The LLMOutput Problem

Large language models are text machines. They generate tokens sequentially, word after word, sentence after sentence. The output is inherently linear. But the knowledge encoded in that text is often hierarchical, relational, and branching. There's a fundamental mismatch between the output format and the information structure.

The Formatting Gap

LLMs try to compensate. They use markdown headers, bullet points, numbered lists. But these are cosmetic patches on a structural problem. A bulleted list inside a paragraph is still a paragraph. Headers inside a scroll are still a scroll. The output medium limits what the model can express.

When you ask an LLM about a complex decision, the model internally represents branching options and weighted outcomes. But the only thing it can output is text. So it flattens the structure into prose, losing the very thing that makes the information useful.

The model knows the answer is a tree. The interface forces it to be an essay.

Structured Output

• Prompt engineering extracts structured data from LLMs
• JSON schemas define the shape of the response
• Tree structures map directly to decision spaces
• The interface renders what the model actually knows, not just what it can type

Bridging the Gap

Persephonie asks the model to think in structures, not paragraphs. The prompt defines a schema: root question, branching options, outcomes with sentiment. The model fills the schema. The interface renders it visually. No flattening, no loss. The user sees the structure the model always had.

This isn't about making AI responses prettier. It's about letting AI express what it actually knows in a format humans can actually use.