I find Claude to be the best writing assistant. I’ve tried gpt, Pi, and many smaller projects. Claude is like having a real conversation about your story. You tell it to stay within the bounds of your descriptions. The project feature is a game changer for authors.
Great post. I found new Claude to be superior to old 3.5 if I broke my scenes into manageable beats and asked it to write 2-3 beats (seeking around 150-200 words per beat) at a time. But the days of getting 1500 words per prompt seem to be over, more or less. The quality of the prose is marginally better than old 3.5 most of the time.
Had some interesting stuff happen when doing scene beats. Claude started cutting off after 300 words and saying like [the rest of the scene will be continued at…. ] I can’t remember what it said but it was essentially saying that it was happy with what it did and didn’t feel like it didn’t need to do more. I was using Novelcrafter at the time 😂 was hilarious and the scene beat cost me ¢0.18 I’d rather use GPT at this point
The new version is awful, outputs are ofen cut off early. Assuming you can get one, I often find myself in a doom loop where I give an instruction, it ignores it and then asks a question. Even if you tell it not to ask a questions and move on, it keeps rewriting. Yesterday I had a issue where it would give an output then recognise it had made the same mistakes. I had this 7 times in a row, I'm not a fan when it started using sarcasm. I can't even get it to write 500 words at times.Typically the outputs get smaller as the length of the prompt gets bigger. So if your writing your own scene beats and want to flesh them out you can end up with a summary that is shorter than what you provided it. Other odd behaviour was having Claude turn several scene beats into 5 scenes within a chapter, the only issue is that there simply wasn't enough material to spread over that length and no need to split into indivdual scenes, three "scenes" it created were short pieces of dialogue in a conversation so should never have been split. As a productivity tool I've abandoned it, previous prompts giving 1500-2000 words are now giving around 700, I have realised that you can tell it to check the output against the prompt to ensure all elements are accounted for, doing this meant it did actually complete the task set. However my big issue with this is that if your paying for a pro sub you are being penalised having to use 3 prompts to get the same output that the previous version gave you. It eats up your limit super quickly. The only way I can use my old prompts is via OpenRouter though it still takes at least 2 attempts, but if you use the webui you end up having to edit the additions back into the generation because when it used artifacts it proceeded to remove content for the addition. My prompts include a instruction to refer back and ensure all key elements have been addressed. However, after generating an output if I ask it to check if all are addressed (as I can see its skipped 50%+), my next message is to check and produce the omissions, it then wants me to confirm, then even after it creates the additions I can see there are further omissions, in the worst instances it literally rephrased part of the original prompt but did so in a way that is worse than what I had originally written.
So with the most recent Sonnet 3.5, it's not going to be immediately apparent what the changes is, but I used it daily before the change and after. The biggest change is the level of language its using. Before, you may have had a slight hint occassionally of purple prose, the only way you are getting that now is if you are telling it that is what you want. It's going to be much dryer emotionally as well. I would suggest dropping to opus and only using Sonnet for final refinements. I don't have any negative feedback on the model at all, but you just have to get used to what its writing depending on your genre. In the actual Anthropic UI which is the only thing I use, I have never had any issues get 2500 words and ever more occassionally. The one complaint I do have overall is that there is a size limit for the Claude projects which I do not have on the Google Gemini Advanced side. The thing to know about it is that as rule, its not going to add extraneous stuff to get to a word count if it can tell the story with fewer words better...which is of course the goal of writing.
I found it much dryer, too. And it definitely shows the difference reading through a text where the model changed two-thirds of the way through. It’s a different author voice coming through, even with my sample text (i use novelcrafter) in the codex.
Great video. I have been up since 2am. The work was completed by 0530. I just finished having this version roll out a script from a fully fleshed-out, outline. I use the Save the Cat, 15 beat method, and the software perfomed, admirably. However, I work by sending in the Cast with arcs, ONE beat at a time after relaying that ALL 15 beats are coming. Thanks so much for all of your work and instruction!
So if I’m working on the story and I have a good outline, I can get Claude the right 3000 words in a chapter of regular and inconsistent basis. This new one I get no more than 2000 words. But I also specify length of the overall project and of the chapters. I also asked it to give in the outline how many words per section and it manages to follow
Was going to write a message asking for this very video. It’s been such a downgrade, terribly upsetting because I went from using it daily for helping me edit (not necessarily write) and it’s like talking to a child now. I really wish they didn’t nuke the creative writing part of this and separated out the computer science stuff from the thing that made it perfect for stories.
