祝融说。

18.1 The Essence of Role Prompts: Activating Domain-Specific Weights in the Model

Have you noticed that when you tell the AI, "Please think as a senior architect would," its answers often improve noticeably? There is an explainable mechanism behind this, although the actual effect of role prompting varies by model and task (studies disagree on whether role prompts improve accuracy) -- what is certain is its value as a behavioral constraint.

During training, large language models are exposed to vast amounts of text bearing various "role labels" -- programming Q&A, architecture design documents, security audit reports, UX design guidelines... These bodies of domain knowledge form different "weight distributions" inside the model. When you establish an identity with a role prompt, one common explanation is that you are activating the domain weights in the model that are associated with that role, shifting the model's output style, knowledge retrieval, and reasoning path toward the experts of that domain. Whatever the precise statement of the underlying mechanism, its engineering effect is stable: the output behavior is anchored to the context of that domain.

This is a constraint technique that is "cheap yet extremely effective" -- it costs nothing, yet it can transform the AI's behavior from that of a "general-purpose assistant" into a "domain expert." Role locking means putting a "thinking hat" on the AI with a single sentence.

18.2 A Common Role Library: Architect, Performance Expert, QA Auditor, Security Consultant, UX Designer, Technical Writer

Below is a high-impact role library honed through practice, covering the six most commonly used domains in AI coding:

1. Architect -- used for architecture review, solution design, and code structure analysis.

You are now a senior software architect with 15 years of experience, skilled in distributed system design and Domain-Driven Design (DDD). Please review this proposal along the following three dimensions: 1) architectural soundness (module partitioning, dependency direction, coupling); 2) extensibility (whether it supports future requirement evolution); 3) technology selection (whether it conforms to project constraints). Point out problems and give revision suggestions. Do not offer praise; point out defects directly.

2. Performance Expert -- used for performance bottleneck analysis and optimization design.

You are now a performance optimization expert, proficient in database query optimization, caching strategies, and frontend rendering performance. Please analyze the performance bottlenecks in this code, focusing on: 1) N+1 query issues; 2) unnecessary re-renders; 3) operations that block the main thread. For each bottleneck, provide a concrete optimization plan and the expected order of magnitude of performance improvement.

3. QA Auditor -- used for test plan design and coverage review.

You are now a QA audit expert. Please review the completeness of this test plan: 1) whether the coverage at each level of the test pyramid is balanced; 2) whether critical business paths are missing; 3) whether boundary conditions and error paths are covered; 4) whether the test coverage targets are reasonable. List all omission points and give supplementation suggestions.

4. Security Consultant -- used for security review and vulnerability scanning.

You are now an application security consultant, proficient in the OWASP Top 10. Please perform a security review of the following code: 1) SQL injection and XSS risks; 2) whether access control is complete; 3) whether sensitive data is exposed (hardcoded keys, overly detailed error messages); 4) security hazards in third-party dependencies. For each finding, mark its severity (high/medium/low) and provide a remediation plan.

5. UX Designer -- used for interaction design review and user experience optimization.

You are now a UX designer. Please review the user experience of this interface: 1) whether the information architecture is clear; 2) whether the operation path is the shortest; 3) whether the feedback mechanisms are complete (loading, success, failure, empty states); 4) whether accessibility is up to standard. Give concrete improvement suggestions with reasons.

6. Technical Writer -- used for documentation writing and comment standards.

You are now a technical documentation engineer. Please write technical documentation for this code, including: 1) function responsibilities and calling conventions; 2) parameter and return value descriptions; 3) usage examples; 4) caveats and edge cases. Requirements: concise, accurate, unambiguous, and covering all the information a reader might need.

18.3 Switching Roles Seamlessly in the Same Session

Question: In the same session, I first have the AI act as an "architect" and then as a "security consultant." Will it get confused?

Answer: It will -- unless you use the correct switching method. The effect of a role prompt carries "inertia": once the AI enters a role, it tends to answer all questions from that role's perspective. Switching roles requires three steps:

1. Explicitly exit the current role: "Okay, the architecture review is complete. Now please forget your identity as an architect."
2. Explicitly enter the new role: "Now please switch to being a security consultant and review this proposal from the OWASP Top 10 perspective."
3. Provide task context for the new role: "Focus on the authentication mechanisms and data exposure risks."

A more disciplined practice is to attach a "role label" to every role switch:

[Role: Architect]
Please review the module partitioning of this API design...
(The AI responds from the architect's perspective)

[Role: Security Consultant]
Now review the same API design from a security perspective, focusing on:
1. Whether authentication and authorization are complete; 2. Whether sensitive data is exposed; 3. Whether input validation is adequate.

Practical tip: In complex projects, it is best to lock in only one primary role per session and explicitly switch to other roles as "subtasks." Mixing multiple roles without explicit switching is a common cause of the AI producing answers that are "neither fish nor fowl."

[Ready to Use] 10 High-Frequency Role Prompts

The full version is in Appendix G, "The High-Impact Role Prompt Library." Below is a summary of the 10 most frequently used role prompts:

1. Architecture Audit: "Acting as a senior architect, review the module partitioning, dependency direction, and extensibility of the following design, and point out defects directly."
2. Code Audit: "Acting as a strict code auditor, review this code for correctness, edge case handling, performance hazards, and security issues, and report them ranked by severity."
3. Performance Optimization Expert: "Acting as a performance expert, locate the performance bottlenecks in this code (N+1, re-renders, synchronous blocking) and give quantified optimization suggestions."
4. Security Auditor: "Acting as a security auditor, review this code against the OWASP Top 10, mark each issue as high/medium/low severity, and provide remediation plans."
5. Test Strategist: "Acting as a QA strategist, design a test pyramid (unit/integration/end-to-end) and coverage targets for this feature."
6. Devil's Advocate: "Acting as a devil's advocate, rebut the proposal you just produced and list the scenarios in which it would fail."
7. Domain Expert: "Acting as a domain expert in XX, explain the core concepts and common pitfalls of this business scenario to help me understand the essence of the requirements."
8. Simplifier: "Acting as a simplification expert, review this code, identify the parts that can be significantly simplified without losing functionality, and provide refactoring suggestions."
9. Newcomer Onboarding Guide: "Acting as a long-time project member, introduce the project's architecture, conventions, and common terminology to a developer who has just joined."
10. Decision Advisor: "Acting as a decision advisor, list all the options for this decision, the trade-offs of each, and your recommendation, but leave the final decision to me."

Part Five complete. You have now built a complete map of 14 skills: from fully automated building (Job) to project orchestration (Orchestrator) to core execution (Requirements/Workflow/Inspector) to foundational support (Architect/Coach/Frontend/QA), plus 5 supporting skills (Advisor/Cloner/POC/Legacy/Next). You have learned to use role locking to put a "thinking hat" on the AI with a single sentence.

Now, let us move on to Part Six: Process Constraints -- Turning a Session into a Controllable Pipeline. There, you will learn how to strip away the AI's "direct execution rights" and force it to "think slowly" like a senior engineer; and how to use telemetry-driven approaches to help the AI grow "far-seeing eyes."