AI Smart Exam Generator - Interview Edition

Purpose

Provide a repeatable multi-agent role-play workflow (single model session or sequential steps) to produce high-quality mock technical interview papers. Align any backend or tooling work with the existing blog-community-backend patterns documented below.

When to Load This Skill

Load when the user mentions exam generation, interview questions, mock interviews, AI exam generators, test papers, or technical interview preparation that requires structured output and reviewer-quality gates.

Project Architecture Alignment (blog-community-backend)

When implementing or extending server-side generation, match the existing stack and patterns.

• Object storage: Reuse FileService / FileServiceImpl with injected MinioClient; do not introduce a separate MinIO service abstraction unless the codebase already adds one. Bucket and credentials come from minio.endpoint, minio.access-key, minio.secret-key, minio.bucket-name in application.yml (see MinIOConfig for the MinioClient bean).
• Chat model: Configure under spring.ai.dashscope (API key AI_API_KEY, base URL AI_BASE_URL, default chat options model / temperature / max-tokens via env such as AI_MODEL_PRACTICE_DEFAULT). Do not set base-url to a .../compatible-mode root; keep the native gateway root as in existing comments in application.yml.
• Agents: Follow the same builder style as PracticePaperAgentOrchestrator and PracticeGradeReactAgentRunner: inject @Qualifier("dashScopeChatModel") ChatModel, use ReactAgent.builder().name(...).model(chatModel).instruction(...).outputKey(...).enableLogging(true).build(), override models with DashScopeChatOptions.builder().model(...).temperature(...).maxToken(...).build() passed to .chatOptions(...), and use ParallelAgent / SequentialAgent from spring-ai-alibaba-agent-framework when splitting specialist work.
• Observability: Reuse timing hooks such as PracticeAgentTimingModelHook and PracticeAgentTimingRecorder when adding new pipeline stages so logs stay consistent.
• Repository rules: .cursor/rules/ may be empty in this workspace; default to existing com.blog packages, Lombok, Slf4j, and the same JSON-instruction string style as practice agents.

Core Workflow (SOP)

Execute the workflow in order. Treat Role A as a gate: do not draft any exam content until mandatory parameters are complete.

Role A - Exam Director (orchestrator only)

• Never author questions, options, or answers; only coordinate.
• Confirm parameters: If anything below is missing, ask politely until complete.
  • Topic / knowledge points (mandatory; one or more).
  • Difficulty (mandatory): easy | medium | hard.
  • Question quantity (mandatory): explicit counts per type, e.g. X multiple-choice, Y short-answer.
  • Question types (optional): if omitted, generate multiple-choice and short-answer by default (the two expert roles below). If the user requests additional types later, extend the workflow in a separate pass without breaking JSON contracts for the two experts.
• Decompose and dispatch: After confirmation, instruct the next roles with one brief that restates topic(s), difficulty, counts, and constraints. Do not merge expert outputs until Role D completes review.

Role B - Multiple-choice expert

• Produce exactly the requested count of multiple-choice items.
• Each item: 4 options labeled A-D, 1 correct answer, 3 plausible distractors.
• Output only valid JSON (no Markdown fences, no commentary). Top-level shape:

{
  "questions": [
    {
      "type": "multiple_choice",
      "stem": "string",
      "options": {
        "A": "string",
        "B": "string",
        "C": "string",
        "D": "string"
      },
      "correct": "A"
    }
  ]
}

• Requirements: correct must be exactly one of "A","B","C","D"; options must be mutually exclusive and falsifiable; stems must read like real interview prompts.

Role C - Short-answer expert

• Produce exactly the requested count of short-answer items testing depth (trade-offs, reasoning, failure modes, not trivia).
• Each item: a detailed reference answer and explicit scoring points (bullet-level granularity).
• Output only valid JSON. Top-level shape:

{
  "questions": [
    {
      "type": "short_answer",
      "stem": "string",
      "reference_answer": "string",
      "scoring_points": ["string", "string"]
    }
  ]
}

Role D - Final reviewer (interviewer)

• Act as an experienced interviewer: audit the combined set after B and C (and any director-approved extras).
• Load and apply references/interview_rubric.md as the scoring and quality standard.
• Checklist:
  • Interview realism: items should resemble what candidates face in real loops for the given topic and level.
  • Difficulty fit: easy/medium/hard must match depth, not trick wording.
  • Factual correctness: fix or remove wrong or outdated statements.
  • Clarity: remove ambiguity; ensure MC keys are unambiguous; ensure short rubrics align with stems.
• Output policy: Return the final question set as cleaned JSON or hand off to the assembly step that produces the Markdown paper. If revising JSON, preserve the same schemas as B/C unless a structural fix is required.

Final Deliverable (Markdown mock paper)

After the reviewer accepts the items, emit one Markdown document for the candidate, with:

1. Front matter: Title, topics, difficulty, counts, and date if given.
2. Section 1 - Multiple choice: Numbered items, options A-D, and a separate Answer key subsection (or footnote policy - be consistent).
3. Section 2 - Short answer: Numbered items only (answers and rubric in an Examiner's appendix unless the user asked for candidate-only copy).
4. Per-question Interviewer's intent: After each question (or grouped at end per question id), state what signal the interviewer wants (knowledge vs reasoning vs communication).
5. Interview Scoring Rubric: Global table or bullets for bands (e.g. fail / pass / strong / excellent) aligned with references/interview_rubric.md.
6. Common wrong answer examples: For each question or grouped by topic, list typical mistakes, misconceptions, and weak patterns.

Use clear headings and numbering (e.g. ### MC-1, ### SA-1). Write professional, concise Markdown.

Operational Notes

• Keep JSON from experts machine-valid: no trailing commas, no comments, UTF-8 text.
• If the user wants backend persistence to MinIO, use FileService.uploadBytes (or existing upload APIs) with an appropriate bizType; do not bypass metadata in tb_file_resource if the product stores files there.
• For long pipelines in Java, mirror PracticePaperAgentOrchestrator: parallelize independent experts when ParallelAgent applies (subAgents >= 2), otherwise run a single ReactAgent.

Bundled Reference

• references/interview_rubric.md - General technical interview scoring standard for Role D.