CD diff review dashboard

The problem

Meta's code review tool, Phabricator, was built for engineers. CDs used it too to review string and design changes before they shipped, but everything about the tool was designed for a different kind of reviewer.

The first problem was the review queue. When a string change was detected, entire CD teams got bulk-added as reviewers. But the queue had no awareness of endorsement status: if a CD had already signed off, the diff still showed as unresolved, and you had to open it to see if anyone had reviewed it yet. Even diffs you'd endorsed yourself stayed in your queue.

The diff detail view was also hard for CDs to parse. Phabricator's diff view showed raw file diffs, CI signals, and engineering context. Finding the actual string changes meant scanning through the code around them. I'd open a diff and often couldn't tell whether it needed content eyes at all, or whether I was looking at an implementation change that happened to touch a string file. I'd often have to open Metamate (our internal AI assistant), paste the diff number, and ask "what am I looking at?" to even get started.

Two problems: a queue full of false positives, and an interface that buried what CDs actually needed to review.

What I built

A CD-specific workflow layer that addressed both.

The queue tracked endorsement status—if any CD had signed off, the diff moved out. No more opening each one to check whether someone had already handled it. The queue also included short AI-generated summaries and tags that helped quickly orient CDs to which Meta surface the diff was touching and what kinds of CD-relevant changes it was making (string changes, accessibility fixes, etc.).

The diff detail interface surfaced CD-relevant context instead of engineering noise. Each diff came with an AI-generated summary written for CD reviewers, not engineers. The dashboard also included a built-in AI assistant, primed with the diff's context and with CD skills (terminology standards, component rules, surface-specific guidance) so reviewers could get answers without leaving the interface. (Phabricator had a built-in AI reviewer too, but it was tuned for engineering.)

What shipped

Live at Meta, accessible to any CD with an internal login. I used it regularly for my own diff reviews, and it made the process noticeably faster. The potential scope was broad: Phabricator handled all engineering diffs across Meta, so any CD working on Facebook, Instagram, WhatsApp, Threads, or internal tooling could have used it.

I soft-launched to my immediate team deliberately—the AI assistant's answers weren't meeting my quality bar yet, and I wasn't willing to do a full announcement until I was confident in the assistant's response quality.

Where it was headed

The main thing in flight was a warm handoff to Metamate, Meta's internal AI assistant. The built-in assistant could answer some questions about a diff, but it didn't have access to internal context or systems and couldn't take actions on diffs. When a reviewer needed to actually fix something rather than just understand it, the assistant would launch Metamate and inject a prompt with the relevant diff context and what the user was trying to do. They'd land in Metamate already oriented, ready to work.

The warm handoff and the response quality were the last pieces before I was ready to take the dashboard to the full CD org. I was laid off before those pieces were complete.