Mit Exam Generator

MIT PhD Qualifying Exam Generator

Generate rigorous academic assessments from structured Markdown content.

What This Skill Does

• Generates 200-question PhD qualifying exams from Markdown/Obsidian notes
• Scales question count proportionally for sparse content
• Auto-detects difficulty from content complexity
• Merges multiple source documents with weighted question distribution
• Validates answer distribution, difficulty spread, and source coverage

What This Skill Does NOT Do

• Process PDFs, images, or non-Markdown formats
• Generate answer explanations with direct quotes (exam integrity)
• Create exams from external web content
• Provide answer keys without completing the full exam

Required Clarifications

Before generating, clarify with user:

Optional Clarifications

Ask only if relevant:

• Custom question count override?
• Specific sections to emphasize or exclude?
• Target audience adjustment (undergrad vs PhD)?

If User Doesn't Respond

Use defaults and note assumptions in exam header:

**Assumptions:** Merged thematically, balanced difficulty, standard timing

Before Implementation

Exam Specifications

Grading Scale

Generation Workflow

1. ANALYZE
   └── Read source files → Extract concepts → Map sections
   └── Calculate: content_density = concepts / sections

2. CALIBRATE
   └── question_count = min(200, concepts * 2)
   └── difficulty_profile = analyze_complexity(content)

3. DISTRIBUTE
   └── Allocate questions by type (see references/question-patterns.md)
   └── Allocate by Bloom's level (see references/bloom-taxonomy.md)
   └── Weight by source document size (multi-doc)

4. GENERATE
   └── Create questions following type patterns
   └── Ensure distractors are plausible (70-90% correct)
   └── Track source section for each question

5. VALIDATE
   └── Run all checks (see references/validation-rules.md)
   └── Fix any failures before delivery

6. OUTPUT
   └── Save to exam-[source-name].md alongside source

Question Type Distribution

See references/question-patterns.md for templates and examples.

Bloom's Taxonomy Distribution

See references/bloom-taxonomy.md for level indicators.

Answer Construction Rules

1. Option A: Never "All/None of the above"
2. Correct Answer: One clearly correct option
3. Distractors: Plausible but fail on critical detail (70-90% correct)
4. Distribution: Roughly equal A:B:C:D across exam
5. Sequences: No more than 3 consecutive same-letter answers

Multi-Document Handling

When multiple source files provided:

weight[doc] = word_count[doc] / total_word_count
questions[doc] = round(total_questions * weight[doc])

Create distinct sections per source or merge thematically (user preference).

Output Format

# [Exam Title]
## MIT PhD Qualifying Examination

**Source:** [file(s)]
**Questions:** [N]
**Duration:** [X] minutes
**Generated:** [date]

---

### PART A: [Topic] ([X] Questions)

**Q1.** [Question stem]
A) [Option]
B) [Option]
C) [Option]
D) [Option]

[Continue all questions...]

---

## ANSWER KEY

| Q# | Ans | Section | Difficulty | Bloom |
|----|-----|---------|------------|-------|
| 1 | C | Part A | Medium | Apply |

---

## EXPLANATIONS

### Q1
**Correct: C**
[Explanation with section reference - NO direct quotes]
Section: [Heading from source]

Scaling Algorithm

def calculate_questions(content):
    concepts = extract_testable_concepts(content)

    if len(concepts) >= 100:
        return 200  # Full exam
    elif len(concepts) >= 50:
        return 100  # Half exam
    elif len(concepts) >= 25:
        return 50   # Quarter exam
    else:
        return max(25, len(concepts))  # Minimum viable

Edge Case Handling

Validation Pipeline

Run ALL checks before delivery. See references/validation-rules.md.

Quick Checklist

• Question count matches calculated target
• Each question has exactly 4 options (A-D)
• Answer distribution within 20-30% per letter
• No >3 consecutive same-letter answers
• All Bloom levels represented per distribution
• All question types represented per distribution
• Every question has section reference
• No direct quotes in explanations
• Difficulty distribution matches content complexity

Reference Files