Update July 11, 2023: The new Claude AI is out, and it’s now free to use, and way, way better. Check out my review here.
Update July 12, 2023: Poe now also offers Claude 2, with its 75,000-word limit and higher intelligence and better grammar, with 30 questions a day for free.
I just tried Poe’s new Claude AI — the one that can handle 100,000 tokens — or about 75,000 words — of text at once. A “token” is a unit of measurement for AI models that’s about three quarters of a word in length. Yeah, I know, they should just switch to words, instead. Who knows how these people think?
Anyway, what it all means is that Claude can summarize or edit a mid-length novel. In one go.
To get started, go to Claude-Instant-100K and, if you haven’t yet, sign up for a new account.
The 100,000-token version is only available on the paid plan, which costs $200 per year — but there’s a seven-day free trial.
Poe is a service from Quora that offers several different chatbots, including Claude, made by Anthropic, and ChatGPT from OpenAI.
I found it to be extremely fast. Maybe that’s because most people don’t know about it yet, though.
Overall, I found it to be very good at summarizing and finding problems in the text. But it wasn’t as good at offering writing suggestions. I have the same issue with ChatGPT — I personally don’t like its writing style. It doesn’t feel very engaging to me. I have a feeling that, by offering the most likely response, it winds up creating things that are too bland and generic. This is where advanced prompt engineering might help — if you have suggestions, please leave them in the comments below! — or maybe it will just improve with time and more training. Claude was also really bad at grammar. By comparison, ChatGPT was much better.
When when it comes to big picture issues — like finding plot holes or improving character arcs — Claude is amazing.
So, let’s get right into some examples and prompts.
The prompt context
Large language models like Claude and ChatGPT work better when you give them a role to play and some contextual information. This is called “prompt engineering.”
Here’s a prompt I like to use:
You’re an expert fiction editor with decades of experience. You have worked with hundreds of best-selling sci-fi and fantasy authors and understand what makes a book engaging. I am a sci-fi author at the beginning of my writing career, and need help with structure, style, grammar, and punctuation. If you understand, please say, “I’m ready. Please paste your text.”
Then just copy-and-paste the entire text of your novel. If it’s less than 75,000 words, of course. If it’s longer, you can divide the novel into parts, or come back later, when Claude can handle longer pieces of text.
I uploaded the text of my first novel, Krim Times. Well, novella. It’s just a little over 17,000 words.
Even though it’s short, ChatGPT would never have been able to handle it all at once.
But it was a breeze for Claude. Here was Claude’s initial response:
Thank you for sharing your fictional story. Here are some key takeaways and reflections:
• The story has an interesting premise of a medieval-themed virtual world with its own economy, politics, and challenges. The setting provides opportunities for suspense, intrigue and problem-solving.
• The characters are an engaging mix of traditional “types” – the security chief, grid admin, publisher, land baron, etc. – but with enough quirks and personality to make them interesting. The protagonist Trask in particular is complex, with his love of food and dislike of exercise in contrast to his sense of duty.
• The story moves at a good pace, with various twists and revelations unfolding throughout the investigation into the griefers. The clues and evidence build in a logical fashion to point the finger at William Lockton as the perpetrator.
• The writing style is straightforward and efficient, fitting for the genre of crime thriller/mystery. However, there are some opportunities for tightening up the prose, smoothing transitions and streamlining descriptions. A professional copy edit would likely be helpful.
• The acknowledgements section highlights the important role that support, feedback and editing can play in developing a work of fiction. It’s clear the author drew on the expertise of a range of people to help bring this story to life.
In summary, this is an entertaining story with potential for further polish and development. I hope my reflections have been helpful and insightful! Please let me know if you have any other questions.
This is all correct. I am a little miffed about Claude saying that the book needs professional copyediting — it did get professional copyediting!
But this is all general information that, except for the copyediting part, I knew already. Let’s get into some actual useful work.