I personally don't mind that it doesn't write out tons of text. Honestly the more you have AI write, the more editing you need to do on the back-end for many reasons. That's completely ignoring the joy aspect of writing, which honestly is the biggest reason for me. I write because I love writing, and the more AI does it for me, the less I enjoy it. Additionally, one thing writing with NC has taught me is that constraint is the real source of true creativity. The more tightly you constrain the AI, the more interesting and truly creative it will be. When you give it a long leash and tell it to write a ridiculous number of words like 1,000+ then you get much less interesting results and you'll have much more editing to do.
I found going in smaller chunks yields pretty good results. However: Seems to reach overall limit for me faster (ie cannot continue until such-and-such time, and I'm a paid subscriber)
Claude may be viable for shot stints but its is plagued by limits both in the chat and project potion. Nothing more frustrating then working on a project to then run into a "you can chat for the next 6 hour" kind of thing and on top of that, there is no way to tell you are encroaching this limit. Until they remove these limitation, Claude is not a recommend imo.
I have the same experience, with new 3.5 output generally capping at 1k words on average. I find the prose to be quite a bit better, especially when instructing it to use both scene and summary narration: old Sonnet 3.5 seemed to only ever want to do dramatic scene narration (action, dialogue, description) even when explicitly asking for summary prose in either style prompt or beats - chat-gpt fared better in this department, mixing narrative modes. However, apart from the limited output, I've noticed the new 3.5 also rushes through the beats sometimes, skipping over crucial things that were mentioned in the outline (like a character saying something, which another then reacts to). So while the limited output can be remedied by generating scenes or stories in smaller parts, the stylistically slightly better prose comes at the cost of glossing over crucial information in the outline sometimes.
I’ve noticed it randomly cutting off answers with the dreaded square bracket summaries, but it’s generally still the best writing assistant and productivity tool for authors.
So as I developer as well as a creative writer I really enjoy Claude 3.5 sonnet. I am curious though with all of the different models available on hugging face if there would be a better suited model that can be ran with Ollama now it allows huggingface models seamlessly.
You bring up a good point about these models being generic and not trained for specific usecases like creative writing. Have you looked into any solutions that are leveraging the open source LLMs specifically fine tuned for creative writings?
Love the video quality. What do you use as an editor? Yes, I think a genre specific model would be amazing although another Copyright nightmare. Have you considered training your own model?
Would you say this is the best LLM to help you write/rewrite with? What would you recommend for that anyway? Something where I do most of the work, like a 90-10 ratio.
For writing, would recommend Erato from NovelAI. It's what I use for writing. It's very different from Claude. It was trained differently and only outputs a few sentences at a time. I think of it like boosting me over hard spots. But it is a bigger learning curve. Once you figure it out, it's great.
It threw me a bit at first .. as I couldn't figure out why it didn't write a whole chapter like the previous 3.5 did... But then I figured out you have to go scene by scene and it normally does quite well. Also you have to tell it to continue if you want to extend a scene... It's just a little more picky and needs a bit more baby sitting than the previous version did.
Seems like you've verified what Kate@Novelcrafter found in a recent live stream - the new version is better at adhering to your prompt. Everyone is used to giving the old version 3 beats and getting a 4th (or 5th) for free. The new version stops after 3. Give it more beats, get longer output. I'm more excited about Haiku 3.5 coming out since 90% of what I use the AI for I do with Haiku.
The prose in the new model is easily better, but... and a big, butt - getting the best results out of the new model takes a fair bit of coaxing and encouraging of the LLM to get. Also, if you are talking to Claude directly expect to run out of prompts quickly due to this, unless you are a masterful prompter. I am not. I suspect this is the combination that's tripping up the angry mob. 1) they don't coax the LLM and 2) they run out of prompts before they reach their writing goals for the session. The latter is really the big one if you are not using open router.
i think its better. im the type of writer who wrote in beats and let it expand in limited scope than to let it write more than a thousand words per prompt.
So you're saying you can get AI to write books for you? I'd love to be able to tell it what kind of story I'd like to read, and then have it write one for me. Sad that it's eliminating authors as artists, but I guess that's technology.
I find Claude to be the best writing assistant. I’ve tried gpt, Pi, and many smaller projects. Claude is like having a real conversation about your story. You tell it to stay within the bounds of your descriptions. The project feature is a game changer for authors.
Great post. I found new Claude to be superior to old 3.5 if I broke my scenes into manageable beats and asked it to write 2-3 beats (seeking around 150-200 words per beat) at a time. But the days of getting 1500 words per prompt seem to be over, more or less. The quality of the prose is marginally better than old 3.5 most of the time.
