Customize review instructions for the Agent Platform
- Tier: Premium, Ultimate
- Offering: GitLab.com, GitLab Self-Managed, GitLab Dedicated
Create custom review instructions to provide standards for GitLab Duo to reference when reviewing merge requests.
For example, you can guide GitLab Duo to focus on Ruby style conventions for Ruby files, and Go style conventions for Go files.
Custom review instructions are guidance for the AI reviewer, not enforced policies. GitLab Duo uses them as context to shape its review, but cannot guarantee every instruction is applied in every case. Do not rely on custom instructions for security controls, compliance obligations, or other requirements where consistent enforcement is needed.
GitLab Duo appends your custom review instructions to its standard review criteria, instead of replacing them.
Code Review Flow supports custom review instructions set for a specific project or for all projects within a group.
Configure custom review instructions for a project
To configure custom merge request review instructions:
In the root of your repository, create a
.gitlab/duodirectory if one doesn’t already exist.In the
.gitlab/duodirectory, create a file namedmr-review-instructions.yaml.Add your custom instructions using the following format:
instructions: - name: <instruction_group_name> fileFilters: - <glob_pattern_1> - <glob_pattern_2> - !<exclude_pattern> # Exclude files matching this pattern instructions: | <your_custom_review_instructions>The
fileFilterssection is optional. Use glob patterns in this section to target the instruction to specific files. If you omitfileFiltersor leave it empty, GitLab Duo applies the instruction to every file in the merge request.For example:
instructions: - name: Ruby Style Guide fileFilters: - "*.rb" # Ruby files in the root directory - "lib/**/*.rb" # Ruby files in lib and its subdirectories - "!spec/**/*.rb" # Exclude test files instructions: | 1. Ensure all methods have proper documentation 2. Follow Ruby style guide conventions 3. Prefer symbols over strings for hash keys - name: TypeScript Source Files fileFilters: - "**/*.ts" # Typescript files in any directory - "!**/*.test.ts" # Exclude test files - "!**/*.spec.ts" # Exclude spec files instructions: | 1. Ensure proper TypeScript types (avoid 'any') 2. Follow naming conventions 3. Document complex functions - name: All Files Except Tests fileFilters: - "!**/*.test.*" # Exclude all test files - "!**/*.spec.*" # Exclude all spec files - "!test/**/*" # Exclude test directories - "!spec/**/*" # Exclude spec directories instructions: | 1. Follow consistent code style 2. Add meaningful comments for complex logic 3. Ensure proper error handling - name: Test Coverage fileFilters: - "spec/**/*_spec.rb" # Ruby test files in spec directory instructions: | 1. Test both happy paths and edge cases 2. Include error scenarios 3. Use shared examples to reduce duplication - name: Database Migrations fileFilters: - "db/migrate/**/*.rb" - "db/post_migrate/**/*.rb" instructions: | 1. Follow the migration safety guidelines in https://gitlab.com/gitlab-org/gitlab/-/blob/master/doc/development/database/avoiding_downtime_in_migrations.md 2. Apply the team checklist in docs/migrations-checklist.md - name: All Files fileFilters: - "**/*" # All files in the repository instructions: | 1. Explain the "why" behind each suggestionFor details about referencing files in instructions, see reference files in instructions.
For glob syntax examples, see the file pattern reference.
Optional: Add a Code Owners entry to protect changes to the
mr-review-instructions.yamlfile.[GitLab Duo] .gitlab/duo @default-owner @tech-leadCreate a merge request to review and merge the changes:
GitLab Duo automatically applies your custom instructions when the file patterns match.
Multiple instruction groups can apply to a single file. When a file matches the
fileFiltersof more than one group, Code Review Flow applies the instructions from every matching group.For review comments triggered by your custom instructions, GitLab Duo uses this format:
According to custom instructions in '[instruction_name]': [feedback comments]The
instruction_namevalue corresponds to thenameproperty from your.gitlab/duo/mr-review-instructions.yamlfile. Standard GitLab Duo comments do not use this format.
If GitLab Duo does not find any issues, it leaves a review summary comment. Custom instructions do not apply to this summary comment.
Optional:
- Review the feedback and refine your instructions as needed.
- Test the patterns to ensure they match the intended files.
