AI-Assisted Code Review Directly in the Development Studio

In practice, the most interesting part is that the review does not run over the entire repository, but over a precisely defined patch file. This allows the model to focus on the actual change, without the surrounding noise, while still allowing the AI agent to use search tools and fetch related parts of the code into the context. The workflow is simple: in Visual Studio with GitHub Copilot, you prepare a prompt that must be run in agent mode, passing it the instruction to use the code review skill: 

By doing this, you save yourself from writing out this prompt, you don't have to remember its exact phrasing or the name of the skill, and in the development environment, you just trigger the prompt (e.g., in Visual Studio, "/" will offer you your prompts), wait a moment, and you already get the code review result. 

Inside the code-review skill, you define the entire workflow of how the code review should take place. For greater consistency, I prepared a simple PowerShell script with which the AI agent extracts the diff between the current branch and the branch you usually merge into: 

param( 
    [string]$Branch = (git rev-parse --abbrev-ref HEAD), 
    [string]$TargetBranch = "origin/dev", 
) 
$patchFile = Join-Path $PSScriptRoot "mr_review.patch" 
git diff "$TargetBranch...$Branch" | Out-File -FilePath $patchFile -Encoding utf8 
# Output path for Agent 
Write-Output $patchFile

The script creates a patch file containing all the changes against the target branch and returns its path. This path is subsequently used by the AI agent in the next step of the review. The skill file can look like this, for example: 

(In this section, you can define your own review rules. Typically, this involves checking security, performance, resource handling, API contracts, changes to application behavior, or internal coding guidelines.)

Thanks to a SKILL file defined in this way, you get a customizable code review over which you have a high degree of control, and you will receive relatively consistent results from the AI for the same model. 

Moreover, it is easy to change the target branch, for instance, when the review is not running against origin/dev but against another branch. You just need to add something like TargetBranch "targetBranchName" to the prompt. 

It is important to keep in mind that the whole process remains under the developer's control. AI is not an "autonomous reviewer" here, but a very fast assistant. It can point out weak spots and suggest specific points for verification, but, of course, it also makes mistakes from time to time. 

From practical experience, this approach has proven to be a very useful addition to a traditional code review. It does not replace an experienced reviewer, but it helps to quickly identify risky areas, shorten the review time, and increase the probability of catching bugs before they are delivered to the main branch.