Summarizing and outlining
I’ve written a lot of books in this series — about a dozen stories, novellas, and novels at this point, most of them yet to be published. I forget what happens in earlier books. And, to be honest, I wasn’t really careful about outlining while I was writing them.
Can Claude help?
Please create a detailed scene list for the book. For each scene, tell me the chapter number, a phrase describing the scene, the date and time of the scene, the location of the scene, the characters in the scene, and a short list of events that take place in the scene.
Here’s an example of how it should look:
SCENE 1: Assassin at window
DATE, TIME AND LOCATION: Mid-day, Leadenhall street
CHARACTERS: Unnamed assassin
• The assassin peers out of a window, seeing potential targets among the local residents and tourists on the street below. The tourists stand out in their colorful clothing. However, the assassin is waiting for a specific target.
Again, the more specific you are with the prompt, the better. Claude already knows the context — I’m writing science fiction, it’s playing the role of an editor. Now I’m asking for exactly what I want it to do, and how I want it to present the information.
Here’s a sample response I got back from Claude:
SCENE 1: Grid hearing
Date: Same day as previous chapters
Location: Grid meeting hall
Characters: Baron, Ademar, witnesses, mediator, Crewe, Osgar, Trask, William
• The Baron testifies that he tortured and killed the editor, confirming Trask’s suspicions
• The process server Crewe recognizes William as the person who abducted him and took him to the Baron
That’s accurate. I’m pretty impressed. This can save me a lot of time.
The only problem is that Claude would trail off after a few chapters and I had to type “continue.” And, it would start loosing its train of thought.
I experimented with having Claude try to do the whole novel all at once, and have it do it chapter-by-chapter, and doing it chapter-by-chapter seems to work better. It looks like Claude loses track of its instructions as it goes along. This will probably improve in future updates.
You can also ask Claude to review its own work:
Please review your scene list for chapter five. Is it accurate and complete? After reviewing, please revise your scene summaries. Remember to include dates, times, locations, and characters as well as a bullet-pointed list of key developments.
Another thing I have a problem with is keeping track of all my characters. I forget how to spell their names, what they look like, when they first show up, their relationships with other characters, their goals and motivations, how they talk — pretty much anything and everything. As I write, I wind up having to go back, over and over again, to try to figure this stuff out. It really slows down my writing process.
Over time, I started to create character lists, so that I could keep track of all this stuff — but then I’d forget to update them!
Let’s see if Claude can help fix that.
Here’s the first prompt:
Give me a list of all the named characters in the book.
Here’s a tip: after it gives you a list, ask, “Did you miss anyone?”
In my case, Claude found a couple of more characters it had missed the first time around.
Then I asked “Did you miss anyone else?” It took me a couple of rounds before Claude finally said, “After a third check, I believe I have now included all named characters mentioned in the story.”
Now you can create the character sheets.
Here’s the prompt:
Now, for each of those characters, create a character sheet. Start with their full name, then give a list of the chapters where they appear. Then list the following: what they look like, when they first show up, their relationships with other characters, their goals and motivations, how they talk, and other key information.
Claude would only do one character at a time for me. I hope it will soon have the ability to follow multi-step instructions. Until then, expect to have to keep prompting it.
You can use the same approach to create a list of settings, a list of themes or motifs, character arcs, key turning points, and so on.
Okay, we’ve got our outline and our character list.
Let’s get to the real meat of the editing process — the structural edit, also known as the content edit.
This is the part of the editing process where you ask questions like “does this make sense?” and “are there any plot holes?” and “do the characters act in a believable way?”
Here’s the first prompt:
You are an experienced structural editor. Take a close look at the plot of the book. Are there any plot holes or loose threads? Are there any points where the developments don’t make sense?
Claude identified four areas in my book that need improving:
- The motives for William being the griefer are not fully explained. His reasoning for trying to sabotage the grid is vague.