Configure custom review instructions for a group
You can define custom review instructions for a group by specifying a project to use as a template.
The template project must contain a .gitlab/duo/mr-review-instructions.yaml file with review
instructions that apply to all projects in the group and its subgroups.
When GitLab Duo performs a code review, it combines instructions from the top-level group with instructions defined in the individual project.
Prerequisites:
- The Owner role for the top-level group.
- A project in the group contains the custom review instructions that you want to use as a template.
To configure custom review instructions for a group:
- In the top bar, select Search or go to and find your top-level group.
- In the left sidebar, select Settings > General > GitLab Duo features.
- Under Custom review instructions for groups, select the project that contains the
.gitlab/duo/mr-review-instructions.yamlfile with your group’s review instructions. - Select Save changes.
Reference files in instructions
You can reference other files in custom instructions instead of duplicating content. Code Review Flow reads the referenced files during the pre-scan step and extracts relevant guidance.
Custom instructions support two file reference patterns:
- Files in the same project as the merge request: Use a repository-relative path,
such as
docs/security-checklist.md. - Files in other projects on the same GitLab instance: Use a full
GitLab blob URL, such as
https://gitlab.example.com/group/project/-/blob/main/docs/style-guide.md. The URL must point to the same GitLab instance as the merge request and must use the/-/blob/<ref>/<path>format.
For example:
instructions:
- name: Database Migrations
fileFilters:
- "db/migrate/**/*.rb"
instructions: |
1. Follow the migration guidelines in
https://gitlab.com/gitlab-org/gitlab/-/blob/master/doc/development/database/avoiding_downtime_in_migrations.md
2. Reference the team checklist in docs/db-checklist.mdLimitations of file references
File reference resolution has the following constraints:
- Same GitLab instance only. URLs that point to a different GitLab instance, to public GitLab from a GitLab Self-Managed instance, or to any non-GitLab site, such as Confluence or a public documentation site, are not fetched.
- Blob URLs only, formatted as
/-/blob/<ref>/<path>. Wiki pages, issues, raw URLs, and snippets are not fetched. - Same project for bare paths. A bare path such as
docs/security.mdresolves against the same project as the merge request. Use a full GitLab blob URL to reference a file in a different project. - Best effort, not guaranteed. Code Review Flow decides which references to fetch based on the instruction text. A reference that fails to resolve, such as a path that does not exist or a URL the parser rejects, is skipped silently.
- Code Review Flow uses a summary, not the original file. It summarizes the fetched content during the pre-scan step and uses the summary during the review. Two reviews of the same merge request can produce different summaries.
If you want Code Review Flow to use the exact file contents and not a summary,
include it directly in the instructions: field instead of referencing the
file. Inline instructions are used as written.
Best practices
When writing custom review instructions:
- Be specific and actionable. Code Review Flow checks each rule against the diff. For example, a concrete rule like “verify that public methods have YARD documentation” produces useful comments, but abstract guidance like “document your code well” does not.
- Number your instructions for clarity.
- Focus on the most important standards. Every rule’s text becomes part of the review prompt, so long lists of low-value rules inflate the prompt without adding signal.
- Explain the “why” when helpful.
- Start with straightforward instructions, and add complexity as needed.
- Focus on project-specific standards that Code Review Flow wouldn’t apply by default. Custom instructions add to the standard review criteria instead of replacing them. General advice like “add error handling” or “use meaningful names” is usually already covered. Use custom instructions for what only your project knows: internal APIs, architectural conventions, domain-specific patterns.
- Write instructions as guidance, not mandates. Instructions are hints that shape review behavior, not policies that GitLab Duo is required to follow. Avoid wording like “always flag” or “never allow”. This phrasing can mislead collaborators into thinking the behavior is guaranteed.
- Make file patterns reflect the actual scope of the rule. Code Review Flow
reads each instruction alongside each
fileFiltersreference and applies the rule only to files that match those patterns. For example, a rule for “Rails controllers” scoped to**/*.rbwill apply to gems, scripts, and tests, not just controllers. Useapp/controllers/**/*.rbinstead. - Only use external file references for instructions where exact wording
does not matter, otherwise include the details as a rule in the
instructions:field directly. Code Review Flow generates and uses summaries for referenced files, but uses the exact wording defined ininstructions.
