LLM Prompt Compiler

The LLM Prompt compiler converts a xtyle definition into a structured Markdown document designed for use as an LLM system prompt. Feed the output to Claude, GPT, or any language model to generate on-brand content without manually describing your brand in every conversation.

Installation

npm install @xtylejs/compiler-llm-prompt

CLI Usage

xtyle compile brand.xtyle.json --target llm-prompt

Output is written to ./out/<name>.brand.md by default.

Programmatic Usage

import { compile } from "@xtylejs/compiler-llm-prompt";

const markdown = compile(definition);

The compile function takes a full xtyle definition and returns a Markdown string.

Output Structure

The compiler walks the entire definition and produces sections for each pillar that has content. Sections are omitted when the definition doesn’t include relevant data.

Section	Source	Content
Identity	verbal.identity	Brand name, tagline, descriptions at multiple lengths
Colors	visual.palette + visual.semantic	Semantic color roles with resolved hex values
Typography	visual.typography	Font families, weights, type scale
Components	visual.components	Component parts, sizes, variants, states
Jargon	verbal.jargon	Domain-specific terms with definitions and deprecated alternatives
Approved Copy	verbal.copy	Pre-approved text fragments keyed by purpose
Boilerplate	verbal.boilerplate	Standard text (legal, disclaimers, footers)
Voice	voice.*	Tone, values, vocabulary, anti-patterns, audience, positioning, art direction, register

Example Output

Given the Midnight Diner definition, the compiler produces:

# Midnight Diner

*open late. no questions asked.*

Late-night ramen from a dark counter in a quiet alley.

## Colors

| Role | Value |
| --- | --- |
| primary | #F5A623 |
| surface | #0D0D0D |
| accent | #C0392B |

## Typography

- **heading:** Barlow Condensed, sans-serif
- **body:** Inter, system-ui, sans-serif

- **normal:** 400
- **bold:** 700

## Jargon

- **the usual** — whatever you had last time, remembered without asking
- **extra fire** — additional chili oil, applied without judgment

## Voice

### Tone
- **energy:** warm, low-key, like the cook who knows your name but doesn't need to chat
- **formality:** casual always. never formal.
- **humor:** deadpan. if you get it, you get it.

### Anti-Patterns
- food photography language
- hospitality industry buzzwords
- exclamation marks

Use Case

The primary use case is feeding brand identity to an LLM as a system prompt:

System: [contents of midnight-diner.brand.md]
User: Write a social media post announcing our new winter menu.

The LLM receives the full brand context — voice, vocabulary constraints, tone, positioning, approved terminology — and generates content that stays on-brand without manual guidance in every prompt.

This works across all three pillars:

• Visual context tells the LLM about the brand’s color palette and typography when generating design-adjacent content
• Verbal tokens provide exact terminology, approved copy, and jargon definitions
• Voice guidance constrains tone, vocabulary, and anti-patterns so the output matches the brand’s personality