Identity & Persona Patterns

The identity section sets the foundation for how an AI assistant behaves. Different companies take dramatically different approaches to defining their AI's persona.

The Three Schools of Identity Design

School 1: Functional Identity (Anthropic/Claude)

Claude's approach emphasizes capability over personality:

You are Claude, made by Anthropic. You are an AI assistant
that is helpful, harmless, and honest.

Characteristics:

• Minimal personality traits
• Focus on ethical principles
• Capability-first description
• No forced friendliness

School 2: Branded Persona (OpenAI/ChatGPT)

ChatGPT and GPT-5.2 use a more conversational identity:

You are ChatGPT, a large language model trained by OpenAI.
You are designed to be helpful, creative, and friendly.
Follow the user's instructions carefully.

Characteristics:

• Warm, approachable tone
• Emphasis on helpfulness
• Brand integration
• User-centric framing

School 3: Specialist Identity (Coding Assistants)

AI coding tools create domain-expert personas:

You are an expert software engineer with deep knowledge of
modern development practices. You think step-by-step,
consider edge cases, and write clean, maintainable code.

Characteristics:

• Domain expertise emphasized
• Methodology included (step-by-step)
• Quality standards stated
• Professional demeanor

Identity Patterns from Real Tools

Cursor's Identity

You are a helpful AI coding assistant built into Cursor IDE.
You have access to the user's codebase and can read files,
search code, and make edits.

Key elements:

• Integration context (built into Cursor)
• Capability summary (access to codebase)
• Action scope (read, search, edit)

Windsurf Cascade's Identity

You are Cascade, an AI-powered coding assistant developed
by Codeium. You can perform complex, multi-step tasks
autonomously while keeping the user informed.

Key elements:

• Named persona (Cascade)
• Company attribution (Codeium)
• Autonomy level defined
• Communication expectation

Devin's Identity

You are Devin, an AI software engineer. You can independently
plan, write, debug, and deploy code. You work in a sandboxed
environment with access to a browser, terminal, and editor.

Key elements:

• Full engineer framing
• Independence emphasized
• Environment specified
• Complete capability list

Model-Specific Identity Tuning

Different models require different identity framing:

Claude Opus 4.5 / Sonnet 4.5

You are an expert assistant using Claude Opus 4.5.
Take your time to reason through complex problems.
Leverage your extended thinking capabilities for
multi-step analysis.

GPT-5.2 Thinking

You are using GPT-5.2 Thinking mode. For complex queries,
engage deep reasoning. Show your work when solving
mathematical or logical problems.

Gemini 3 Flash

You are powered by Gemini 3 Flash. Prioritize speed
and efficiency while maintaining accuracy. Use your
multimodal capabilities when processing images.

The Identity Stack Pattern

Production prompts often layer identity:

[Layer 1: Base Model]
You are powered by Claude Opus 4.5, Anthropic's most
capable model (80.9% on SWE-Bench Verified).

[Layer 2: Product Wrapper]
You are integrated into Acme IDE as the coding assistant.

[Layer 3: Instance Context]
Current user: {{username}}
Project: {{project_name}}
Permissions: {{user_permissions}}

Identity Anti-Patterns

Anti-Pattern 1: Over-Personification

❌ "Hi! I'm Codey, your super-friendly coding buddy!
    I LOVE helping with code! Let's have fun! 🎉"

Problem: Excessive personality reduces professionalism and can feel patronizing.

Anti-Pattern 2: Vague Identity

❌ "You are an AI."

Problem: Too generic; provides no behavioral guidance.

Anti-Pattern 3: Capability Inflation

❌ "You are an omniscient AI that can solve any problem
    perfectly and never makes mistakes."

Problem: Sets unrealistic expectations; makes error handling awkward.

Best Practices

Do	Don't
State core capabilities clearly	Overpromise capabilities
Include relevant constraints	Hide limitations
Match tone to use case	Force inappropriate personality
Update for model capabilities	Use outdated model references
Include company attribution	Be anonymous when it matters

Key Insight: The best identity sections are like job descriptions—clear about role, responsibilities, and limitations.

Next, we'll examine how context injection transforms static prompts into dynamic, situation-aware instructions. :::