This article is based on our YouTube video - if you would prefer to watch a tutorial on how to use these features, you can do so here.
Adding Models and Changing Parameters
It is simple to add new models to your prompts.
Take a new or cloned prompt
Go to models
Press add model
Select your model from the model drop menu
Name your model in the name second (the system will suggest a name should you leave this field blank when you pick a model)
Change your parameters. As a primer:
Temperature: how coherent or creative a model is. This usually is a scale between 0 and 1, but some models allow values higher than that. The “colder” the temperature (so closer to zero), the more statistical and less creative its responses will be. this is good for when you want a model to accurately respond with facts instead of coming up with stuff.
Top_p: a threshold for which words are even in the running to be selected based on their combined likelihood. Again, this is a value between 0 and 1, and can be thought of a percentage between 0% and 100% of the full knowledge of the model’s brain. 0.9 is a good starting point: the weirdest words won’t being picked, but there is still a good selection.
Max Tokens: if in doubt, 2048 is good. Again, this differs between models, so look up their max output size on pages like the OpenRouter model card. Models like Claude 3 actually have a context size of about 4,000 tokens.
Frequency Penalty: If you experience repetitive outputs, or the AI is looping the text, this is a good parameter to tweak. It configures how often it uses the same words or phrases in the output, the higher the number, the more you discourage the AI from repetition of common words/phrases. Again this differs between models, some range from 0 and 2 where 0 is no penalty at all, whereas other models also go between 0 and 2, but 1 means the “middle” ground (so you can either be more or less repetitive)
Presence Penalty: Same as the frequency penalty, but this time higher numbers encourage the model to not use words/tokens that already appear in its input. It encourages the AI not to repeat your beats word for word. But it also will use less information of your world building, character details and so forth (since you actively encourage the model to ignore those details).
Prompt Editing
In the instructions tab of your cloned prompt, you can edit the text as you please. Some simple prompt changes to try out:
Change the “primer” sentence to include the genre of the book you are writing.
Add or remove rules in the system message to more clearly reflect your style of writing.
Change the number of words that you prompt the AI to write in a scene beat
Add in examples for the AI to follow (caveat: this works better for the larger context models)
Altering the function calls
Finally, for those who want super duper control, you can change the programming a bit. Think of these function calls as placeholders, that Novelcrafter replaces with actual content once you call the AI.
Here are five to get started, that you could add to your prompts to pull up more details.
<storySoFar>{context.storySoFar}</storySoFar>to get the chapter summaries for your story prior to the chapter you are work on.
{context.codex}to take into account codex entries that are mentioned within the prompt.
{snippets.get("Snippet Name")}to pull up a snippet as a reference, for example if you have a sample chapter of your writing style.
{context.wordsBefore(#)}and{context.wordsAfter(#)}the AI will read the words before and after
{context.wordsBeforeAcrossScenes(#)}for the AI to read words in previous scenes/chapters
You can see a list of all available functions when you type in the open-curly braces "{". Navigate the list with your up and down arrow keys. Hit Tab or Enter when you want to insert the function you like. If the list doesn’t show for you, you can also use the shortcut of CTRL+SPACE while you’re in the curly braces to bring it back up.
Some functions might require extra arguments, which usually are self-explanatory (like with the wordsBefore needing a number or snippets.get needing a name).