Transcription
Okay, I think we can all agree that AI isn't capable of writing. It can generate summaries, explain things, and it can even imitate good writing. But when it comes to writing something novel, new which is supposed to have an emotional impact on actual human beings, it AI fails just miserably.
Do we even want AI to be able to do that instead of real writers? That's another story. Not what I want to discuss here. For any professional or aspiring writer, there is a more practical and timely question on the table. Can AI be used as an editor? Can it help you improve your text, your own text you you crafted yourself? Be it a fantasy novel, a non-fiction book, or a long form post? Can it spot weaknesses in your writing? Help improve your style, fix grammar, and make better decisions as an author.
I'm uh 46 now based in Berlin and I've been a ghost writer and book editor for almost uh two decades. I worked on dozens of books, corporate and personal projects and uh once I uh even managed a book project that the client published with a total run of exceeding 1 million copies. It was a bit of a crazy story. And since JP launched, like everyone else, I've been experimenting with AI a lot, using all kinds of models and all kinds of chats and custom tools and building even my own. And here is my take. AI can definitely be used as an editor and it can be quite good at it. Surprisingly good. But why surprisingly? It's a large language model after all. So understanding language is basically its job description, right? But there is a catch. There is always a catch. If you blindly ask even the most advanced LLM to give you feedback or improve your style, it will like 100% fail. If you want to harness all the power of this technology, you have to know how to work with it.
So, um here I want to share my own framework I've developed for using AI as an editor based on my experience in publishing, the way I think about LMS, the core principles I follow, common gotchas I've noticed, all the way to the actual prompts I use. I hope it will be helpful whether you are a writer just starting out or someone with a lot of writing experience might be not much experience uh with AI. My way of working with AI as a writer is obviously not the only way. I'm curious to learn how you use AI as well. Please share your thoughts in in the comments.
Before we start, I want to address an ethical question first. You might say, are we now taking away work from editors and replacing them with AI, which by the way has been trained uh on a data gathered in a questionable way to to put it to put it mildly, right? So, let's let's let's pause here first. Um, editing is an expensive service. Not everyone can afford it. If you have an editor, usually that means you have a publisher who pays them. And if you self-publish or just write for a blog, you are paying out of your own pocket or not paying at all. Like not many people can can do that. They priced out of quality editing. So in many cases when you use AI, you are not replacing a real editor's work. You are not replacing humans. Even if you do have an editor, using AI might still be very beneficial. I've worked as an editor on many projects and I know for a fact that the the quality of my output as an editor depends heavily on the quality of the input. If there are too many uh low-level problems that need fixing uh my energy and time which are limited get spent there and I can focus on more subtle issues. If I get a clean well thoughtout draft I can focus on something more intricate and the output will be much much better. Finally, editors themselves can use AI to speed up their process and delegate the boring grant work, patching, missing commas, catching accidental repetition. So they can focus on task that require real skill and real deep thinking. So uh on balance if you apply AI to editing, we are not replacing human editors. We empower them. It's not a zero sum game. Editor editors get to work on more exciting stuff. Many people who had never had access to a quality editing now get at least some degree of it. Readers get better writing in the end. Um, everyone wins.
With that out of the way, let's now talk about how AI works, which is critical to understand, at least conceptually. I've noticed that some people have certain misconceptions about LLMs. And I think um the biggest is the idea that the more context you give AI, the better the result will be. Essentially, AI is a very sophisticated autocomplete, a statistical machine that predicts the most likely word to come after a given set of words. So, it follows that the more context AI has, the worse it performs. Give it the phrase when it rains and it easily predicts it pours. But give it half a novel, stop in the middle of a chapter and ask it to continue and it will start generating gibberish because there is no correct answer anymore. There are many possibilities with similarly low likelihood. The practical implication of this is that you have to go in the opposite direction of what most people do. Limit what you ask of AI as much as possible. And uh this works on two levels. The narrower the job, the better the result. The simplest example u if you ask a guy to check grammar in general, it's very likely to miss a lot of issues. But if you ask it to check specifically only like punctuation mistakes, the likelihood of it missing a comma will be greatly uh reduced. The same applies to how much text you give AI to analyze. The larger the text, the worse the result is. Give it an entire book and ask it to check punctuation and even in this case it will likely miss a lot of commas. But give it a single chapter and the result becomes much more consistent. Give it just a few paragraphs, a tenth of a chapter and it performs even better. So we have to keep this in mind.
