Standardizing AI-Assisted Development with Github Copilot
Context
In a large-scale environment with multiple Scrum teams contributing to a shared repository maintaining consistency is a challenge. Our team follows restrict requirement for Accessibility (A11y), Responsive Web Design (RWD), and standardized Material UI (MUI) implementation.
Problem
While GitHub Copilot increased our productivity, it introduced "consistency debt". Because Copilot learns from existing code, it frequently suggested legacy anti-patterns or generated manual CSS for layouts that should have been handled by MUI components. This led to bloated PRs and manual review cycles.
My Role
I identified recurring inconsistencies in PRs and explored ways to reinforce coding standards through automation. Since the team was already using GitHub Copilot to enhance productivity, I proposed leveraging Copilot Instructions to guide code generation and reviews.
Challenges
- The Echo Chamber Effect: Copilot mirrored legacy anti-patterns found in the codebase.
- Library Neglect: AI often bypassed our MUI component library, recreating layouts from scratch.
- Platform Limitations: Unlike GitHub, Azure DevOps lacks native AI-assisted code review capabilities as of I'm writting this (February 2026).
Approach
I researched and implemented a structured .github/copilot-instructions.md (and related configurations):
| Instructions | Purpose |
|---|---|
| General | Global standards across the tech stack. |
| Repository Specific | Enforcing the use of MUI components over custom CSS and adhering to specific project folder structures. |
| Commit logic | Standardizing commit messages format to ensure documented code. |
| Code Review | Prompts to help developers self-review based on the standard of the specific repo. |
Knowledge Transfer (KT): I lead a team KT across scrum teams with total of 45 attendees. I emphasize the "Human-in-the-Loop" philosophy. The Golden Rule: Never commit code you do not fully understand.
Outcomes and Impact
The implementation of these instructions transformed our workflow:
- Faster PR Cycles: Team feedback indicated that code review dropped to an average of 30 minutes per PR.
- Enhanced Consistency: A noticeable decrease in "custom CSS" bloat as Copilot began prioritizing MUI-native patterns.
- Standardized Quality: Reduced back-and-forth during the review phase regarding basic naming and structural conventions.
Learnings
- The Optimization: We discovered that including "Optimization" in the default instructions even we prompt to provide judgement cause the AI to over-engineer simple logic.
- Judgement vs. Automation: I learned that instructions requiring judgement like (performance vs. readablity) are better suited for template prompt rather than Default System Instructions. This give developers control over when to optimize.