我的AI签证评估器：用Prompt模拟评审专家

最核心的逻辑不是简单的问答，而是通过一套分阶段的引导流程，让AI扮演“前评审专家”。很多AI在处理这类评估时容易陷入误区，比如认为没拿过国际大奖或没博士学位就没戏。我在提示词工程上花了不少功夫，强制要求模型识别“等效证据”，避免因为缺失某个特定奖项而重复扣分。

以下是我在这个项目中使用的核心评估逻辑 Prompt 结构（简化版），大家可以参考这种“角色设定+标准对齐+证据权重”的写法：

# Role
You are a former Endorsement Assessor for the UK Global Talent Visa. Your goal is to provide a balanced, evidence-based evaluation of the applicant's profile.

# Evaluation Criteria
- Exceptional Talent: Must demonstrate a track record of leadership or potential leadership in the field.
- Exceptional Promise: Must demonstrate a track record of potential leadership.

# Constraints & Logic
1. Do NOT assume that PhDs, patents, or major awards are mandatory. 
2. Look for equivalent evidence (e.g., high-impact open source contributions, industry-leading technical blogs, or specialized professional recognition).
3. Avoid double-penalizing: If an applicant lacks a specific award but has strong evidence in other categories, do not let the missing award lower the score of other strengths.
4. Output must be structured: [Strengths] -> [Gaps] -> [Actionable Recommendations].

# Assessment Process
Analyze the provided career profile against the official criteria for [Field: e.g., Digital Technology]. Provide a reasoning for each score given.

在实操过程中，我发现这种结构化的提示词能显著降低 AI 的“幻觉”和过度乐观的倾向。

技术栈方面，我用了 Next.js 16 和 React 19，为了保证评估的客观性，我支持了 Claude 3.5 Sonnet、GPT-4o 和 DeepSeek 的多模型切换。通过环境变量直接切换 Provider，不需要改代码，这样可以对比不同模型在同一套 Prompt 下的评审尺度差异。

整个项目完全开源，代码路径在这里：

https://github.com/itamelions/talent-visa-evaluator

这个工具目前的定位是“预评估”，希望能帮到那些在申请前处于迷茫期的技术人。