# All green. Still wrong.

> AI reviews how code is written, not whether it is right.

- **Format:** short video, 9 steps, ~51 seconds
- **Topic:** GitHub Copilot code review limits — a refund pull request passes analyze, tests, the layering rule and an AI review Skill, and still refunds the wrong amount, because AI reviews how code is written, not whether it is the right code.
- **Author:** Suthahar Jegatheesan (MSDEVBUILD)
- **Category:** GitHub Copilot · copilot
- **Published:** 2026-03-02
- **Tags:** githubcopilot, codereview, aicoding, softwareengineering, pullrequest, flutter, devops, msdevbuild
- **Canonical URL:** https://blog.msdevbuild.com/shorts/copilot-code-review-how-not-whether/

---
## What you'll learn

- What an AI review Skill reliably catches
- Why a clean, tested pull request can still be wrong
- Where the human reviewer still has to decide

## Understand it one step at a time

### 1. A refund pull request

A use case, a button and three tests. It goes up for review.

### 2. Every check is green

Analyze clean, tests passing, and the review Skill finds no blocking issues.

### 3. Three days later

Promo customers are refunded more than they paid.

### 4. Why every gate missed it

The tests had no promo, so subtotal and total were the same number.

### 5. Two jobs in one review

Mechanical: is it written correctly? Judgment: is it the correct code?

### 6. What the Skill is good at

Rule-based, tireless, consistent: the checklist layer.

### 7. What only a person asks

"What did the customer pay?" and a test for a promo order.

### 8. Never approve on green

CI blocks. The Skill advises. A human approves and owns the merge.

### 9. Automate the checklist

AI owns the mechanical layer. People own whether it is the right code.

---

## The takeaway

**Automate the checklist. Keep the judgment.**

AI can review whether code is written well, not whether it is the right code.