- Trask seems to rely heavily on vague suspicions and circumstantial evidence rather than concrete clues to identify William as the griefer.
- Some characters act in ways that don’t seem fully logical given the situation, such as William trying to physically escape capture rather than using his administrative powers. However, without the full context of these scenes, I cannot say for certain if these developments constitute plot holes.
- Some coincidences or conveniences drive the plot forward, such as Crewe happening to recognize William and the burnt newspaper happening to implicate him. However, these types of plot devices are common and not necessarily problematic without more context.
Claude seems to think that the text I gave it is an excerpt rather than the full book. It is the whole thing — it’s supposed to be short. It’s a novella. But other than that, Claude is spot-on. I feel a little embarrassed now about the book. But hey, it was my first one!
And, again, you should follow up by asking Claude if it there are any other areas that need improving — did it miss anything?
Let’s see how Claude would recommend I fix things, starting with the first problem — William’s motivation.
Here’s my prompt:
And here’s what Claude suggested:
Here are some suggestions for improving William’s motivations for being the griefer:
• Give William a clearer personal stake in the outcome. For example, he could be motivated by greed, wanting to buy the grid for cheap after destabilizing it. Or he may have a grudge against specific individuals he aims to drive out. Personal motivations tend to make goals more compelling.
• Tie William’s motives to the broader themes of the story. For instance, if a theme is virtual vs. reality, William could want to prove that the virtual world is ultimately meaningless. Or if a theme is lawlessness, he may believe the grid needs more order and this is his twisted way of enforcing it. Connecting to themes grounds motives in the work’s ideas.
• Show William’s frustrations with his administrative role and desire for more power. He may resent being just a “cog in the machine” and see destabilizing the grid as a way to gain recognition and promotions in the company that owns it. Jealousy and ambition can drive characters’ actions.
• Reveal William’s motives in flashbacks that explain his history and why he wants to disrupt the grid. Personal history and past grievances are common sources of motivation. Flashbacks could provide this context for his present actions.
• Have other characters speculate on William’s reasons, giving the audience different theories even if none are ultimately confirmed. This could make William’s true motives (when revealed) more surprising and meaningful.
These are all excellent! I would totally pay money for a human being to tell this to me. This is the kind of feedback that you’d normally only get from a very good structural editor, or from a critique group with a deep understanding of your writing, or maybe a very, very, very good beta reader or writing partner.
And if these suggestions aren’t enough for you, you can ask Claude for more ideas. Eventually, it will start to repeat itself, so you might need to guide it. For example, I might ask for ten ideas about personal grudges that William could have against people.
Here’s just one of many suggestions Claude had: “William once had an extremely minor disagreement with the owner over some trivial matter, like the proper way to fold a napkin. The owner made an offhand comment that wounded William’s pride, inspiring a quest for petty vengeance to prove the owner wrong.”
Once you decide on a plan, you can ask Claude for some ways of implementing it.
The flashback scene sounds good to me. But where would I put it?
This is solid advice. And it would have taken me at least half an hour to figure out where to put this scene.
Okay, now what should the flashback scene contain, other than the napkin bit? Here’s what Claude said:
Yes, the scene revealing William’s “napkin disagreement” motive through a flashback could potentially serve additional purposes to improve the story:
• The scene could introduce some comedy or lightness. By depicting the petty, ridiculous nature of William’s grudge, the flashback could elicit humour and provide a brief break in the plot’s suspense. This comedy could help underline the absurdity of William’s actions.
• The flashback could reveal details about the grid owner’s character. By showing how he innocently yet carelessly insulted William, the scene could flesh out the owner and provide more depth and nuance to his role in the story.
• Revealing the owner’s role in inspiring William’s grudge could make him slightly more sympathetic and complicit in the events. This additional nuance and complexity would enrich the plot and characterization.
• The scene could also include details about the company culture, politics or inner workings that lend more realism and world-building. Mundane office elements in the flashback could add texture.