Had some interesting stuff happen when doing scene beats. Claude started cutting off after 300 words and saying like [the rest of the scene will be continued at…. ] I can’t remember what it said but it was essentially saying that it was happy with what it did and didn’t feel like it didn’t need to do more. I was using Novelcrafter at the time 😂 was hilarious and the scene beat cost me ¢0.18 I’d rather use GPT at this point
what are you using now if not Novelcrafter? just curious
Because I generate beat by beat, the only thing I noted was the improvement in the prose. I have been very impressed and satisfied with the 3.5
Did you check if you’re in “Concise” mode. I think that may be the problem. It’s currently the default for some users.
The new version is awful, outputs are ofen cut off early. Assuming you can get one, I often find myself in a doom loop where I give an instruction, it ignores it and then asks a question. Even if you tell it not to ask a questions and move on, it keeps rewriting. Yesterday I had a issue where it would give an output then recognise it had made the same mistakes. I had this 7 times in a row, I'm not a fan when it started using sarcasm. I can't even get it to write 500 words at times.Typically the outputs get smaller as the length of the prompt gets bigger. So if your writing your own scene beats and want to flesh them out you can end up with a summary that is shorter than what you provided it. Other odd behaviour was having Claude turn several scene beats into 5 scenes within a chapter, the only issue is that there simply wasn't enough material to spread over that length and no need to split into indivdual scenes, three "scenes" it created were short pieces of dialogue in a conversation so should never have been split. As a productivity tool I've abandoned it, previous prompts giving 1500-2000 words are now giving around 700, I have realised that you can tell it to check the output against the prompt to ensure all elements are accounted for, doing this meant it did actually complete the task set. However my big issue with this is that if your paying for a pro sub you are being penalised having to use 3 prompts to get the same output that the previous version gave you. It eats up your limit super quickly. The only way I can use my old prompts is via OpenRouter though it still takes at least 2 attempts, but if you use the webui you end up having to edit the additions back into the generation because when it used artifacts it proceeded to remove content for the addition. My prompts include a instruction to refer back and ensure all key elements have been addressed. However, after generating an output if I ask it to check if all are addressed (as I can see its skipped 50%+), my next message is to check and produce the omissions, it then wants me to confirm, then even after it creates the additions I can see there are further omissions, in the worst instances it literally rephrased part of the original prompt but did so in a way that is worse than what I had originally written.
So with the most recent Sonnet 3.5, it's not going to be immediately apparent what the changes is, but I used it daily before the change and after. The biggest change is the level of language its using. Before, you may have had a slight hint occassionally of purple prose, the only way you are getting that now is if you are telling it that is what you want. It's going to be much dryer emotionally as well. I would suggest dropping to opus and only using Sonnet for final refinements. I don't have any negative feedback on the model at all, but you just have to get used to what its writing depending on your genre. In the actual Anthropic UI which is the only thing I use, I have never had any issues get 2500 words and ever more occassionally. The one complaint I do have overall is that there is a size limit for the Claude projects which I do not have on the Google Gemini Advanced side.
The thing to know about it is that as rule, its not going to add extraneous stuff to get to a word count if it can tell the story with fewer words better...which is of course the goal of writing.
I found it much dryer, too. And it definitely shows the difference reading through a text where the model changed two-thirds of the way through. It’s a different author voice coming through, even with my sample text (i use novelcrafter) in the codex.
Great video. I have been up since 2am. The work was completed by 0530. I just finished having this version roll out a script from a fully fleshed-out, outline. I use the Save the Cat, 15 beat method, and the software perfomed, admirably. However, I work by sending in the Cast with arcs, ONE beat at a time after relaying that ALL 15 beats are coming. Thanks so much for all of your work and instruction!
So if I’m working on the story and I have a good outline, I can get Claude the right 3000 words in a chapter of regular and inconsistent basis. This new one I get no more than 2000 words. But I also specify length of the overall project and of the chapters. I also asked it to give in the outline how many words per section and it manages to follow
Was going to write a message asking for this very video. It’s been such a downgrade, terribly upsetting because I went from using it daily for helping me edit (not necessarily write) and it’s like talking to a child now. I really wish they didn’t nuke the creative writing part of this and separated out the computer science stuff from the thing that made it perfect for stories.
I personally don't mind that it doesn't write out tons of text. Honestly the more you have AI write, the more editing you need to do on the back-end for many reasons. That's completely ignoring the joy aspect of writing, which honestly is the biggest reason for me. I write because I love writing, and the more AI does it for me, the less I enjoy it.
Additionally, one thing writing with NC has taught me is that constraint is the real source of true creativity. The more tightly you constrain the AI, the more interesting and truly creative it will be. When you give it a long leash and tell it to write a ridiculous number of words like 1,000+ then you get much less interesting results and you'll have much more editing to do.
