Handoff Brief — Exam Study Tool (personal build)

# Handoff Brief — Exam Study Tool (personal build)

For the building AI: This is a complete, self-contained spec. Build from it cold. You do NOT need any prior conversation history. Read the Safety rails (§5) carefully — they are non-negotiable and define what this tool must NOT do.

---

§1 Who it's for + the goal

The end user is a board-certified nuclear-medicine physician, sub-specialized in neurology, preparing for a nuclear-medicine neuro super-specialization exam. She is ADHD and already a heavy, comfortable Gemini user (assume AI-fluent; design for her, not for a beginner).

Goal: a personal exam-practice / study tool that drills her on board-style questions generated from her own study materials, and — crucially — points her back to the exact source (document + page / slide) so she knows where to study deeper when she gets something wrong.

This is a study tool, not a clinical tool (see §5).

§2 Form factor — decided, don't re-litigate

A web app: HTML + JavaScript. Build it as a PWA (installable, offline-capable).
v0 = a single self-contained HTML file with the question bank embedded as JSON.

No server, no hosting, no accounts — she opens it on her laptop or phone, fully private, works offline. Grow into a hosted/synced PWA only if she asks.

Responsive: must work well on phone (she'll drill in 5-minute gaps — micro-study

is what an ADHD brain sustains best) and on laptop.

NOT native mobile, NOT desktop-only. One codebase, every screen.

§3 The core design

1. Validated question bank from HER materials. An AI ingests her PDFs / slide decks once and generates a bank of board-style Q&A grounded in those documents. The bank is a reviewable artifact (e.g. a JSON file) — NOT freshly generated per session. This is what makes it trustworthy for a high-stakes exam: the content can be spot-checked, and it doesn't hallucinate live. Regenerate/expand when her materials change.

Recommended generator: Gemini — she already uses it, and its long-context

document handling is ideal for ingesting big PDFs/decks. (The builder AI can be anything; the bank-generator is best as Gemini.) 2. Source references on every question (the headline feature). Each question is tagged with its origin: {document, page} or {deck, slide}. When she answers (especially wrong), show "From: Guideline X, p.42 — review this" and, if feasible, link/open that page or slide. This is the "study more when she needs to" guidance — it removes the what do I review next? decision entirely. 3. Weak-area tracking + spaced repetition. Remember what she misses, re-surface it, and roll the misses up by topic/source → "you keep missing dopaminergic imaging → here are the 3 slides on it." Spaced repetition on missed items (active recall is the engine ADHD brains respond to). 4. ADHD-shaped UX: short sets (10–20 questions), instant feedback, visible progress, one-tap start (low activation energy beats a 600-page textbook), clean and uncluttered.

§4 Data shape (suggested)

A question-bank JSON, e.g.: `json { "questions": [ { "id": "q001", "topic": "dopaminergic imaging", "stem": "A DaTscan shows ... Most likely diagnosis?", "choices": ["A ...", "B ...", "C ...", "D ..."], "answer": "B", "explanation": "Because ... (grounded in the source)", "source": {"document": "Neuro NM Guidelines 2025", "page": 42} } ] } ` The runtime app is static — it just serves and scores this bank. (Optional later: a "explain this deeper" button that calls Gemini live — but the core loop needs no live AI.)

§5 SAFETY RAILS — non-negotiable, they travel with this spec

1. This is a STUDY tool, not a diagnostic tool. It must NEVER read real patient scans, never provide a clinical diagnosis, never be used in patient care. Practice / teaching cases only. 2. Ground everything in HER validated materials. Cite, don't invent. Every question and explanation must trace to a source in her documents. Do not generate medical facts from the model's own training — only from her materials. 3. The model can misread a doc it's looking at — so the generated bank must be spot-checkable (sources cited per question) and she remains the authority. Build for "cite-and-verify," not "trust-and-pray." 4. Do NOT have the AI "read" nuclear-medicine images to determine answers. General vision models hallucinate on nuclear-med imaging. For any image-based question, ground the answer in her annotated teaching cases where she/the source provides the ground truth — the AI scaffolds the question around a known answer; it does not diagnose the image.

§6 What you need from her before building

Her source materials (PDFs, slide decks, her notes) — the corpus.
Her exam format (single-best-answer MCQ? multi-select? image-based?), preferred

session length, and the topic list for the exam.

Then: generate the validated bank (Gemini) → build the single-file PWA (§2) around it.

§7 Build philosophy

Compose, don't over-build. Start with the single self-contained HTML file + an embedded validated bank. Ship that, let her use it, add sync / spaced-rep polish / live "explain deeper" only if she wants them. Thin tool, real value, safe by construction.

--- *Brief by lab-ovh (the architect cell), 2026-06-07, for hand-off to a fresh builder AI — keeps the personal build out of lab's substrate context while carrying the full spec + guardrails.*