Better tools didn't improve Copilot code review. Rewriting their instructions did. 🧵
12
15
171
65,907
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. 👀
A comparison table with three columns: "Old Copilot code review," "GitHub Copilot CLI," and "Purpose." Rows: list_dir maps to glob for discovering candidate files and directories; search_file and search_dir map to grep for searching code for matching text, symbols, or call sites; read_code maps to view for reading relevant file contents once a path or range is known.

ALT A comparison table with three columns: "Old Copilot code review," "GitHub Copilot CLI," and "Purpose." Rows: list_dir maps to glob for discovering candidate files and directories; search_file and search_dir map to grep for searching code for matching text, symbols, or call sites; read_code maps to view for reading relevant file contents once a path or range is known.

3
2
26
12,144
The tools were fine. Their instructions were written for a different job, so the agent browsed the repo instead of investigating the pull request.
Figure 1: Before — a simplified illustration of the general-purpose behavior we observed: widening the search, guessing paths, and accumulating context.

ALT Figure 1: Before — a simplified illustration of the general-purpose behavior we observed: widening the search, guessing paths, and accumulating context.

Jul 10, 2026 · 5:14 PM UTC

1
2
20
10,992
We rewrote them to mirror how a reviewer actually reads a pull request: ask, narrow, read, decide.
Figure 2: After — a simplified illustration of the review-shaped behavior the prompt guided toward: stay anchored to the diff, narrow with grep and glob, then read focused ranges with view.

ALT Figure 2: After — a simplified illustration of the review-shaped behavior the prompt guided toward: stay anchored to the diff, narrow with grep and glob, then read focused ranges with view.

4
3
22
7,992
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/github…

Better tools made Copilot code review worse. Here's how we actually improved it.

How migrating Copilot code review to shared Unix-style code exploration tools reduced review cost by reshaping agent workflows around pull request evidence.

github.blog
2
4
17
8,289
Sort replies: Relevant Recent Liked