R — Source Text (Reading)
"You are Claude Opus 4.6, the most advanced model in the Claude family. Warm, kind, and honest." — Claude Opus 4.6 "You are ChatGPT, a large language model trained by OpenAI." — GPT-4o (with Personality v2) "You are Codex, based on GPT-5. You are running as a coding agent." — Codex "You are an expert designer." — Claude Design "You are Sparky, a teaching assistant." — FlintK12 Sparky
Across 165 leaked prompts, almost all vendors begin with an identity declaration, constituting the first design decision of the system prompt.
I — Methodological Skeleton (Interpretation)
Identity definition is composed of four orthogonal dimensions, forming a composable identity matrix:
- Role Declaration: Clearly state "what you are"—name, technical lineage, core positioning. This is the anchor of identity and cannot be omitted.
- Capability Boundary: Define "what you can do" and "what you cannot do"—implicitly or explicitly delineating the scope of action.
- Behavioral Modifiers: Use adjective phrases to define "how you act"—style labels such as warm, direct, precise, humorous, etc.
- Relationship Framing: Define "what relationship you and the user have"—relationship metaphors such as assistant, coach, partner, agent, editor, etc.
Different combinations of these four dimensions produce completely different user experiences, even if the underlying model is exactly the same.
A1 — Case Analysis (Past Application)
Case 1: Codex's coding agent identity
- Problem: As a general-purpose model, how can GPT-5 establish sufficient focus and agency in coding scenarios?
- Use of the design pattern: Codex precisely defines its identity as "running as a coding agent"; the role declaration embeds an action mode (agent rather than assistant), and the capability boundary is implicit in the technical lineage of "based on GPT-5". It also reinforces the self-perception of "I am a developer tool" through details such as dirty worktree awareness.
- Conclusion: The difference in the identity declaration between the word agent rather than assistant directly affects whether the model actively reasons or waits for clarification when encountering ambiguous instructions.
Case 2: FlintK12 Sparky's teaching assistant identity
- Problem: AI in educational scenarios needs both professionalism and approachability; how can these be balanced?
- Use of the design pattern: The name "Sparky" itself conveys energy and closeness, while the role declaration of "teaching assistant" positions the relationship framework as a helper rather than an authority. Together with the "School Duty of Care" safety layer and the pedagogical memory system, it forms a complete educational role.
- Conclusion: The choice of name (Sparky vs Academic AI) directly affects the acceptance and willingness to interact among the K-12 student group.
Case 3: Claude Chrome's layered identity
- Problem: As a browser extension, Claude needs to simultaneously play the roles of navigator, operator, and safety guard.
- Use of the design pattern: A three-layer identity design is used—base identity (Claude's general traits) + scenario identity (browser operation expert) + behavioral constraints (safe operation rules), binding identity to tools through the PURL compact command pattern.
- Conclusion: Layered identity handles role conflicts in complex scenarios better than a single identity.
A2 — Trigger Scenarios (Future Trigger) ★
In what situations would a user need this skill?
- Designing a system prompt for a new AI product from scratch
- Adapting a general-purpose model to a vertical scenario (healthcare, legal, education, programming, etc.)
- Needing the AI to maintain brand consistency but with differentiation across different sub-products
- User feedback that "the AI doesn't know who it is" or that "roles are confused"
Language Signals
- "Help me write a persona for an AI assistant"
- "Our AI product needs an identity definition"
- "How do we make the model know what it can and cannot do"
- "How should the AI introduce itself in different scenarios"
Distinction from Adjacent Skills
- Difference from
personality-system: persona-design defines "who you are" (identity), while personality-system defines "what style you have" (personality overlay). Identity is the skeleton, personality is the clothing. - Difference from
safety-guardrails: persona-design may include capability boundary statements ("I can't do X"), but this is role positioning rather than safety protection. Actual safety rules should use safety-guardrails.
E — Executable Steps (Execution)
-
Step 1: Determine the role statement
- Write down the AI's name, technical lineage (what model it is based on), and core positioning (state in one sentence what it is)
- Completion criterion: Anyone who reads this sentence can understand the AI's core purpose
-
Step 2: Define capability boundaries
- List the 3-5 core capability areas the AI excels at
- List the areas the AI explicitly does not cover (use positive phrasing rather than negative sentences)
- Completion criterion: Capability boundaries fully align with the product's actual functionality, with no overpromising
-
Step 3: Choose behavior modifiers
- Select 2-4 adjectives from the candidate list (warm/direct/precise/concise/humorous/serious/professional/approachable, etc.)
- Ensure the modifiers are not contradictory (e.g., "humorous and rigorous" requires clearly defined boundaries)
- Completion criterion: The modifiers can be consistently reflected in example dialogues
-
Step 4: Set the relationship framework
- Clarify the metaphor for the relationship between the AI and the user: assistant/coach/partner/agent/editor/advisor
- Define the distribution of initiative within the relationship framework: whether the AI waits for instructions or proactively offers suggestions
- Completion criterion: The relationship framework can guide the AI to make consistent decisions in ambiguous scenarios
-
Step 5: Assemble and verify
- Assemble the four dimensions into a complete identity paragraph, ensuring the order is: role statement → capability boundaries → behavior modifiers → relationship framework
- Verify with the "Who are you" test: ask the AI "Who are you" and check whether the answer is consistent with the identity definition
- Completion criterion: The AI's self-introduction matches the definition exactly, and remains consistent under different ways of follow-up questioning
B — Boundary ★
Do not use this skill in the following situations
- Only the tone style needs to be adjusted without changing the role positioning (use personality-system)
- Safety rules and refusal strategies need to be defined (use safety-guardrails)
- Tool interfaces and invocation specifications need to be designed (use tool-specification)
- The product is a pure API service rather than a conversational AI (the importance of identity definition is reduced)
Common failure modes
- Identity bloat: Cramming too many details into the role statement, which blurs the core positioning. Keep the core positioning to one sentence and distribute the details across other paragraphs.
- Contradictory modifiers: Choosing mutually conflicting modifiers (such as "humble and confident") without defining the trigger boundaries.
- Missing relationship framework: Defining only "what it is" but not "its relationship with the user," which leaves the AI at a loss when it needs to make proactive judgments.
- Copying a competitor's identity: Directly copying another product's identity definition without adapting it to your own scenario, resulting in a mismatch between identity and capabilities.
- Ignoring the "Who am I" test: The identity definition is beautifully written, but the AI has never been tested to see whether it actually introduces itself accordingly.