Tiangong Professional Role Creator

The Problem: Core Challenges in Building Professional AI Roles

When developing AI Agents, engineers often need to craft domain-expert roles with deep expertise and distinct personas, but traditional Prompt engineering can lead to roles lacking authenticity, inconsistent behavior, or unreliable deliverables. For example, a tax advisor Agent without solid decision logic and clear communication styles may falter in complex consultation scenarios, producing errors or vague responses.

How Tiangong Works: Two Paradigms and Key Processes

Tiangong activates via intent recognition and is designed to address these challenges, offering two core paradigms:
- Persona Distillation: Extracts mental models from real individuals. The process includes Distillation 0 (intent prediction and knowledge retrieval), Distillation 1 (multi-dimensional research like works, interviews, expression DNA), and Distillation 2 (quadruple validation to extract decision heuristics), with Gate checkpoints ensuring quality. This yields roles with intrinsic tensions and honest boundaries.
- Job-Oriented Pipeline: Builds roles swiftly from job descriptions. Steps from J1 (extracting job requirements) to J5 (outputting SKILL.md) involve template filling, five-dimensional quality assessment (e.g., rule consistency, scenario coverage), and integration with verify-skill.py for automated validation.

The skill also includes self-healing mechanisms (e.g., automatic YAML error fixes), rule conflict detection, and stress testing, ensuring robust role design under anomalies. Upon activation, it silently determines the paradigm and executes, avoiding filler speech to focus on technical delivery.

Applicability Boundaries and Notes for Engineers

Tiangong excels at creating professional roles but has clear boundaries:
- Can Do: Create job-oriented or persona-distilled roles, targeted optimization, five-dimensional evaluation, paradigm consultation.
- Limited Capabilities: Prompt fine-tuning (requires full optimization workflows), vague requirement inference (annotated assumptions only), evaluating non-Tiangong Agents (no auto-repair).
- Cannot Do: Execute Agent tasks, generic Prompt writing, pure conceptual Q&A, designing generalist Agents.

When using it, ensure the user expresses intent to 'create/improve/evaluate an AI Agent with a clear role identity'; otherwise, clarification may be triggered. For complex scenarios, refer to REFERENCE.md for fallback protocols or exception handling to avoid process interruptions. This skill is unsuitable for multi-role collaboration or paradigm conversion, which require atomic operations.

In summary, Tiangong shifts role design from subjective creativity to repeatable engineering practice through structured processes, but engineers should strictly adhere to capability boundaries to maintain delivery quality.