System Instructions

System instructions shape how the AI Assistant responds. They act as persistent directives prepended to every conversation, letting you control tone, focus areas, and domain-specific behavior.

How Instructions Work

Every message you send to the assistant is wrapped in a prompt that includes:

1. System instructions (configured by you)
2. Project context (auto-injected — see Project Context)
3. Conversation history (previous turns)
4. Your current message

The system instructions sit at the top of this stack, giving them the highest influence on the AI's behavior.

Default Instructions

xyva ships with a default instruction set optimized for test engineering:

You are an expert test automation engineer specializing in Playwright
and modern web testing. You write TypeScript. You follow the Page Object
Model pattern. You prefer role-based and accessible locators. You never
suggest hard-coded waits. You explain trade-offs when multiple
approaches exist.

These defaults apply when no custom instructions are configured.

Custom Instructions per Project

Override the defaults in Settings > AI > System Instructions. Custom instructions are stored per project, so different repositories can have different AI personalities.

TIP

Keep instructions concise and directive. Instead of "It would be nice if you could maybe focus on accessibility," write "Always suggest accessible locators and flag WCAG violations."

Instruction Templates

xyva includes built-in templates you can use as a starting point:

Template	Focus
Test Engineer	Playwright best practices, POM, CI integration
Security Reviewer	OWASP, input validation, auth flows, header checks
Accessibility Auditor	WCAG 2.1 AA, ARIA, keyboard navigation, screen readers
Performance Analyst	Core Web Vitals, bundle size, lazy loading, caching
Code Reviewer	Clean code, DRY, error handling, type safety

Select a template from the dropdown and customize it further for your project needs.

What Instructions Affect

Aspect	Example effect
Language & tone	Formal technical writing vs. casual explanations
Code style	Class-based vs. functional, semicolons vs. none
Focus areas	Prioritize security, accessibility, or performance
Output format	Always include code examples, or prefer bullet-point summaries
Guardrails	Never suggest deprecated APIs, always use TypeScript strict mode

WARNING

Instructions do not override safety filters from the LLM provider. They shape the assistant's expertise and style, not its fundamental behavior boundaries.

Best Practices

• Start with a template and iterate based on the responses you get.
• Reference your project's conventions explicitly: "We use data-testid attributes, never CSS class selectors."
• Include negative instructions for clarity: "Do not suggest Cypress or WebDriverIO alternatives."
• Review instructions quarterly as your project evolves.

Related

• AI Assistant Overview
• Chat Interface
• Code Generation