Outliner

Reads a prose document and generates a StoryCAD outline from it using Semantic Kernel and an LLM. This is the inverse of the normal StoryCAD workflow: instead of building an outline first and writing prose from it, the Outliner reverse-engineers a structural outline from finished or draft text.

[!NOTE] This sample requires an OpenAI API key. Set the OPENAI_API_KEY environment variable before running.

View source on GitHub

What It Demonstrates

The sample accepts a .txt, .docx, or .pdf file and sends the full prose to an LLM with a structured system prompt (Prompts/OnePassSystemPrompt.md). The LLM returns a JSON response describing the story’s characters, settings, problems, and scenes. The OutlineBuilder class then constructs a live StoryCAD outline from that response using the StoryCADLib API, and saves it as a .stbx file.

This is a Uno desktop application targeting both net10.0-windows10.0.22621 and net10.0-desktop, so it runs on Windows (WinUI) and macOS (Skia).

API Methods Used

Method	Purpose
CreateEmptyOutline	Create a new story from template
AddElement	Add Character, Setting, Problem, Scene elements
UpdateElementProperties	Set title, author, premise, and other overview fields
AddCastMember	Link characters to scenes

Semantic Kernel Components

Component	Purpose
Kernel.CreateBuilder().AddOpenAIChatCompletion()	Configure SK with OpenAI
IChatCompletionService	Single-shot chat completion
ChatHistory	System prompt + full prose as user message

How to Run

export OPENAI_API_KEY="your-key-here"

Build and run on the desktop target:

cd samples/Outliner/Outliner
dotnet build -f net10.0-desktop
dotnet run -f net10.0-desktop

On Windows, build the WinUI target:

cd samples/Outliner/Outliner
dotnet build -f net10.0-windows10.0.22621 -p:Platform=x64

From the repo root, .\scripts\build-outliner.cmd runs this build.

When the app opens, select a prose file (.txt, .docx, or .pdf) and an output path for the .stbx. The app extracts the text, calls the LLM, and builds the outline. A .raw.json file is written alongside the output with the full LLM response for inspection.

You can optionally set the model (defaults to gpt-4o-mini):

export OPENAI_MODEL="gpt-4o"

Input Requirements

• The full prose must fit within the model’s context window. For most short stories and novelettes, gpt-4o-mini (128k context) is sufficient.
• .docx files are read via DocumentFormat.OpenXml; .pdf files via PdfPig. Complex formatting is stripped; plain text is what reaches the LLM.