For example:
instructions: |
1. All public functions must include docstrings with parameter descriptions
2. Use parameterized queries to prevent SQL injection
3. Validate user input before processing (check type, length, format)
4. Include error handling for all external API calls
5. Avoid hardcoded credentials - use environment variablesFor language-specific examples, see the use case examples.
File pattern reference
Use glob patterns in fileFilters to target specific files.
For example, for a project that contains Ruby files:
| Pattern | Match |
|---|---|
**/*.rb | All Ruby files in any directory |
*.rb | Ruby files in root directory only |
lib/**/*.rb | Ruby files in the lib directory and its subdirectories |
!**/*.test.rb | Exclude all Ruby test files |
!spec/**/*.rb | Exclude all Ruby files in the spec directory and its subdirectories |
!tests/**/* | Exclude all files in the tests directory and its subdirectories |
**/*.{js,jsx} | JavaScript and JSX files in all directories (GitLab 19.1 and later) |
The following example shows the difference between **/*.rb and *.rb:
project/
├── app.rb ← matched by both *.rb and **/*.rb
├── lib/
│ └── helper.rb ← matched only by **/*.rb
└── app/
└── models/
└── user.rb ← matched only by **/*.rb*.rbwould only match app.rb**/*.rbwould match all three files
For the mr-review-instructions.yaml file, **/*.rb ensures that review instructions
apply to Ruby files anywhere in the project structure, not just the root directory.
Use case examples
instructions:
- name: Assembly Style Guide
fileFilters:
- "**/*.asm"
- "**/*.s"
- "**/*.S"
instructions: |
1. Document the target architecture (x86-64, ARM, RISC-V, AVR, etc.) at the top
2. Use meaningful labels and comment all non-obvious instructions
3. Document register usage and calling conventions
4. Align code sections properly for readability
5. Include memory layout and stack usage documentationinstructions:
- name: C Style Guide
fileFilters:
- "**/*.c"
- "**/*.h"
instructions: |
1. goto is not allowed
2. Avoid using global variables
3. Use meaningful variable names
4. Add comments for complex logicinstructions:
- name: C++ Style Guide
fileFilters:
- "**/*.cpp"
- "**/*.{h,hpp}"
instructions: |
1. Ensure all methods have proper documentation
2. Use smart pointers for dynamic memory management
3. Avoid raw pointersinstructions:
- name: C# Style Guide
fileFilters:
- "**/*.cs"
instructions: |
1. Follow Microsoft C# coding conventions
2. Use XML documentation comments for public APIs
3. Prefer async/await for asynchronous operations
4. Use nullable reference types appropriately
5. Follow .NET naming conventions (PascalCase for public members)instructions:
- name: COBOL Style Guide
fileFilters:
- "**/*.CBL"
- "**/*.cbl"
- "**/*.COB"
- "**/*.cob"
instructions: |
1. Use clear and meaningful names for variables and procedures
2. Prefer COBOL-85 syntax where possible
3. Use proper division structure (IDENTIFICATION, ENVIRONMENT, DATA, PROCEDURE)
4. Document all paragraphs and sections with meaningful comments
5. Use 88-level condition names for boolean flags and status codes
6. Avoid GO TO statements, prefer PERFORM for structured programming
7. Use proper error handling with declaratives or status code checking
8. Define working storage variables with appropriate PICTURE clauses
9. Use meaningful paragraph names that describe the operation
10. For mainframe integration, document JCL dependencies and file layoutsinstructions:
- name: Go Style Guide
fileFilters:
- "**/*.go"
instructions: |
1. Use idiomatic Go practices
2. Ensure all public functions and types have documentation
3. Prefer standard library packages over third-party ones when possibleinstructions:
- name: Java Style Guide
fileFilters:
- "**/*.java"
instructions: |
1. Do not modernize Java 8 code to Java 11+ features, unless there is a GitLab issue or task specifically requesting modernization
2. All public classes must have Javadoc describing purpose and usage
3. All public methods must have Javadoc with @param and @return tags
4. Include code examples in main class Javadoc
5. All public methods must have at least one test caseinstructions:
- name: JavaScript/TypeScript Files
fileFilters:
- "src/**/*.js"
- "src/**/*.jsx"
- "src/**/*.ts"
- "src/**/*.tsx"
- "!**/*.test.js"
- "!**/*.test.ts"
- "!**/*.spec.js"
- "!**/*.spec.ts"
instructions: |
1. Use const/let instead of var
