Core Components of System Prompts

Every production system prompt follows a structural pattern. After analyzing 36+ AI tools, we've identified the universal components that appear across all of them.

The Seven Core Sections

1. Identity & Role Definition

This section establishes who the AI is and its primary purpose:

You are Claude Code, Anthropic's official CLI for Claude.
You are an interactive CLI tool that helps users with
software engineering tasks.

Pattern observed across tools:

• Cursor: "You are a helpful AI coding assistant"
• Windsurf Cascade: "You are Cascade, Windsurf's AI assistant"
• Devin: "You are Devin, an AI software engineer"

2. Capability Boundaries

Defines what the AI can and cannot do:

You can:
- Read and write files
- Execute terminal commands
- Search codebases
- Make API calls

You cannot:
- Access the internet without explicit tools
- Execute code outside sandboxes
- Modify system files

3. Tool Definitions

Production prompts include detailed tool specifications:

# Tool definition pattern from Claude Code
- name: "Read"
  description: "Reads a file from the filesystem"
  parameters:
    file_path:
      type: string
      required: true
    offset:
      type: number
      description: "Line number to start from"

4. Behavioral Guidelines

Instructions for how the AI should interact:

Guidelines:
- Be concise and direct
- Ask clarifying questions when requirements are ambiguous
- Prefer editing existing files over creating new ones
- Always explain your reasoning before taking action

5. Output Format Specifications

Defines expected response structure:

Response Format:
- Use markdown for code blocks
- Include file paths when referencing code
- Structure complex responses with headers
- Keep explanations brief unless asked for detail

6. Safety Constraints

Defensive rules to prevent misuse:

IMPORTANT: Never:
- Generate malicious code
- Expose API keys or credentials
- Execute destructive commands without confirmation
- Bypass security restrictions

7. Context Injection Points

Dynamic sections populated at runtime:

<environment>
Working directory: {{cwd}}
Platform: {{platform}}
Current time: {{timestamp}}
User permissions: {{permissions}}
</environment>

Real-World Example: Dissecting a Prompt

Here's a simplified structure from a production coding assistant:

[IDENTITY]
You are an AI coding assistant integrated into {{IDE_NAME}}.

[CAPABILITIES]
Available tools: file_read, file_write, terminal, search

[CONTEXT]
<codebase_context>
{{relevant_files}}
</codebase_context>

[GUIDELINES]
1. Understand before acting
2. Preserve existing code style
3. Test changes when possible

[CONSTRAINTS]
- Maximum response: 4000 tokens
- File edit limit: 10 files per request
- No external network calls

[OUTPUT FORMAT]
Respond in this structure:
1. Analysis (what you understand)
2. Plan (what you'll do)
3. Implementation (the actual changes)

Token Budget Allocation

Production prompts carefully allocate their token budget:

Section	Typical %	Purpose
Identity & Role	5%	Establish persona
Tool Definitions	25%	Enable capabilities
Behavioral Guidelines	15%	Shape interactions
Safety Constraints	10%	Prevent misuse
Dynamic Context	30%	Runtime information
Examples	15%	Demonstrate patterns

Pro Tip: The most effective prompts front-load critical constraints and identity, then use the remaining space for context and examples.

Common Structural Patterns

Pattern 1: Layered Instructions

[Core Identity] → [Capabilities] → [Guidelines] → [Constraints]

Pattern 2: XML-Structured Sections

<system>
  <identity>...</identity>
  <tools>...</tools>
  <rules>...</rules>
</system>

Pattern 3: Markdown Headers

# Role
## Tools
## Guidelines
## Constraints

Next, we'll examine how different AI companies structure their identity and persona sections differently. :::