ghost

role: Concept development, Systems architecture, Agent tooling, Engineering
date: Apr 2026–present
team: Brand, Creative, Product
my part: I created Ghost’s portable brand format and built its task-aware retrieval layer so agents can get relevant guidance within a context budget.

brand judgment before generation

Agents now make screens, emails, and copy. Without a brand’s stance, they fall back on the model’s defaults. I designed and built Ghost to give them that judgment before they start working.

I designed Ghost’s portable brand format and built its task-aware retrieval layer. Teams record their voice, principles, and decisions about trust alongside concrete materials in a .ghost/ package. Agents retrieve the guidance relevant to their task within a context budget, rather than loading the whole brand each time.

give the next agent the decision

Correcting one output doesn’t teach the next agent. I made the package a place to record the underlying judgment: when to celebrate, how to deliver bad news, or when brand presence should recede.

Authors write those decisions in Markdown, describe when they apply, and reference existing components, assets, or examples. The package travels with the repo.

Your brand package

Write guidance in Markdown: when it applies, what to do, and which materials to use.

.ghost/
  manifest.yml
  glossary.md
  brand.md
  principle.density.md
  voice.md

Ghost

Your agent uses the CLI to read the package and retrieve guidance.

ghost gather: Get the complete guidance menu.
ghost pull <ids>: Get selected guidance and materials.

requests

context

Your agent

Chooses the guidance that fits your task and uses it to make the work.

Installed ghost skill

Instructions for reading the package and applying its guidance.

less guidance to load

37%

less brand-guidance text in the Square comparison

Ghost’s selected guidance averaged 99.4 KB against a 156.9 KB DESIGN.md containing the full guidance. Across both brands and models, the reduction averaged 30.4%. Human designers reviewed all results and confirmed they met the quality bar.

what agents make with it

Each brief below went through that selection flow against the same package. The outputs explore different situations and interactions within a shared design language.

internal adoption

I presented Ghost in a leadership deep dive and at a company all-hands, with a positive response to both. I followed up with an educational session on its architecture.

Since then, internal usage has risen sharply. Product teams have reached out to test integrations.

Project site · Source on GitHub · Development log

Open-source development preview.