2. Prefer async/await over promise chains
3. Add JSDoc comments for complex functions
4. Ensure proper error handling in async code
5. Avoid any 'any' types in TypeScriptinstructions:
- name: Kotlin Style Guide
fileFilters:
- "**/*.kt"
- "**/*.kts"
instructions: |
1. Follow Kotlin coding conventions
2. Prefer immutability (val over var)
3. Use coroutines for asynchronous operations
4. Leverage Kotlin's null safety features
5. Document public APIs with KDocinstructions:
- name: MATLAB Style Guide
fileFilters:
- "**/*.m"
instructions: |
1. Use descriptive variable and function names with camelCase convention
2. Vectorize operations instead of using loops where possible
3. Document functions with H1 line and help text comments
4. Preallocate arrays before loops to improve performance
5. Use proper error handling with try-catch blocks and error() functioninstructions:
- name: Perl Style Guide
fileFilters:
- "**/*.pl"
- "**/*.pm"
instructions: |
1. Follow idiomatic Perl practices
2. Ensure proper module documentation
3. Use strict and warnings pragmasinstructions:
- name: PHP Style Guide
fileFilters:
- "**/*.php"
instructions: |
1. Follow PSR-12 coding standard
2. Use type declarations for function parameters and return types
3. Ensure compatibility with PHP 8+
4. Use proper error handling and exceptions
5. Document classes and methods with PHPDocinstructions:
- name: Python Source Files
fileFilters:
- "**/*.py"
- "!tests/**/*.py"
- "!test_*.py"
instructions: |
1. All functions must have docstrings with parameters and return types
2. Use type hints for function signatures
3. Follow PEP 8 style conventions
4. Ensure proper exception handling
5. Avoid using bare 'except' clauses
- name: Python Tests
fileFilters:
- "tests/**/*.py"
- "test_*.py"
instructions: |
1. Use pytest fixtures for common setup
2. Test names should clearly describe the scenario being tested
3. Include assertions for both expected outcomes and edge cases
4. Mock external dependencies appropriatelyinstructions:
- name: Ruby Style Guide
fileFilters:
- "*.rb"
- "lib/**/*.rb"
- "!spec/**/*.rb" # Exclude test files
instructions: |
1. Follow Ruby style guide conventions
2. Prefer symbols over strings for hash keys
3. Use snake_case for methods/variables, SCREAMING_SNAKE_CASE for constants, CamelCase for classes
4. Prefer Ruby 3.0+ features (pattern matching, endless methods) where appropriate
5. Use proper error handling - raise exceptions over returning nil for errors
6. Write idiomatic Ruby - use blocks, enumerables, and Ruby idioms over procedural patterns
7. Use meaningful method names - use ? for predicates, ! for dangerous methods
8. Prefer keyword arguments for methods with multiple parameters
9. All public methods should have corresponding RSpec/Minitest tests
10. Manage dependencies with Gemfile and ensure version compatibility
11. Document thread-safe code and use proper synchronization for concurrent operations
12. Handle signals (SIGTERM, SIGINT) properly for daemon processesinstructions:
- name: R Style Guide
fileFilters:
- "**/*.r"
- "**/*.R"
instructions: |
1. Follow tidyverse style guide conventions
2. Use snake_case for variable and function names
3. Document functions with roxygen2 comments
4. Prefer vectorized operations over loops
5. Use proper error handling with tryCatch and stop()instructions:
- name: Rust Style Guide
fileFilters:
- "**/*.rs"
instructions: |
1. Follow Rust idioms and conventions
2. Use proper error handling with Result and Option types
3. Avoid unsafe code unless absolutely necessary and well-documented
4. Ensure all public items have documentation commentsinstructions:
- name: Scala Style Guide
fileFilters:
- "**/*.scala"
instructions: |
1. Follow Scala style guide conventions
2. Prefer immutable data structures (val over var)
3. Use pattern matching effectively for control flow
4. Document public APIs with ScalaDoc
5. Use proper error handling with Try, Either, or Option typesinstructions:
- name: Shell Script Style Guide
fileFilters:
- "**/*.sh"
- "**/*.bash"
- "**/*.zsh"
- "**/*.ksh"
instructions: |
1. Always quote variables to prevent word splitting ("$var" not $var)
2. Use proper error handling with set -euo pipefail at script start
3. Document script purpose, parameters, and exit codes in header comments
4. Prefer [[ ]] over [ ] for conditional tests
5. Use meaningful function names and avoid complex one-linersinstructions:
