Better tools didn't improve Copilot code review. Rewriting their instructions did. 🧵

We swapped Copilot code review onto the same shared tools that power the Copilot CLI (grep, glob, view), expecting a clean upgrade. Instead, reviews did more work and caught fewer issues. 👀

The tools were fine. Their instructions were written for a different job, so the agent browsed the repo instead of investigating the pull request.

We rewrote them to mirror how a reviewer actually reads a pull request: ask, narrow, read, decide.

The result: roughly 20% lower average review cost, while maintaining the same review quality. ✅ The bigger lesson: for an agent, tool instructions are as load-bearing as your API docs. Here's how we traced the regression to the instructions ⬇️ github.blog/ai-and-ml/gith…

@github that 20% drop from one instruction change is wild. what was the actual diff — a single missing line, or something structural?