• The flashback could feature other characters who are never seen again, demonstrating the wider scope of the virtual world beyond the main plot. Fleeting appearances by minor characters can make a setting feel fuller.
You might be tempted at this point to ask Claude to just write the scene for you.
I admit it, I did it. I asked it to write the whole scene.
And I’m glad to say that Claude got everything wrong — it got my tone wrong, it got the characters wrong, it even got the format wrong — it wrote it as a screenplay. That’s not to say that, with sufficient prompting, Claude couldn’t do a better job. Eventually, Claude and ChatGPT probably will be able to write scenes that just fit, and on the first try. So, as writers, we need to start thinking ahead about what it even means to be a writer. But that’s the subject for another article. Or many articles.
Still, even without the ability to write well itself, Claude is pretty good at structural editing.
Now, let’s move on. Can it do other kinds of editing?
The next stage in the editing process is editing for style, though it’s sometimes lumped in as part of content editing.
I like to think of it as a separate stage. In structural editing, you make sure that all the piece of the book are there, and they’re where they’re supposed to be and everything makes sense.
Then, in the style editing stage, you look at how the writing flows.
For example, some sections of your book may be too rushed, and could use a bit more description. Others might feel too padded, with text that doesn’t add anything to the story and can be cut out.
Or you might write in short, easy-to-digest sentences and then have a passage in the middle that’s dense and incomprehensible. Maybe you had a bad day or not enough coffee.
Or you might start out funny, then have the humor trail off over the course of the book. Or you might be going for a fast-paced book, and have a long section that really drags.
Or you might have a character speak with an accent in the first chapter, and forget all about the accent by the end of the book.
Here’s the prompt:
You’re an expert style editor with decades of experience. You have worked with hundreds of best-selling sci-fi and fantasy authors and understand what makes a book readable. I am a sci-fi author at the beginning of my writing career, and need help with my writing style. What are the top three problems with this book, as far as the stylistic aspects of the writing are concerned? I’m thinking about things like pacing, tone, and amount of description.
Apparently my scenes had lots of places that needed work with description.
- Chapter 1 – The introduction of Krim and its setting. Very little sensory detail is provided about what this virtual world actually “feels” like for characters.
- Chapter 1 – The Barley Mow Inn and its inhabitants. Adding atmospheric details about this key location could improve its presentation.
- Chapter 2 – The shooter’s perspective and actions. More information about what the assassin experiences could build tension.
- Chapter 3 – Trask’s movements and search. His actions are told plainly; showing them through his perspective could engage the reader more.
- Chapter 5 – The grid hearing. Sensory details that bring the chaotic atmosphere to life would make this pivotal scene more visceral.
And on and on. It was a little disheartening. But also — Claude was totally right about all these.
With each of these items, I can ask Claude for suggestions about how to make the fixes. But I didn’t, since I still have that flashback scene to write!
Meanwhile, here are some more style questions you can ask Claude:
- Does each scene contain sufficient description?
- Is each new character sufficiently described?
- Is the characterization believable and consistent?
- Are there any sections where I’m telling, rather than showing?
- Is the narrative voice consistent? Does it change unexpectedly between scenes?
- Are there overly complex or convoluted sentences that could be simplified?
- Is dialog realistic and natural sounding for the characters and setting?
- Are perspective shifts, flashbacks and other devices handled in a smooth manner?
- Does each chapter end in a way that pulls the reader into the next part?
- Are plot points delivered in an engaging manner that creates suspense and keeps the story moving forward?
- Is subtext and symbolism used deliberately to add depth, or does it feel heavy-handed?
- Does the story build in complexity at a reasonable pace, neither too fast nor too slow?
- Does the writing pull the reader directly into the story world, creating a sense of immediacy and experience?
- Are proper names, locations and other key details introduced in a way that avoids confusion?
- Does all foreshadowing actually come to pass in the story?
- Are transitions between time periods clear and consistent?
