Fictional Writing with AI support

tl;dr LLMs help me become a better writer because they bring knowledge and availability to the table I otherwise could never get my hands on.

Context

• I write fictional short stories and non-fictional pieces.

• For my newsletter I write in English as a non-native speaker.

• I derive pleasure from the writing process and when the results reflect stylistic beauty and craftsmanship.

• I do not write for output volume, to earn money or wide publicity paired with the goal of investing as little resources as possible

• I do not have much money to spend on this (the kids are eating well though)

• It makes absolutely no sense for me to create that one perfect prompt which is able to spill out a perfect text. If you came here for that, turn around and leave.

“Vibe” writing (not what you think)

about 1-2 hours for 2,000 words

I am not very good at chess because my capability for planning several branching steps ahead simply… sucks. Same applies for larger projects, my mind overflows with optional pathways and it all becomes a tangled ball of rails and several trains of thoughts.

So, writing for me is vibe writing, just without the AI. The first draft is reasonably messy. It’s mainly about ideas, narration, rhythm, pace and if it resonates with my gut feeling.

If I cannot pin that down to a reasonably complete, readable draft I may start again at the beginning or decide that I drifted off so far that I am actually writing about something else. Sometimes I realize that I my idea does not have enough substance, than I drop it completely.
Research is normally down to ‘extrapolate’ if my concepts are not completely unhinged but it can become a devastating rabbit hole at this stage – rendering the time frame above unrealistic.

The amount of errors I introduce in this phase is ridiculous and embarrassing. Their becomes There, I miss the last ‘e’ in here, and you have no idea how many times I wrote 'of' instead of 'off' or vice versa.

In general the writing itself is unpolished and uneven. There may be some gems, but since words tend to hang in my head longer than they should I also introduce a lot of repetitions.

Intermediate: Reviewing & Redacting

The clearer I am about what I want to write, the better works the craft itself. I have more restraint, I explain more consistently, I write clearer sentences and I make far less typos and grammar mistakes.

Messy drafts can result in rewriting whole paragraphs and chapters. This can be very frustrating when I realize that my core message sucks and I could not pin it down – leading me back to re-imagining the draft.

Rubberduck AI (finally.)

At this point I really have a strong desire to share my work to receive feedback. This point meaning 1 a.m. But unless you are insanely wealthy there is no option for that. Next working day could be a possibility but how many people do you know who could instantly review your piece and…

• brings extremely wide T-shaped knowledge (which actually is more like [ ] shaped)

• has serious language skills (grammar, style, asf.)

• reads all the entries in a series and sees the context

• gives you feedback within minutes

• wants to read the piece

• the last: does not ask a living wage (which at least I cannot pay as an underdog writer)

Now, this is when good LLMs shine. And I am talking about Claude Opus 4.6 in particular (Sonnet works as well, but it’s a bit too much on the sugarcoating side or as Christina Wodtke (Öffnet in neuem Fenster) puts it ‘it’s more of a sycophant’). I tried local LLMs as well on a RTX 3080 and 64 GB RAM (bought before it got touched by Midas) but I got mixed results and Proton’s Lumo made far too many mistakes to be helpful. If you want to try the following stuff, I tested with Claude on free tokens. I switched to the paid version later on.

The Prompt: Right and Wrong

So, the rather obvious thing to do: let the LLM check for typos, grammar and in my case weird idioms that do not exist, do not translate well or come off as simply borderline awkward. Upload your text (haven’t tried URLs but I don’t want to publish before finalization) and either use general prompt instruction or the chat itself. This is my first instruction for the mosaic novel I am writing:

I want you to check attentively for typos and grammatical errors. Keep in mind that some awkwardness still might be very much intentional as stylistic devices.

As a rather ambitious and playful writer (which is not the same as good) I have a knack for playing with words and stylistic devices. I found out that Claude was sometimes flagging those to a point where it bothered me (Yeah, you’re right Claude, but…) . I had to tell it to interpret them correctly. I collected those in a separate markdown file, explained them shortly and uploaded that as well. I could do this because Claude told me why it flagged something. Then added this:

Please consider the also uploaded list of stylistic devices to fully understand why I use certain phrases.

This works remarkably well. Once I explained a stylistic device used at one location in my texts, it tried to explain others I had not specifically mentioned in the file yet.

