A tool for Human-AI collaborative writing of speculative fiction stories.
ollama pull mistralnpm installnpm run devinstall dependencies
npm installcompile and hot-reload for development
npm run devcompile and minify for production
npm run buildlint with ESLint
npm run lintOpen the .env file and set the VITE_MODEL and VITE_API_URL environment variables accordingly. When using Ollama you'll need to pull the model before using it.
VITE_MODEL=mistral
VITE_API_URL=http://localhost:11434/api/generateStory templates serve as a starting point for writers. They can be static text, text generated by a LLM, or a mix of both.
Story templates are json files located in src/assets/templates and imported in src/assets/templates/index.js.
Each story template requires a name and one or more actions. Each action requires a type (static or generate) and a template. Depending on the type the template will either just return the template text or send it to the LLM to generate the text response.
A simple template for static text:
{
"name": "static design template",
"actions": [
{
"type": "static",
"template": "In a world where technology has advanced beyond our wildest dreams, design practices have evolved to keep pace. As we step into the future, designers are no longer confined by the limitations of physical materials and two-dimensional screens."
}
]
}A simple template for text generation
{
"name": "generated design template",
"actions": [
{
"type": "generate",
"template": "Write the first paragraph of a story on the state of design practice in 2100."
}
]
}For easy customization you can specify variable in env and envoke them in the template string using the syntax ::variable::.
{
"name": "generated design template",
"actions": [
{
"type": "generate",
"template": "Write the first paragraph of a story on the state of ::topic:: practice in ::year::."
}
],
"env": {
"year": 2100,
"topic": "design"
}
}It is also possible to chain multiple actions. The example below inserts generated text into a static template. The first action uses the bind property which binds the generated response to the specified variable reply. The static action inserts that variable by using ::reply::
{
"name": "mixed design template",
"actions": [
{
"type": "generate",
"template": "Kim is a designer living in the year 2100. A friend asks her what she does all day. What does she respond?",
"bind": "reply"
},
{
"type": "static",
"template": "Kim explains \"::reply:: Is this something you'd like to do as well, Alex?\""
}
]
}Prompts allow writers to interact with the LLM. They are similarily structured to story templates and can currently either be executed on selected text or new lines.
Like story templates they require a name and a list of actions. Additionally a trigger selection or new-line and a mode replace or append must be specified.
Here's a simple example to continue the story. The response will be appended at the end of the existing text. Here you can use the flag ::full::, which will be replaced by the full story text.
{
"name": "continue writing",
"trigger": "new-line",
"mode": "append",
"actions": [
{
"type": "generate",
"template": "Continue the following story, which is delimited with triple backticks. \n\nStory: ```::full::```"
}
]
}Note: Notice in the example above the use of triple backticks around ::full::. When prompting it is important to write clear and specific instructions, especially when prompts get long. One tactic is using delimiters to clearly indicate distinct parts of the input. Delimiters can be anything like triple backticks (```), triple dashes (---), angle brackets (< >) or XML tags (
Here's a simple example to shorten a selected text. The response will replace the selected text. ::selection:: in the template string will be replaced with the user selected text. ::selection:: is only available if the trigger is set to selection. Here you can also use ::full:: which is replaced by the full story text. ::full:: is available on all triggers.
{
"name": "condense",
"trigger": "selection",
"mode": "replace",
"actions": [
{
"type": "generate",
"template": "Shorten the following text: ::selection::"
}
]
}Here's a simple example to make the model further elaborate the idea (term, concept, etc.) behind a selected text. The response will be appended at the end of the story. ::selection:: in the template string will be replaced with the user selected text. Notice that you can also use ::full::, which is replaced by the full story text. ::full:: is available on all triggers.
{
"name": "elaborate",
"trigger": "selection",
"mode": "append",
"actions": [
{
"type": "generate",
"template": "Continue the following story further elaborating the aspect addressed in the following text. Text: '::selection::'. \n\nStory: ::full::"
}
]
}Again, it is possible to chain multiple actions. The action type options allows users to select from given options. The bind property is used to define a variable option which the selected option will be assigned to. In the second action the selected option is inserted into the template string using ::option::
{
"name": "push forward",
"trigger": "selection",
"mode": "replace",
"actions": [
{
"type": "options",
"options": ["10 years", "100 years", "1000 years", "10000 years"],
"bind": "option"
},
{
"type": "generate",
"template": "rewrite the following text as if it was set ::option:: in the future ::selection::"
}
]
}Alternatively, options can also be generated using the action type generate-options. The example below first offers static options, followed by generated options and finally uses both selections to generate text.
When using generate options it is crucial to state in the template string that the answer should be returned in json along with a set of keys. Additionally the keys must be listed under the keys property. These values are used to parse and verify the LLM response. Once the user chooses an option the keys and their respective value become available as variables. The name property identifies the key which is used for presenting the options to the user.
{
"name": "diverge",
"trigger": "new-line",
"mode": "append",
"actions": [
{
"type": "options",
"options": ["poltical", "societal", "cultural"],
"bind": "option"
},
{
"type": "generate options",
"template": "::full:: \n\n suggest three topic ideas to delvelop the above story further while focussing on ::option:: aspects. provide them in json with the following keys: topic, description",
"keys": ["topic", "description"],
"name": "topic"
},
{
"type": "generate",
"template": "::full:: \n\n continue the story above with one paragraph that focusses on ::option:: aspects and the topic of ::topic:: (::description::)"
}
]
}It might be more efficinet to generate options and final text response in one step. This can be achieved by rephrasing the generate options template string and closing with a static action that simply returns the value of one of the keys.
{
"name": "diverge alternative",
"trigger": "new-line",
"mode": "append",
"actions": [
{
"type": "options",
"options": ["positive", "neutral", "negative"],
"bind": "tone"
},
{
"type": "generate options",
"template": "::full:: \n\n suggest three story coninuations to further tell the above story while keeping the tone ::tone::. limit the length of the continuations to 10 words each. provide them in json with the following keys: title, continuation",
"keys": ["title", "continuation"],
"name": "title"
},
{
"type": "static",
"template": "::continuation::"
}
]
}VSCode + Volar (and disable Vetur) + TypeScript Vue Plugin (Volar).