Lol, you threw me for a loop. Between the slight lisp from the orthodontic elastics and the edit cuts, I thought at first you had AI cloned yourself 😄
I found going in smaller chunks yields pretty good results. However: Seems to reach overall limit for me faster (ie cannot continue until such-and-such time, and I'm a paid subscriber)
Claude may be viable for shot stints but its is plagued by limits both in the chat and project potion. Nothing more frustrating then working on a project to then run into a "you can chat for the next 6 hour" kind of thing and on top of that, there is no way to tell you are encroaching this limit. Until they remove these limitation, Claude is not a recommend imo.
I have the same experience, with new 3.5 output generally capping at 1k words on average. I find the prose to be quite a bit better, especially when instructing it to use both scene and summary narration: old Sonnet 3.5 seemed to only ever want to do dramatic scene narration (action, dialogue, description) even when explicitly asking for summary prose in either style prompt or beats - chat-gpt fared better in this department, mixing narrative modes. However, apart from the limited output, I've noticed the new 3.5 also rushes through the beats sometimes, skipping over crucial things that were mentioned in the outline (like a character saying something, which another then reacts to). So while the limited output can be remedied by generating scenes or stories in smaller parts, the stylistically slightly better prose comes at the cost of glossing over crucial information in the outline sometimes.
I’ve noticed it randomly cutting off answers with the dreaded square bracket summaries, but it’s generally still the best writing assistant and productivity tool for authors.
So as I developer as well as a creative writer I really enjoy Claude 3.5 sonnet. I am curious though with all of the different models available on hugging face if there would be a better suited model that can be ran with Ollama now it allows huggingface models seamlessly.
You bring up a good point about these models being generic and not trained for specific usecases like creative writing. Have you looked into any solutions that are leveraging the open source LLMs specifically fine tuned for creative writings?
Love the video quality. What do you use as an editor?
Yes, I think a genre specific model would be amazing although another Copyright nightmare. Have you considered training your own model?
I hope as things get redesigned, LLM creators keep accessibility in mind. Not everyone can use their voice to control a computer.
Would you say this is the best LLM to help you write/rewrite with? What would you recommend for that anyway? Something where I do most of the work, like a 90-10 ratio.
For writing, would recommend Erato from NovelAI. It's what I use for writing. It's very different from Claude. It was trained differently and only outputs a few sentences at a time. I think of it like boosting me over hard spots. But it is a bigger learning curve. Once you figure it out, it's great.
Thanks for paving the way into a future of AI writing.
It would be nice if they added version numbers and the option to swap between them instead of writing over the current model.
It threw me a bit at first .. as I couldn't figure out why it didn't write a whole chapter like the previous 3.5 did... But then I figured out you have to go scene by scene and it normally does quite well. Also you have to tell it to continue if you want to extend a scene... It's just a little more picky and needs a bit more baby sitting than the previous version did.
Seems like you've verified what Kate@Novelcrafter found in a recent live stream - the new version is better at adhering to your prompt. Everyone is used to giving the old version 3 beats and getting a 4th (or 5th) for free. The new version stops after 3. Give it more beats, get longer output. I'm more excited about Haiku 3.5 coming out since 90% of what I use the AI for I do with Haiku.
I just tried the new 3.1 free version. I'm really impressed. There's hope for ai yet.
The prose in the new model is easily better, but... and a big, butt - getting the best results out of the new model takes a fair bit of coaxing and encouraging of the LLM to get. Also, if you are talking to Claude directly expect to run out of prompts quickly due to this, unless you are a masterful prompter. I am not.
I suspect this is the combination that's tripping up the angry mob. 1) they don't coax the LLM and 2) they run out of prompts before they reach their writing goals for the session. The latter is really the big one if you are not using open router.
I agree. I am finding my scene beats now are not going much beyond 300 words.
I'll have to check.
True - not long responses. That's a problem but the writing is better
Close the doors behind you 🤣
Great video as always!
Yea, I saw unhappy posts on Reddit.
There are always unhappy posts about Claude on Reddit
i think its better. im the type of writer who wrote in beats and let it expand in limited scope than to let it write more than a thousand words per prompt.
I still find claude still can't code effectively, I often use claude for the first attempt and then finish it off in Gemini or ChatGPT.
So you're saying you can get AI to write books for you? I'd love to be able to tell it what kind of story I'd like to read, and then have it write one for me. Sad that it's eliminating authors as artists, but I guess that's technology.
Nah. It’s just another way to get a creative vision on the page. Same as if I used dictation vs. a typewriter.
It's completely broken. Doesn't follow custom instructions. Writing is like AI and doesn't flow.
I think if you give most AI all the info. they are all pretty good assistants. But as far as writing. they seem to get worse and worse.
Anybody making money with your stories? Claude and the RUclipsr are making...the rest please hands up....
Just know that if AI writes your books you cannot legally copyright it because you didn't write it. Have fun.
Oh shut it.
This is not true.