The Prompt: Quality and Craft

Writing my short stories, which are all supposed to work on their own, but provide a larger universe, I wanted to make sure that everything I write holds together as consistent and flawlessly throughout.

I want you to take a thorough look at scientific elements, continuity and inner logic and structure of world-building.

And since my stories are very character-driven, I wanted those to be especially rich and believable.

Motivation and actions of characters should be plausible.

It’s also important to me not to recycle old ideas and fall into common traps of lazy writing and research.

You may also check for typical tropes and if they are navigated sufficiently well.

And the last bit relates to the writing itself

Feedback on the craft of writing itself (pacing, structuring,...) is also very welcome.

The results were amazing. The stuff Claude covered from scientific details and continuity and tropes to craft are measured in literally years of writing experience and studying creative works.

I also have some ideas for long character arcs and how the general world works. I put those into extra files as well and let Claude check against them to make sure it fits to the larger picture I am building.

You may consider the document for world knowledge and characters after your assessment. This information will be unavailable to the reader. But you can check if the plot and your assumptions still check out.

This is especially helpful when you want the LLM to stop extrapolating and bicker about dangling ends and loose threads.

The Prompt: Playful and Personal

As said before, I am very playful person and I love referencing pop-culture and easter eggs.¹ I put in another instructions to test if the LLM can catch those.

If you want you can point to highlights and pop cultural references and easter eggs.

And Claude does. Not all of them and sometimes they are duds. And some I do explain in the style notes. Still, in general it finds refs not even my closest friends would catch on a good day.

To round things up I want Claude not to be nice. I don’t need nice, I want to be good at what I do. And the last sentence is there because Claude did start to slightly hallucinate here and there.

Be honest and diligent, no need to sugarcoat. Make sure not to confuse text passages in between the stories.

Result: Reviewing & Redacting (again)

about 1-2 hours for 2,000 words

After I give Claude a draft and ask it to assess it, a more or less vicious circle begins. It points out all the stuff it finds, tells me the reason, I fix it. I upload again, reset its chat context (try that with a human) and the next round of feedback begins.

It finds hard typos and errors, I correct them, and then it finds more in the next run. At first this highly irritated me. Partly because I did not realize I was introducing new errors when changing text, but also because it behaves very human – it’s not its sole job to find all the errors, so it does not. Rarely it would dangle back and forth and criticize stuff which itself just proposed – the risk of chat window attention spans (but I do know some humans with the same problem).

Craftsmanship, world-building and science are the most frustrating and rewarding areas to get to a level where Claude is satisfied. Frustrating because many times a small change is not enough. Rewarding because Claude is really diligent in many things regarding this and supports me in my learning experience to write better, consistent worlds which are enjoyable to read.

When I am satisfied and Claude is mostly fine with it, I publish it. This is it. To get to the quality I achieve with using Claude I’d normally would need weeks, for research and gathering feedback alone.

The Verdict

The most important thing I had to learn when writing with AI support was the following: You need to know what you want to write and what you want to share with the world. Your voice.

The good thing about ‘the voice’ is: the LLM can help you finding it – by reflecting about the intention of your own words. In this, it’s quite like a human who wants to spend more time with you. So, it won’t crush you but also won’t say you’re perfect. It will motivate you to go for the next edit to really excel. And you can manipulate it – for good or bad by priming it. It’s not deterministically right – you need to have vision and take responsibility (Öffnet in neuem Fenster) (Link also provided by C. Wodtke. That woman is resourceful).

If you don’t find that voice, you will be likely writing for the audience of Two. That’s you and the LLM of your choice. To get real feedback and a decent assessment how your writing and ideas indeed qualify, you still need to get out there and let others read your works. This is actually the hardest part for me: growing my audience and finding people who are interested AND give me their opinion.

Until then, you won’t know. But to that point AI can be an excellent rubberduck.

Ethical Final Note

Using Generative AI means you are drawing on the collected (and stolen) knowledge and creativity of all human creators ever. You also profit from people being exploited while those models get trained. It also has a severe environmental impact if not taken care of properly.

Those are some of the reasons why I publish my stories for free or as cheap as possible. If you use Generative AI as well, please consider giving back, otherwise you are part of the problem.

1. You can call yourself lucky I do not use footnotes anymore. They break the reader’s flow. But boy, do I love those. ↩