Case Study · 01
Building AI-assisted content infrastructure at a global scale.
Content audits at Meta were slow, fragmented, and dependent on institutional knowledge. Auditing a term or concept across our content ecosystem took weeks of coordination across teams, and the results were rarely complete.
Manual audits only find what you already know to look for. Teams relied on people’s memory and on documentation that wasn’t always reliable, so whole surfaces or scattered instances got missed. That meant late-stage delays and post-launch fixes. It’s an even bigger problem for content in regulated spaces, where a missed term is a compliance and legal liability.
AI tools had the same gaps, plus false confidence. A query could come back empty just because the tool didn’t know where to look or how to search a particular repository. When it did return results, it implied “that’s everything” after searching only what was convenient. Unless you were an expert in that content, you’d never know what it missed. If you did, re-prompting ate up the time AI was supposed to save.
When I was asked to run yet another content audit in a heavily regulated area, I decided to build a systematic fix. The fix needed to leverage AI to move faster, and most critically, to scale institutional knowledge to anyone running a content audit.
As an independent initiative with no engineering support, I used Claude Code to build a multi-agent skill that encodes institutional knowledge of where content lives across dozens of codebases and CMSs, and how to search each one for specific terms and concepts. I published it company-wide so anyone running an audit inherits that expertise.
The core technical challenge was that search logic isn’t uniform across the systems where content lives. So I built a modular architecture of up to 12 specialized agents, each designed around one system’s search logic, searching dozens of codebases and CMSs in parallel.
Instead of literal keyword matching, the agents search by concept, flagging variations and related terms that keyword searches miss.
I also built in persistence and quality control. Specialized evaluation agents check every result. When a query comes back empty, they treat it as a signal to reformulate and try again, using that agent’s knowledge of its system rather than a generic retry.
Before rolling out the skill company-wide, I ran a pilot with content and design peers, who stress-tested the tool against systems they knew well to validate accuracy and build trust.
The skill delivers two outputs:
I reduced a process that took weeks of cross-team coordination to about 30 minutes for one person.
For privacy and compliance work, the skill adds a safety net that makes sure audits are complete from the start. Its analysis also flags legacy language, terminology discrepancies, and parallel workstreams affecting the same content, which gives siloed teams early visibility into conflicting projects.
Most importantly, the skill scales institutional knowledge, so anyone can run an audit with more complete results.
Shortly after launch, a legal team needed to find every remaining link to a deprecated site to address an immediate legal risk. The skill adapted easily to search for URL patterns instead of terms or concepts. I helped them find every instance across our codebases and CMSs in minutes, work that would have required engineering support and far more time. It showed the skill works as versatile infrastructure.