Given this let's now think about how we can split the jobs we give an LLM when editing a text. Here I want to discuss the traditional editing process which might be helpful to understand. Essentially every book is edited in um three main stages. The first is what I call the creative editing stage. This is the most important and most complex one. This is where you try to make the text work in general. Here you tackle creative questions, voice, metaphors, word choice and so on. Uh the second stage is polishing another read of the text where you fine-tune. Once you're happy with how everything works creatively, you give it another go. See if parts can be trimmed. If there are repetitions, if we sections can be made stronger. You don't make big changes here, but you are still working on the on the creative side of things. And then finally comes proofing. This is where you double check everything. Grammar, facts, inconsistencies and so on and so on. I think it's very important to understand that every stage is layered and the editor always does a number of things simultaneously. Proofing is the most obvious stage to demonstrate this. Here you have you pay attention to a number of completely different things. Punctuation, types, spelling, tenses and so on and so on. All this will be very relevant to how we apply LLMs.
Let's also keep in mind that edits aren't black and white. Different editors will have different opinions about the same sentence. One editor will say it's crap. Another editor will say it's brilliant. Literature, luckily for us, isn't math. And in real life, authors often disagree with certain suggestions from their editors and sometimes fight them. Working with your editor, is a tricky process, even when you have similar views in general on what makes a good text. When declining your editor's suggestions, you have to be polite. Editors try to do their best after all. Right? So, you have to explain your position and find the right words, which can be tiresome to be honest. With AI, by the way, it's much simpler. You just disregard the corrections you don't like. You are not hurting anyone's feelings. AI is not conscious. There is no AGI here. It's just autocomplete. So, you should treat it as a tool. By rule of thumb, around 50% of the comments AI gives uh can and should be disregarded. That number might be higher for creative tasks and lower for grammar. If you blindly apply any suggestions from AI, you risk sanitizing your story, especially when you are early in the creative stages. So this is important to keep in mind. Even when you use AI, you still have to make all the decisions. You have to trust your intuition. AI is there to highlight potential problems you might have overlooked, but you are the judge who delivers the final verdict.
Okay, let's get to editing. Finally, our starting point. You have a text you are happy with. As an author, you have done everything you wanted and you think it's working as a story. It's done. As a text, it maybe needs work. Based on everything we discussed, I think you already know how we are going to approach this. We will split the editing process not just into three broad stages but into smaller laser focused tasks and we will work on different problems one by one. For an even better result each time we'll be given an the LLM a limited part of the text no more than a chapter. You might consider even splitting each chapter into a few parts. I'll show you all my prompts but um to be honest I don't even think it's important to copy them word for word. If you understand the process conceptually, you can create your own prompts very easily or you can just ask AI to create prompts for you. LMS are quite good at it at this point. If you still want to check out my prompts, watch till the end and I'll explain where you can grab them.
One last thing before we discuss all the prompts in detail. Edits very much depend on their context. If you are trying to get rid of stock phrases, this will mean different things depending on your genre. A crime fiction novel is one thing. A literary fiction novel whose whole point might be being innovative with language is a different beast. The bar for stock phrases will be completely different. So a professional editor will treat text differently depending on its genre regardless of their personal taste. That's part of being a professional. And this means you have to give AI a reference of what your genre is and you are what you are trying to achieve where your bar is. This applies to every prompt we'll be discussing in this video, especially when you are still early in your editing process.
