Content Check
An AI agent that automates quality checks within design and eng workflows
Project type: Agentic system design
Company: Meta
Product context
The problem
WhatsApp ships strings to 2B+ users in 60+ languages. A single typo or missing word affects not just English; it propagates globally, across countless surfaces.
We lacked a singular, formalized way to preempt these errors or address them post-ship.
The solution
Create an automated system that can scan, review and fix strings in code.
Product goals
-
Build an agent based on existing content guidelines
-
Train on existing codebase
-
Expand to proactive model
Content requirements
The basics: Spelling, punctuation, grammar, capitalization, missing words
Product-specific: Voice & tone, terminology, display rules, component-based guidelines
Higher-order: System-level thinking (e.g., cross-surface consistency), FOA & external patterns
Content strategy
Knowledge ingestion
Existing internal resources
-
Content style guide
-
Voice & tone guidelines
-
WhatsApp design system
-
Terminology glossary
-
Word lists
Training
Golden datasets
Created Bad/Good string pair examples to provide clear quality benchmarks.
Rationale
Supplemented examples with the guidance behind each choice.
Validating & refining
Clarification
Managed false positives, refining rules to include more nuanced guidance.
"Do not apply"
Implemented logic to protect user-set entities like business names.
Human discretion
Identifying where rules-based logic may fail and CD arbitration is needed.
Outcomes
Scale
The content guidelines have expanded to hundreds of lines covering grammar, spelling, platform terms, verb forms, localization and content standards.
We have used the skill to sweep every user-facing string across WhatsApp on iOS, Android, and web. Over just a few weeks, it found and committed hundreds of fixes, propagating to translation corrections in the tens of thousands.
Production
We merged this skill with our design review skill so they can run in tandem on every diff published. This automates quality checks and catches mistakes before they ship.