Finally, we come to the last stage of the editing process — the copyedit and proofread.
I decided to have Claude look at a book that hadn’t gone through any copyediting yet, the first draft of the ninth book in my Krim series, Bridge Over the River Krim. It’s quite a bit longer — nearly 59,000 words. Not my longest book, but getting up there. I wanted to see how Claude would handle longer text. It handled it just fine.
First, I started a new chat by clicking on the little broom icon to the left of the prompt box. The old chat was still there, if I scrolled up, but now there was a “context cleared” message indicating that I’m starting a new conversation. Unlike with ChatGPT, Poe and Claude don’t have a chat history yet. I’m hoping that will change. For $200 a year, I better have a chat history.
I went through the thing about being an experienced editor, uploading the text of the book, and having it give me its initial thought about it. It liked it. Thanks, Claude!
Then I gave it this prompt:
You’re an expert copyeditor with decades of experience. You have worked with hundreds of best-selling sci-fi and fantasy authors and understand what makes a book readable. I am a sci-fi author at the beginning of my writing career, and need help with grammar, grammar, punctuation, word choice, tense, point of view, passive voice, and other technical writing issues. What are the top problems with this book?
Here’s what it had for me:
Based on my review of the text, I think the top three technical writing issues are:
- Inconsistent point of view – The story shifts between third person limited point of view focused on different characters and occasional jumps into omniscient point of view, making it confusing at times for the reader. Fixing the point of view to be consistent throughout would likely improve clarity.
- Overuse of passive voice – Many sentences are written in the passive voice, which can make the prose seem convoluted and wordy. Reducing passive voice constructs in favor of more active sentences would tighten up the writing.
- Imprecise word choice – Though the writing is generally clear, some words and phrases seem imprecise or general when more specific language could bring the story to life for readers. Replacing some vague or generalized words with more precise and concrete word choices would improve the imagery.
Fixing these main issues would likely involve:
- Choosing a consistent point of view and sticking to it
- Running the text through a passive voice checker to identify opportunities to replace passive constructions with active verbs
- Performing a “precision edit” to identify general language that could be made more specific and descriptive
I totally agree with those recommendations, though I’d argue that the point of view falls more under style editing than copyediting. Still, it’s a valid suggestions, and I should do something about it.
Let’s look at the first item, point of view. Having read the book, which point of view do you recommend — limited third person, or omniscient?
Based on what I read in the book excerpt, I would recommend sticking consistently to limited third person point of view. Here are the reasons:
- Limited third person is a more common and readable point of view. Shifting between limited third person focused on individual characters and occasional omniscient slip-ups is confusing for readers. Sticking to one consistent point of view improves clarity.
- The story is primarily driven by the experiences of a few key characters. Limited third person allows us to get inside the heads of those central characters while maintaining an objective narrative tone.
- Omniscient point of view is harder to pull off in a compelling way. Since we don’t spend a prolonged time inside any one character’s head, it can be difficult to form an emotional connection with the story.
- The occasional omniscient intrusions involve summary-level observations (e.g. “there were castles and crowns made of marchpane”). These could easily be revised to align with a more limited perspective focused on one character.
These are all pretty good reasons. I think I’ll take its advice and replace omniscient point of view with limited third person.
Please go through the entire book, and find instances of omniscient point of view. Please give me the chapter number and paragraph text for each instance.
Claude found four spots in the book with omniscient point of view. When I asked it, “Are there any others? Did you miss any?” it found a couple of other instances, though they were a little borderline.
Just to be sure that it caught them all, I asked the same question again and it came up with a few more. Well, there’s my to-do list for next week!
Okay. Let’s keep going to the next item on the list of things that need to be fixed.
Let’s move on to passive voice. Please go through the entire book, and find instances of passive voice. Please give me the chapter number and paragraph text for each instance.
It gave me a few examples. Again, I asked it if it missed any, and it did. I could probably spend a whole day on this, just asking Claude to find examples of passive voice and then fixing them.