Let's move on stage number one, creative editing. Here I run four main checks and occasionally a fun one. For creative editing, I usually start with a prompt that focuses on vocabulary precision, scanning the text for weak words like nice and interesting, good, and generic verbs, and figuring out if there are better options. I have to warn you that the accuracy isn't super high. AI tends to overthink to add unnecessary complex words, and if you blindly follow every suggestion, it will be very detrimental to the text. This might be especially harmful if your text aims to imitate daily speech where a simple vocabulary is the whole point. A deliberate stylistic choice. But even in such cases, a prompt like this still helps me spot big words and I didn't use deliberately and simply missed. Usually about 10 uh 30% of suggestions are useful here.
Number two, next I check stock phrases, tired metaphors, and dead languages general in general any overused word or expression. Same caveat. When I write a text that imitates how people actually speak, it's impossible to avoid some set phrases. As far as I can tell, AI cannot reliably differentiate between a good set use of set phrase and a bad one. So, apply your best judgment. Still, it helps spot things you missed. At the very least, it highlights every stock phrase you used and makes you pause and think, shouldn't be there. What if I swap it for this or that? Will it become better or worse? 20 30% accuracy is the best you can hope here. But that uh 20 30% can make a real difference.
Number three, another highle check is the famous show don't tell. The common advice to to express feelings through action and feeling telling details instead of he was afraid or the furniture was cheap. It's really good advice if you don't overuse it. As with any rule in writing, you have to know where to break the rule. If you are writing a stream of consciousness type of text, this show don't tell idea doesn't apply to you obviously. And sometimes you have to tell to keep the story going, especially if you are writing genre fiction. But then again, I I will highlight every telling part in the text and you'll have to decide if you want to rewrite it or keep it as it is.
And number four, I'm if I'm working on non-fiction, I have a prompt that checks the logic and the arguments now how they are grounded. This catches logical weaknesses like jumping to conclusions based on random facts. The edger the danger is that it can push you to hedge too much. eding based on this account reportedly etc in every paragraph and dancing around things instead of being direct which can be detrimental but if you are aware of the danger checking logic is a good idea for non-fiction in my opinion and um now the fun part I've also been experimenting with asking I to imitate different writers and assess my text from their point of view it's quite uh entertaining the trick is to pick writers who left a big enough body of work and advice on how to write there aren't that many otherwise AI will just hallucinate the the usual choices are Hemingway Steven King Nabokov or Czechov warning good uh give it a try just don't take it too seriously as you probably noticed for stage one creative editing I disregard 70 80% of the suggestions coming from an LLM early on AI can be helpful but its benefits and accuracy are much more tangible in the second and third stages than in the first.
So, let's talk about stage two. In a traditional publishing process, it's a second read of the text. Sometimes it's done by a new editor who comes with fresh eyes and a unique set of skills and a different goal to polish a story that has been considered finished in general at least. First, if it's non-fiction, I have a fact checker. I ask AI to fact check every text fact-based statement. A fact checker is also helpful for historical or sci-fi fiction that's heavily based on real facts. And I have to say this is where AI already beats human editors by a wide like wild margin. A true story. I've been working on a book that had been already published through a traditional publisher. So it went through a writer, an editor, another editor, the whole deal. I ran I ran this fact checker prompts on it and it found many mistakes that human editors just missed. I guess I I guess this is similar to coding right now where AI finds bugs and vulnerabilities in websites and applications that have been there for decades. It was really astonishing and humbling to see how many things the editors missed. They were just they were doing their best. But you can't know anything, can you? Unless you are an AI, of course.
Another thing I I almost always do is a redundancy sweep. Uh finding ways to trim the text and make it more concise. It's well known among among editors that every text benefits from at least 10% trimming. This works on a few levels. Words that add no meaning, sometimes whole sentences, occasionally paragraphs. Um a word of caution, sometimes AI cuts might be might break the flow. For me, the music of the text, the melody of the text is very important. Sometimes I agree a word isn't necessary and adds no meaning, but I need it for the melody, so I keep it. But I usually accept 50 to 70% of the suggestions.
