Exam Intelligence

Exam Intelligence

Use this skill when a student asks to analyze past exams, identify SOS material, find repeated exercises or solution patterns, compare exam years, decide what to study first, build a target-grade or cram plan, or generate pattern-based practice.

Historical frequency is not a prediction probability. Say, for example, “Taylor appeared in 9 of 11 analyzed exam variants (81.8% historical occurrence),” never “Taylor has an 81.8% chance of appearing.”

Operating boundary

The agent performs semantic work: inspect source files, preserve exact problem text, split every subquestion, understand notation, select topic/pattern/method, estimate qualitative difficulty and confidence, and link relevant solutions. Python performs deterministic work: validate, count, group, rank, cache, detect duplicate candidates, create plans, and generate reports. The core needs no API key and course material stays local unless the user explicitly configures an external service.

Read only the material needed for the current stage. Do not load an entire large course into context. Prefer native or parsed text, inspect pages/images when needed, and use OCR only as a fallback. Never silently repair an unreadable symbol; lower confidence and mark it for review.

Workflow

1. Inventory sources with exam-intelligence add PATH --type exam|notes|solution|book.
2. Inspect .exam-intelligence/sources.json and cached text under .exam-intelligence/cache/text/ progressively.
3. Extract exams and independent subquestions into the contract documented in references/extraction.md. Preserve source file, page and raw text.
4. Classify each question using existing patterns first. Pattern IDs describe a transferable solution strategy, not constants or a particular exam year.
5. Validate low-confidence assignments and alternatives. Set needs_review through confidence below 0.70; never invent marks.
6. Write { "exams": [...], "questions": [...], "patterns": [...] } JSON and run exam-intelligence import-data FILE.
7. Run exam-intelligence doctor, then exam-intelligence analyze and exam-intelligence report.
8. Inspect solutions/notes only for the highest-value patterns. Prefer professor solutions, solved exams and course notes over generic knowledge. Surface multiple solution methods when supplied sources conflict.
9. Use exam-intelligence plan --target GRADE --hours HOURS or cram.
10. Explain uncertainty and trace recommendations back to historical questions.

Manual corrections made with set-pattern must not be overwritten by a later agent import. Add genuinely new course patterns centrally only after checking for semantic equivalence. Generated exercises must be labelled “Generated practice question” and must never masquerade as historical material.

Consult references/methodology.md, references/taxonomy.md, references/scoring.md, references/extraction.md, and references/output-format.md as needed.