Agent-First Brand Kit

by Little Plains


1. What This Is

Traditional brand guidelines are designed for humans to read. They exist as PDFs, slide decks, and Figma files. They work well for design teams reviewing brand standards, but they are structurally useless when an AI agent needs to generate on-brand content.

This document defines the architecture for an agent-readable brand system. The core idea is simple: brand knowledge should be stored in structured, indexed chunks that an AI agent can selectively retrieve based on the task at hand. Instead of dumping an entire brand guide into a prompt, the system allows an agent to pull only the specific positioning, voice, or constraint data it needs for a given job.

The framework has three layers:

This is the technical specification for how those layers work together.


2. Chunk Architecture

2.1 What Is a Chunk

A chunk is a single file containing one discrete unit of brand knowledge. It has two parts: structured front matter (YAML) and body content (Markdown, YAML, or JSON depending on the data type).

The target size is 300–500 tokens per chunk. This constraint is deliberate. It keeps each chunk small enough to fit alongside other chunks in a prompt without consuming excessive context, while being large enough to contain meaningful, self-contained information.

Design Principle: A chunk should answer one question well. If you find yourself covering two distinct topics, split it into two chunks.

2.2 Chunk Types

Not all brand knowledge has the same shape. The system supports three content formats, each suited to different data types:

Format Extension Best For Example
Structured Data .yaml Positioning, personas, product registry, terminology, constraints brand-positioning.yaml
Narrative Content .md Voice guidelines, brand story, messaging translations, use cases voice-core.md
System Data .json Visual system specs, color values, typography tokens, spacing scales visual-system.json