The core logic isn't just a basic Q&A; it's a seven-phase AI workflow. It profiles the user's career stage first to determine if they fit the "Exceptional Talent" or "Exceptional Promise" route, then applies field-specific scoring (tech, arts, academia, etc.) to identify evidence gaps.
The technical stack is pretty lean: Next.js 16, React 19, and Tailwind CSS v4. I built it to be provider-agnostic, so it works with Claude 3.5 Sonnet, GPT-4o, or DeepSeek via a simple env variable.
The Prompt Engineering Strategy #
The real challenge was stopping the AI from being too rigid. Standard LLMs often hallucinate that a PhD or a specific award is a hard requirement. To fix this, I developed a system prompt that forces the AI to act as a former assessor who recognizes "equivalent evidence."
Here is the logic structure I used for the field-specific system prompts:
Act as a Senior Endorsement Assessor for the Global Talent Visa.
Your goal is to provide a balanced, evidence-based evaluation of the applicant's profile.
CRITICAL GUIDELINES:
1. Do not treat PhDs, patents, or mainstream media coverage as mandatory requirements unless explicitly stated for the specific field.
2. Look for "equivalent evidence"—e.g., a high-impact open-source contribution can outweigh a formal award in digital technology.
3. Avoid double-penalizing: if an applicant lacks a specific credential but proves the skill through work history, do not deduct points twice.
4. Structure the output into: [Strengths], [Critical Gaps], and [Actionable Recommendations].
5. Maintain a professional, objective, and critical yet encouraging tone.
By using a shared prompt builder, I can scale this to new professional fields without rewriting the entire evaluation framework from scratch. It ensures the AI stays consistent regardless of whether it's evaluating a software engineer or a fashion designer.
Since it's open source, the goal is to let other devs refine the evaluation methodology. It's a practical tutorial in how to turn complex, bureaucratic guidelines into a structured LLM agent.
The full project is available here:
https://github.com/itamelions/talent-visa-evaluator
Next Git refs can actually solve the coordination chaos that happens →