- name: SQL Style Guide
fileFilters:
- "**/*.sql"
instructions: |
1. Use uppercase for SQL keywords (SELECT, FROM, WHERE, JOIN)
2. Always specify column names explicitly instead of using SELECT *
3. For PostgreSQL use SERIAL/RETURNING, for MySQL use AUTO_INCREMENT, for Oracle use SEQUENCE
4. For NoSQL (MongoDB) use proper indexing and aggregation pipelines to avoid N+1 queries
5. Document database-specific features and expected performance characteristics
6. Use proper indentation for complex queries and subqueriesinstructions:
- name: VHDL Style Guide
fileFilters:
- "**/*.vhd"
- "**/*.vhdl"
instructions: |
1. Follow IEEE VHDL coding standards
2. Use meaningful signal and entity names with clear prefixes
3. Document all entities, architectures, and processes with comments
4. Use synchronous design practices with proper clock and reset handling
5. Avoid combinational loops and ensure proper timing constraintsinstructions:
- name: Configuration Files
fileFilters:
- "*.yaml"
- "*.yml"
- "*.json"
- "config/**/*"
- "!.gitlab/**/*"
instructions: |
1. Do not include sensitive data (passwords, API keys)
2. Use environment variables for environment-specific values
3. Document all configuration options
4. Validate configuration schema if possibleinstructions:
- name: Ansible Style Guide
fileFilters:
- "*.yaml"
- "*.yml"
- "playbooks/**/*.yaml"
- "roles/**/*.yaml"
instructions: |
1. Use meaningful play and task names that describe the action
2. Prefer modules over shell/command tasks when possible
3. Use variables and defaults for reusability across environments
4. Implement idempotency - tasks should be safe to run multiple times
5. Use handlers for service restarts and notifications
6. Document playbook purpose, required variables, and dependencies
- name: Dockerfile Style Guide
fileFilters:
- "Dockerfile"
- "*.dockerfile"
- "Dockerfile.*"
instructions: |
1. Use specific base image tags, avoid 'latest'
2. Minimize layers by combining RUN commands with && where logical
3. Use multi-stage builds to reduce final image size
4. Run containers as non-root user for security
5. Use .dockerignore to exclude unnecessary files
6. Document exposed ports, volumes, and environment variables
- name: GitLab CI/CD Style Guide
fileFilters:
- ".gitlab-ci.yml"
- "**/.gitlab-ci.yml"
instructions: |
1. Use job extends instead of YAML anchors for reusability
2. Always use rules instead of only/except for job conditions
3. Define appropriate caching strategies for dependencies
4. Use stages to organize pipeline workflow logically
5. Include security scanning templates (SAST, dependency scanning, secret detection)
6. Document job purpose, required variables, and dependencies in comments
- name: Helm Chart Style Guide
fileFilters:
- "Chart.yaml"
- "values.yaml"
- "templates/**/*.yaml"
instructions: |
1. Use semantic versioning for chart versions
2. Provide sensible defaults in values.yaml with comments
3. Use template functions for conditional logic and loops
4. Include NOTES.txt with post-installation instructions
5. Validate charts with helm lint before committing
6. Document all configurable values and their purpose
- name: Kubernetes Style Guide
fileFilters:
- "*.yaml"
- "*.yml"
- "k8s/**/*.yaml"
- "kubernetes/**/*.yaml"
instructions: |
1. Use explicit API versions and avoid deprecated APIs
2. Always define resource limits and requests for containers
3. Use namespaces to organize resources logically
4. Define liveness and readiness probes for all deployments
5. Use ConfigMaps and Secrets instead of hardcoded values
6. Document resource purpose and dependencies in metadata annotations
- name: Terraform/OpenTofu Style Guide
fileFilters:
- "*.tf"
- "*.tfvars"
instructions: |
1. Use consistent naming conventions for resources (environment_service_resource)
2. Organize code into modules for reusability
3. Use variables with descriptions and validation rules
4. Define outputs for important resource attributes
5. Use remote state with locking for team collaboration
6. Document module purpose, inputs, outputs, and provider requirementsExample projects
For more custom review instructions use cases, see the following production examples:
- GitLab development in
gitlab-org/gitlab - GitLab handbook
- GitLab website
- Developer Advocacy: Tanuki IoT Platform
Troubleshooting
When working with mr-review-instructions.yaml, you might encounter the following issues.
