agentic-exam-prep

A Claude Code skill for assembling open-book exam-prep packs the way evnchn runs them: an iterative MCQ drill loop (graduate-correct → revamp-wrong → pull-from-source → drain), a print-pack architecture (tag-freeze + round addenda + print-shop-grade two-stage PDF), course-material mining, NotebookLM cram pods, and subagent trust-calibration lessons.

Course-agnostic in mechanics, tuned to HKUST-style open-book finals. Not a general study-advice skill — it encodes one person's working system. Two worked applications of it: COMP4021 (web programming) and COMP4431 (DSP).

Showcase — what it produced (and why it worked)

Run end-to-end on two HKUST courses — COMP4021 (Internet Computing) and COMP4431 (Multimedia Computing) — as a showcase of learning ability ("Agentic Evan" = me + my agents), not of a tool acing a course. Grades aside, the point is the depth of engagement.

The interesting part is deliberately not the source-gathering. Pulling material from a course LMS is a commodity — closed-form, non-agentic solutions exist, and the content is proprietary coursework anyway. That's plumbing, not the point. The value is what happens after the material is in hand:

1. Targeted revision, not re-reading. The drill loop grades each attempt by reasoning quality, not just the letter: a confident-correct answer graduates, a lucky-guess or answer-by-elimination gets retested in a fresh shape, a wrong answer gets revamped against its specific misconception. A parallel failure-shape track (Set 0.x) extrapolates where the same cognitive error will bite next on un-drilled material. Revision lands on the actual gap instead of spraying the whole syllabus.

2. Agent-authored notes — and the pedagogy that makes it a win. The headline isn't "the agent did the work." Courses often assign note-writing precisely because transcribing is the memorization device. Handing that transcription to the agent doesn't skip the learning — it reallocates the time: the hours that would go to copying slides into notes instead go to practicing, understanding, and using the concepts — engagement that lives above rote memorization. The agent freed the human to do the deeper work, and that reallocation — from transcription to understanding — is the thesis this skill is really about.

(Method-level only — no real course content lives here; the proprietary material stays off the repo.)

The two things that matter most

1. Triage by reasoning quality, not just correctness. A right answer reached by lucky guess gets retested, not graduated — the Thinking: line is the signal. Each drill miss is also a cheatsheet-completeness signal: patch the sheet where the miss landed.
2. Canvas is strictly read-only. canvas.ust.hk quizzes count for grade; GET only, never POST/PUT/PATCH. The practice-quiz harvest technique mutates state and is operator-approval-only, never agent-initiated.

Install (Claude Code)

git clone https://github.com/evnchn-agentic/agentic-exam-prep.git \
  ~/.claude/skills/agentic-exam-prep

Notes

The mechanics are course-agnostic; the depth axes and source material are re-designed per course (see the two worked applications above). The Canvas read-only rule is non-negotiable — quizzes count for grade.