Then a repetitions check. I ask AI to scan for repetitions. Be very careful. I, for example, use repetitions all the time as a stylistic device, repeating words in in close proximity. I've tried tweaking prompts to make AI distinguish uh my deliberate repetitions from accidental ones and it cannot do it reliably. It will it will flag everything that repeats and you have to decide whether it's a real problem or an imaginary one. Sometimes it's also better to repeat a word than to force something very unnatural in the text. So accuracy around 30%. If you're working on something really sophisticated where the bar is very high, that's uh that's what I there is what I call a zero repeat scan. You could try. It imitates what editors at highbrow magazines do. It's generally not a good idea to use a rare very distinctive word multiple times in one text. If you use a word like gargantuan at the beginning and again at the end, unless it's deliberate, it jumps out at the reader and that's not good. You have to ask you can ask AI to focus specifically on this spaced repetitions of marked distinctive words. By the way, it's one case where breaking text into like parts uh isn't possible since LM needs to like read the entire body. That's four main prompts I use when polishing a text.
Then comes proofing the simplest stage and this is where you can rely on AI the most. Accuracy increases to maybe 70 80%. But you still have to follow the rules. small chunks of text, very narrow tasks. Before anything else, a consistency check, checking that place names are spelled correctly, that's the that there is no mismatch between American and British English, that's a character isn't spelled one way in one place and differently in another, and all the mismatches uh that can appear.
Check number two is tense precision. This is very relevant for English and I assume for German too and many other languages. In English tenses are definitely complicated, especially when your story is in the past and you have to shift from the past into the even deeper past. Keep in mind that AI tends to overuse the past perfect. So you have to steer it toward the more readable version signal the shift in the the deeper past then move to the simple past once the shift has happened. At least that's what I usually do. Uh another one is a an article check. uh articles are especially especially tricky in English because sometimes there are no clear rules. It's more about gut feeling and it differs across versions of English. Consider whether it's worth asking AI to focus on them. Then punctuation ask you to check your commas and so on separately. It will never miss an Oxford comma. You will be covered there. And what I run uh at the very end after all the other checks, it's a general grammar check. This flies in the face of all the principles that I just shared, but sometimes it's a helpful final overview that catches something that all the previous uh checks missed.
I suspect you are now thinking this seems pretty complicated. It's about a dozen prompts. Yes, it's a lot. It can take up to an hour or two to process all the feedback for a chapter. But still, I dare it's 10 times 100 times faster than working with a human editor where you wait days or weeks for edits. work through all the corrections, send it back and add your comments. And the quality isn't there yet, but uh as we already discussed, it's still uh it's still a big step forward.
Okay, I promise to share the prompts themselves, right? Uh you can find them at a small website I built, grabaprompt.com. It's absolutely free. You don't have to sign up or anything. You can simply drop your text and construct the final prompt with a click to use it in an LLM of your choice leveraging your current subscription whatever it is like claude chat GP or anything else.
One last thing when you start working on your text you will probably feel that to doing this in a chat interface is very very cumbersome. a lot of going back and forth between the chat and your editor. And when you have a follow-up question, for example, for a few items on the list, it becomes very hard to keep track of them and discuss them with AI one by one. I got tired of of it and built my own tool, my own very minimalistic text editor that leaves feedback as comments on the document. They look just like comments, for example, in Google Docs. The most important part is that each comment can be turned into a separate chat and you can discuss the the suggestions. You can click any comment and discuss it and it even breaks longer text in chunks automatically. The app works with markdown files on your local machine with obsidian folder folders as well. Other way the text stays on your machine not saved on the server and you can use different AI providers clince models. There are even local models available so nothing will ever leave your computer. And most of the prompts I discussed here are already built in, but you can easily set up your own. It's fully customizable and it's the whole point. If you want to check it out, go to icanwrite.app. It's free to try and to use with uh local models. Let me know what you think uh about the app and the the ideas I shared. And happy writing.