Finally, I wanted it to look at grammar and punctuation.
And, finally, I found an editing task Claude couldn’t do. It found grammar and punctuation problems where there weren’t any and even made up paragraphs that didn’t occur in the text. For example, it told me that I occasionally misspelled “Quimby” as “Quimbye.”
Now, I write with a grammar checker turned on, so my text doesn’t generally have a lot of grammar errors in it. So maybe the reason it couldn’t find any grammar mistakes to fix was because there weren’t any.
But there are plenty of examples of errors online! I grabbed a few sentences.
Take this sentence: “We cleaned all the kitchen while our parents were out to dinner.”
On its first try, Claude told me that the word “cleaned” was singular and the subject was plural, so the verb should be “cleaned.” I asked again, and this time it found no mistakes. The third time around, it spotted the problem — “all the kitchen” — but suggested “all kitchen” as the fix, which was just as bad. Finally, on the sixth try, it suggested replacing “all the kitchen” with “all the kitchens” which, is, at least, grammatically correct, though probably not what the author intended to say.
In the sentence, “He starts work everyday at 8 a.m.,” Claude didn’t notice that “everyday” should be two separate words instead of one.
In the sentence, “Less than 50 people showed up for the presentation,” Claude didn’t notice that the word “less” should be “fewer.”
Then I asked it to check the sentence “I could just lay down and go to sleep” — and got the error message “Sorry, you’ve exceeded your monthly usage limit for this bot.”
Yup, you only get 100 questions per month. Per month!
This limit is for the beta period, because it’s an experimental early access model, and the limit is subject to change. But still. Only 100 messages!
Will I pay for this?
First, I want the interface to be improved. Having to scroll back through your entire chat history gets unwieldy very quickly.
Second, I went through my 100 monthly messages in a couple of hours, just writing this article. I didn’t even get into the meat of editing either of my books!
In fact, I’m cancelling my trial right now, before I accidentally get billed $200.
But if you’re a writer or an editor, you should still check out Claude-Instant-100K. Or, at least, check out the free trial.
If you have a novel that needs either a structural edit or a style edit, Claude can give you a head start. Even if it only catches half of your problems, you’ll still be making life a lot easier for your beta readers, critique group partners, or professional editors. That’s going to save you time and money.
I also really appreciated its ability to generate a complete outline of a book and create character lists. This is an extremely useful skill.
At the very least, see what it can do. If you do wind up liking Claude, $200 a year comes out to around $17 a month. If you write a lot of books, and craft your prompts carefully so that you don’t use more than 100 a month, you might be able to get a lot of value out of it. And the prompt and length limits will probably increase, too.
If you’re an editor, you should learn how to use Claude either as a first-pass assistant or a second pair of eyes, to see if you missed anything major. It’s not a replacement for a grammar checker. It’s not a replacement for a human editor. It’s just another tool in your toolbelt that can definitely help you be more productive.
Me, I’m going to wait and see what the competition has in store.
Now that Claude can read and analyze an entire book, we’ll probably soon see ChatGPT and other large language models do the same. After all, this is extremely useful functionality and the AI space is advancing by leaps and bounds.
Oh, and remember those three sentences with grammar errors that Claude couldn’t fix? ChatGPT 4 fixed them all for me, and on the first try, without breaking a sweat. So if you do want to use a large language model as a grammar checker, I suggest using ChatGPT, and just pasting in one scene or short section at a time.
Edited by Melody Friedenthal
MetaStellar editor and publisher Maria Korolov is a science fiction novelist, writing stories set in a future virtual world. And, during the day, she is an award-winning freelance technology journalist who covers artificial intelligence, cybersecurity and enterprise virtual reality. See her Amazon author page here and follow her on Twitter, Facebook, or LinkedIn. Email her at [email protected]. She is also the editor and publisher of Hypergrid Business, one of the top global sites covering virtual reality.