Custom Instructions

Overview

Instructions are the rules that guide agent behavior. They live in the instructions array in drevon.config.json and are injected into every agent’s config file.

Instruction Structure

{
  "instructions": [
    {
      "id": "unique-id",
      "title": "Human-Readable Title",
      "content": "The actual instruction text that agents will follow.",
      "alwaysApply": true,
      "globs": ["**/*.ts"]
    }
  ]
}

Field	Required	Description
`id`	Yes	Unique identifier (kebab-case recommended)
`title`	Yes	Display title
`content`	Yes	The instruction text
`alwaysApply`	No	Whether to always apply (default: `true`)
`globs`	No	File patterns — used by Cursor for conditional rules

Writing Effective Instructions

Be Specific and Actionable

{
  "id": "error-handling",
  "title": "Error Handling",
  "content": "Use custom error classes extending BaseError. Always include error codes. Log errors with structured JSON. Never catch and silently ignore exceptions."
}

{
  "id": "errors",
  "title": "Errors",
  "content": "Handle errors properly."
}

Use Domain-Specific Rules

{
  "instructions": [
    {
      "id": "api-conventions",
      "title": "API Conventions",
      "content": "REST endpoints use kebab-case URLs. Always version APIs (/v1/). Return JSON with { data, error, meta } envelope. Use HTTP status codes correctly (201 for create, 204 for delete)."
    },
    {
      "id": "database-rules",
      "title": "Database Rules",
      "content": "Use Prisma for all database access. Never write raw SQL in application code. All queries go through the repository layer. Use transactions for multi-step operations."
    },
    {
      "id": "security",
      "title": "Security Standards",
      "content": "Validate all input with Zod schemas. Sanitize output to prevent XSS. Use parameterized queries (Prisma handles this). Never log sensitive data (passwords, tokens, PII)."
    }
  ]
}

Leverage Globs for Cursor

Cursor’s .mdc format supports conditional rules that only activate for specific file types:

{
  "instructions": [
    {
      "id": "react-components",
      "title": "React Component Guidelines",
      "content": "Use functional components only. Prefer composition over inheritance. Use forwardRef for reusable components. Co-locate styles, tests, and stories.",
      "globs": ["**/*.tsx", "components/**"],
      "alwaysApply": false
    },
    {
      "id": "test-patterns",
      "title": "Test Patterns",
      "content": "Use AAA pattern (Arrange, Act, Assert). One assertion per test. Use descriptive test names. Mock external dependencies, not internal modules.",
      "globs": ["**/*.test.ts", "**/*.spec.ts"],
      "alwaysApply": false
    },
    {
      "id": "config-files",
      "title": "Config File Rules",
      "content": "Add comments explaining non-obvious settings. Keep configs minimal. Don't duplicate defaults.",
      "globs": ["*.config.*", ".*rc*"],
      "alwaysApply": false
    }
  ]
}

Glob-targeted instructions generate separate .mdc files for Cursor. Other agents receive all instructions regardless of globs, since they don’t support conditional activation.

Example: Full-Stack Project

{
  "instructions": [
    {
      "id": "memory-protocol",
      "title": "Memory Protocol",
      "content": "Read all files in .drevon/memory/ at the start of every session. Write to them after significant actions.",
      "alwaysApply": true
    },
    {
      "id": "tech-stack",
      "title": "Tech Stack",
      "content": "Next.js 15 (App Router), TypeScript strict mode, Tailwind CSS, Prisma ORM, PostgreSQL, Vercel deployment. Use these unless explicitly told otherwise.",
      "alwaysApply": true
    },
    {
      "id": "frontend-rules",
      "title": "Frontend Rules",
      "content": "Use server components by default. Only add 'use client' when interactivity is needed. Use next/image for all images. Keep pages thin — business logic in server actions or API routes.",
      "globs": ["app/**", "components/**"],
      "alwaysApply": false
    },
    {
      "id": "backend-rules",
      "title": "Backend Rules",
      "content": "Route handlers in app/api/. Use Zod for request validation. Return consistent response shapes. Handle errors with try/catch and proper status codes.",
      "globs": ["app/api/**", "lib/services/**"],
      "alwaysApply": false
    }
  ]
}

Tips

Keep instructions focused

Each instruction should cover one concern. Don’t combine unrelated rules in a single instruction — use multiple instructions instead.

Reference existing docs

If you have a style guide or architecture doc, reference its path:

"content": "Follow the architecture described in docs/ARCHITECTURE.md. Read it before making structural changes."

Use memory for evolving context

Static instructions go in instructions. Dynamic, evolving context belongs in memory files. Use instructions for rules, memory for state.

Don't over-constrain

Too many rigid rules make agents less useful. Focus on the important guardrails and let agents use their judgment for the rest.

​Overview

​Instruction Structure

​Writing Effective Instructions

​Be Specific and Actionable

​Use Domain-Specific Rules

​Leverage Globs for Cursor

​Example: Full-Stack Project

​Tips

Overview

Instruction Structure

Writing Effective Instructions

Be Specific and Actionable

Use Domain-Specific Rules

Leverage Globs for Cursor

Example: Full-Stack Project

Tips