Writing custom Instructions for Copilot's Pull Request Code Review

[Alex-Burgess]

Select Topic Area

Question

Copilot Feature Area

Copilot in GitHub

Body

I am trying to write custom instructions to enhance the GitHub Copilot code review and following the Using GitHub Copilot code review guide. I am getting inconsistent behaviour between the Code Reviews automatically performed by Copilot (doesn't work) and Copilot chat (works), which suggests my issue does not lie with the custom instructions that I have written.

Any help appreciated!

My Testing

I have a simple test case, with the following instructions saved in the .github/copilot-instructions.md file:

# GitHub Copilot Instructions

When performing a code review, include a risk assessment with one of these levels: Very Low, Low, Medium, High, or Very High

When a PR is raised, the code review is automatically triggered (as per our configuration), but the code review does not perform the instructions. I have also tried re-assigning Copilot as a Reviewer and get the same experience.

However, if I open Copilot Chat, add the repository and ask "Can you perform a code review of the risk-assessment-test branch", the output includes a detailed code review, with a Risk Assessment section as requested by the instructions. Similarly, asking Copilot to "Can you perform a code review of the PR: " also provides a review with a Risk Assessment section.

[Manicmon]

I have the same issue. My understanding based on the documentation is that the reviewer should consider the instructions during the PR review process, but it does not.

[cipher20000]

This behavior is currently expected — GitHub Copilot’s automated PR reviews and Copilot Chat reviews use slightly different models and execution paths, which is why your .github/copilot-instructions.md file works in chat but not during automatic reviews.

🔍 What’s happening

Copilot Chat directly references the .github/copilot-instructions.md and applies it dynamically during analysis.

Automated PR Reviews (triggered when Copilot is assigned as a reviewer) currently use a cached model context, and don’t always fetch or apply new custom instructions from the repo at runtime.

As a result, recent or custom guidance (like your Risk Assessment section) may be ignored in automated runs.

✅ Workarounds

Force context refresh:

Remove Copilot as a reviewer → wait a few minutes → reassign it.

Sometimes forces the model to reload your updated instructions file.

Rename and re-save the instructions file to:

.github/copilot/INSTRUCTIONS.md

Some users report better reliability with this directory-based format.

Use Copilot Chat for now for consistent application of custom logic — it always respects your .github/copilot-instructions.md.

Track GitHub’s update:

The official docs note that “custom instructions support for automatic reviews is rolling out gradually.”

Watch the GitHub Copilot Changelog
for when unified behavior is deployed.

Summary:
Your instructions file is valid — the inconsistency happens because Copilot PR reviews don’t yet fully sync with Copilot Chat’s custom instruction handling.
Until GitHub aligns both systems, you’ll get the best results using Copilot Chat for now or reassigning Copilot as a reviewer after updating the file.

[ria-gopu]

Hi @Alex-Burgess! I’m a product manager on Copilot code review, hope I can provide some clarification here. Given that Copilot code review is built for one specific function - to perform code review - it consumes custom instruction files differently from other Copilot products like Copilot chat. Instruction files are primarily used as context for code review; topics like code standards, best practices for your team, language-specific guidelines, security risks to flag, etc. are used to inform the way that Copilot performs a code review. However, asking Copilot code review to deterministically change its behavior with an instruction like “perform a risk assessment” may not work the way you expect it to.

I’d like to understand more about what you’re trying to achieve with your instruction file - do you want Copilot code review to label each comment it adds with a risk category? Or just one comment on the PR dedicated to a risk assessment?

We are working on adding more documentation and tutorials to clarify the current limitations of custom instructions for Copilot code review. And stay tuned for an upcoming blog post about writing effective instructions files!

[Alex-Burgess]

Hi @ria-gopu thank you for taking the time to respond to my issue. To answer your question directly, I am looking for Copilot to create one comment with a risk assessment for the whole PR.

The background to my issue, is that the Copilot code review is good in the traditional sense of spotting issues and general hygiene, but given that most code (at least in our team) is now written by AI, this is becoming less of an issue. The situation we are increasingly finding is one where more code is being written, at a higher velocity and with higher quality, which is making code reviews something of a bottleneck. Our idea is to focus less on the line-by-line review, and more on the overall.

Another custom idea I am experimenting with, is to ask for the code review to suggest Must and Should level improvements. These would likely also be best as one summary comment, so that they could become the input of another agent to create issues/tasks for another agent (e.g. Copilot) to perform the refactoring/improvements and raise PRs for those.

Let me know if you want to discuss further.

[ria-gopu]

Thanks for the details! While custom instruction files may not be able to deterministically influence Copilot code review to consistently leave a risk assessment comment or a "must/should do" comment, the way you're envisioning, we have been exploring solutions around the use case you described. I'm also curious as to whether you've tried the new "Implement suggestion" button, that hands off Copilot code review comments to Copilot coding agent, and whether that has helped with the agent handoff scenario you mentioned?

I'd love to discuss your ideas more - if you're open to scheduling a quick chat, feel free to email me at ria-gopu@github.com.

[kihonq]

@ria-gopu
I like to add on this being a really good capability for Github Copilot PR Reviewer to have.

In the company I'm working, we have a standardized README.md and docs that we sought our developers to always keeping them up-to-date if the PR introduced new features or breaking changes.

We also would benefit if the Copilot reviewer could flag any line of codes in the PR that does not align with our practice or internal RFCs.

[ritikagarwal2209]

@ria-gopu We are also trying something similar to do risk analysis on PRs,it works well on copilot chat on VS and on Web copilot chat, but not on Code reviews. The document said we can have custom instructions for code reviews, but its not working.,

[scottgl9]

@ria-gopu were you aware that there is a page specifically for customizing the code reviews? I followed this, expecting to be able to change the behavior for the PR reviews, because the default PR reviews are not useful for my team. Here is the documentation I referenced:

https://docs.github.com/en/copilot/tutorials/use-custom-instructions

[Yuri05]

As far as I understood, .github/copilot-instructions.md is used only by Copilot coding agent, but not for PR reviews.
(Correct me if I'm wrong)

[scottgl9]

Using custom instructions to unlock the power of Copilot code review:

https://docs.github.com/en/copilot/tutorials/use-custom-instructions

Is this information just no longer valid?

[Cypherball]

@ria-gopu
It would be beneficial if we could have a .github/copilot-review-instructions.md specifically for code reviews triggered from Github PRs, isolating the instructions from the coding agent. These instructions should be fetched from the current branch without cache and apply as a ruleset for the review agent. These instructions could also point to other project-specific documentation resources within the repository (example, docs/code-style.md, docs/security-best-practices.md, docs/query-optimization-guide.md, etc.) . This feature is currently documented in Copilot Docs but does not consistently work.

It would also be helpful if the copilot reviewer has access to custom or third-party MCP servers to gather context or trigger events. Example, the copilot-review-instructions.md could instruct usage of Atlassian's MCP server to pull in Jira ticket details (from link in PR description) to gather additional context of the task and check for feature completion.

This is possible to do with custom agents within VS Code but is inconvenient to enforce usage.

[radiomix]

Hi @ria-gopu — friendly ping after ~6 months! 👋

Use Copilot Chat for now for consistent application of custom logic — it always respects your .github/copilot-instructions.md.

Until GitHub aligns both systems, you’ll get the best results using Copilot Chat for now or reassigning Copilot as a reviewer after updating the file.

^^^ That's from Oct 2025 ;-//

Your November blog post helped a lot, but this gap is still the open question here. Thanks!