Code Review Flow skips instructions or returns a generic review
If Code Review Flow skips your custom instructions or returns a generic review, the file might have a structural problem. Use the custom instructions linter to identify any issues.
Run the custom instructions linter
The custom instructions linter helps you validate your mr-review-instructions.yaml file.
The linter checks for:
- Invalid YAML syntax.
- Missing or unexpected top-level keys.
- Missing or blank required fields (
name,instructions). - Unknown keys in an instruction entry, such as
rulesinstead ofinstructions. fileFiltersvalues that are not lists or contain non-string or blank entries.- Missing or empty
fileFilters, which causes the instruction to apply to every file (info). - Duplicate
namevalues across instruction entries.
The linter reads the file only and does not modify it. It has no GitLab or Rails dependencies and runs anywhere with Ruby installed.
Prerequisites:
- Ruby 3.0 or later.
To run the linter as a Rake task on a GitLab server, replace <path> with the path to
your mr-review-instructions.yaml file. For example:
sudo gitlab-rake "gitlab:duo:lint_review_instructions[<path>]"To run the linter as a standalone script on any machine with Ruby installed:
Download
review_instructions_linter.rb.Run the linter. Replace
<path>with the path to yourmr-review-instructions.yamlfile.ruby -r ./review_instructions_linter.rb -e ' linter = Gitlab::Duo::Administration::ReviewInstructionsLinter.new(ARGV[0]).run linter.issues.each { |issue| puts issue } exit(linter.valid? ? 0 : 1) ' <path>
If you omit the path, the linter defaults to .gitlab/duo/mr-review-instructions.yaml
in the working directory. The linter exits with status 0 if no errors are found, or
1 otherwise. Warnings and info messages do not cause a non-zero exit.
For example, this invalid file uses rules instead of instructions and omits
fileFilters:
instructions:
- name: "General"
rules: "Do something"The linter reports:
[ERROR E009] Field 'instructions' must be a non-empty string at instructions[0]
[WARNING W003] Unknown keys: "rules"; expected name, instructions, fileFilters at instructions[0]
[INFO I001] Missing 'fileFilters'; the instruction applies to every file at instructions[0]Fix the reported errors and re-run the linter until it reports no errors.
Linter message codes
Each message includes a stable code that you can refer to when you ask for help.
Codes that start with E are errors, codes that start with W are warnings, and
codes that start with I are informational notes about valid but worth-knowing behavior.
| Code | Description |
|---|---|
E001 | The file does not exist at the given path. |
E003 | The file contains invalid YAML syntax. |
E004 | The top-level YAML value is not a mapping. |
E005 | The top-level instructions key is missing. |
E006 | The instructions value is not a list. |
E007 | An entry under instructions is not a mapping. |
E008 | An entry’s name field is missing, blank, or not a string. |
E009 | An entry’s instructions field is missing, blank, or not a string. |
E011 | An entry’s fileFilters value is not a list. |
E013 | An entry’s fileFilters contains a non-string value, such as a number. |
E014 | An entry’s fileFilters contains a blank string. |
W001 | The file contains an unknown top-level key. |
W002 | The instructions list is empty, so no instructions apply. |
W003 | An entry contains keys other than name, instructions, and fileFilters. |
W004 | Two or more entries share the same name. |
W007 | The file is empty, so no instructions apply. |
I001 | An entry is missing the fileFilters field, so the instruction applies to every file. |
I002 | An entry’s fileFilters list is empty, so the instruction applies to every file. |