facebook · content strategy · AI

Creating an AI-driven CD audit agent

As the only content designer on Facebook Core Identity, I created an AI agent named Spruce that let me scale myself so I could focus on current projects while cleaning up outdated content across Profile.

Tools: Claude Code, internal agents for analytics and diff creation

Prompt architecture is proprietary and covered by NDA. This page shows how Spruce behaves and what it produces.

How Spruce works

Severity rubric

High severity

Blocks assistive technology or translation. Vague or misleading CTA labels. High-traffic surface.

Moderate severity

Inconsistent terminology or lower-traffic surface.

Low severity

Casing, punctuation or style drift that doesn’t block comprehension.

My role

Encoded CD expertise in the prompt as well as severity levels
Trained Spruce to hand off work to other agents that complete different jobs, such as analyzing string liveness and pulling daily views
Ensured Spruce submitted diffs as drafts and didn't land code directly without review

Results

Prompt tests to ensure accuracy

Profile directories in codebase scanned weekly

54%

Increase in diff accuracy over prompt lifecycle

Increased CD coverage exponentially across Profile
Encouraged cross-functional collaboration with eng
Created reusable prompt pattern that scaled to partner teams