AI was able to tell me immediately: “No, those mountains are not named in the books. Would you like me to suggest names based on other locations nearby?”

Tools like NotebookLM, designed specifically for digesting and querying documents, are particularly well-suited for this kind of work. By helping you maintain a consistent and detailed world, the AI frees your mental energy for the creative heavy lifting.

The AI is an infinitely patient editor too. I had some old scans of a typewritten manuscript from decades ago. I had used OCR to extract the text but the scan was so poor that it came out horribly garbled – or “4orr161y gar6e$” as the OCR would have rendered it. With sinking heart I resigned myself to an afternoon comparing the OCR version to the original scans and recreating the text that way. But hang on a second – character recognition is exactly the kind of thing neural nets cut their teeth on, and accurate text prediction is how LLMs work. I gave the garbled version to Claude and said: “I am trying to extract text from a scanned document that is low-res and full of amendments by hand. This confused the OCR so the resulting text file is rife with artefacts and missing or garbled sections. Can you take this and try to reconstruct the original text? Don't add or change any of the intended content, just try to deduce and restore the original.”

Claude took less than a minute, replying with the cleaned-up manuscript and a note: “I see this is a scenario for the RuneQuest roleplaying game. Some sections were particularly damaged in the OCR, so I made reasonable inferences based on context and RuneQuest rule conventions. For example, I formatted all the character statistics consistently and reconstructed tables that appeared to be broken in the OCR.”

Another case: my Italian publisher emailed to ask about one of my books that they were considering making an offer for. Translations are expensive – well, they are if they’re done right. Deciding whether to shell out for that would be easier once the publisher had some feedback from their readers. Not all the Italian readers were able to assess the original English text, though. A few minutes with Google Translate in Documents mode and I was able to send them the book in Italian. Granted that was a very rough translation; I know because I reversed it to see what it looked like in English. Not quite “an invisible idiot” but not elegant prose either. It wouldn’t do for publication, but it was good enough for the publisher to decide whether to hire a human translator for the job.

I asked Gemini for some thoughts on this, incidentally. “Using large language models for initial translations can make your work accessible to a much wider readership,” it said, “allowing you to gauge interest in new markets before investing in full human translation. This is particularly useful for authors with extensive backlists or those writing in niche genres.”

I am exactly that kind of author — as Gemini, from having worked with me, knows full well. And by the way I don’t recommend Gemini for authors. It’s good at code and maths, but it has by far the most stilted prose style of all the LLMs. You’d almost think it was paid by the word. Claude, ChatGPT and DeepSeek are all much better for the kinds of task we’re interested in. I haven’t used Llama or Copilot much, but I’m sure they’re just as good. Try them all out and see which you like.

Lastly there’s simple brainstorming. Have you ever been stuck on a story problem, and as you start explaining it to somebody else you suddenly see how to fix it? The other person doesn’t even need to say much. What they do say may be irrelevant to solving the problem, but it’s enough if they get you thinking about it from another angle.

The beauty of doing this with an LLM is that it is endlessly patient and it knows everything on the Internet. The only downside is what I call “banjo duel daze”, in reference to the sequence in John Boorman’s film Deliverance (video above), whereby the LLM responds so rapidly and completely that after a few minutes you have to stagger off and take a break. Until we can pay for a direct brain upgrade (and no, I’m never going near Neuralink’s products!) we may just have to accept that our robot assistants may be able to tie a girdle round the earth in forty milliseconds, but at least they still need us to ask the right questions.