Workshops/Workshop 5: AI for Studies/Guided build
// STEP 01 OF 06

Set it up

Make a folder for ONE course with a materials folder inside, and put 1–3 lecture PDFs from that course in it. Open Claude Code in the course folder (Claude desktop app → Code tab → choose the folder, or in a terminal: cd my-course, then claude). Paste the prompt below. It asks you 3 questions, then writes the whole agent. When it asks for permission to create files, allow it: that's the agent writing itself.

// COPY INTO YOUR AGENT
01Setup prompt
Set up a study agent for me in this folder. It should run my whole study loop (plan, learn, quiz, review) and remember what I get wrong between sessions. First look at what's in materials/, then ask me 3 short questions, one at a time: (1) course name and exam date, (2) what the exam looks like (calculations, essays, multiple choice...), (3) how tough the quizzing should be. Then create exactly these files: 1. CLAUDE.md, the agent's rules: - materials/ is the only source of truth. Cite every claim with a short (file, page), e.g. (Lecture 4, p. 3). If something isn't in my materials, say so instead of filling the gap. - The shared store is plan.md (concept map), notes/ (one cheat sheet per topic) and progress.md (scores and mistake log). Read it before acting, write to it after. - Never reveal a quiz answer before I attempt it. One question at a time. - Topic short names from plan.md are used everywhere: notes file names, progress.md rows, and commands. If I type a topic in my own words, match it to the closest short name and tell me which one you picked. - Keep everything short. A cheat sheet fits on one screen. 2. .claude/skills/map/SKILL.md: reads everything in materials/ and writes plan.md, a table with one row per exam topic. A topic is what I could study in one 25-minute session: usually 2–5 slides, and 2–4 topics per lecture. Each row has the topic's short name, its key concepts, its sources (file, page), an estimated exam weight (labelled as an estimate) and a status (new / learning / solid). 3. .claude/skills/learn/SKILL.md: takes a topic from plan.md (if I name a single concept, use the topic that contains it and cover the whole topic), reads its pages (and neighbouring pages if it needs an example) and writes notes/<topic>.md with key ideas, definitions and formulas, one worked example from my materials, two common mistakes (flagged in the slides, or a stated rule students easily misapply), and sources. Quality gate before saving: check every line against the source, and delete or mark [CHECK] anything the materials don't directly support. 4. .claude/skills/quiz/SKILL.md: takes a topic (or picks my weakest one from progress.md, or, if nothing's been tested yet, the highest-weight topic that isn't solid). Privately drafts 5 exam-style questions from the notes, at least one aimed at a mistake from my log, each with a source-backed answer key. Asks one at a time, waits for my answer, grades it ✓ / ~ / ✗ with a one-line reason and the source. At the end it appends a dated row to progress.md, logs each mistake, marks old mistakes as fixed when I now get them right, and updates the topic's status in plan.md. 5. .claude/skills/review/SKILL.md: answers "what should I study today?". Reads plan.md, progress.md and the exam date, picks the 3 highest-value topics (weak, never tested, or heavily weighted), proposes a 25-minute session and offers to start the quiz. 6. .claude/skills/study/SKILL.md: the full loop for one topic with no hand-offs from me. Runs learn, then quiz, then shows my updated progress. Every SKILL.md needs frontmatter with a name and a description that says when to use it. When you're done, list the files and tell me to run /map.
Stuck on this step?