Course notes

You are turning a student's notes from one lecture into the study notes anyone in the course would want: the whole lecture in its own order, with the hard parts explained properly. The student's pages are the input, not the content. They tell you what the lecture said, which parts the student found hard, and which misreadings are plausible. The book is written for any reader and never quotes, mentions or addresses the original notes or their author.

Never add from memory. Every fact cites the course's own material first: the lecture's slides, then the course textbook or study guide, then the papers the lecture teaches.

The output is a Letter-size PDF in the house style of the technical-manual kit (cream paper, Manrope / Source Sans / Source Serif italic / JetBrains Mono): cited body text, focus callouts for the hard topics, "Common mistake" callouts, and pointers forward. One lecture is one part; the book grows over the term.

What stays private: the transcript, the gap ledger and the questions about the student's own pages live in research/ and in the conversation, not in the book.

Files in this skill

Prerequisites

Check these before step 1, and ask the user to install anything missing before the work starts.

Without pdftoppm the build still writes the PDF but skips the page images, so the pages cannot be looked at. If the user declines to install it, say in the hand-over that the pages were not checked.

Workflow

0. The course (once per course)

Settle, from the request or by asking (one question at a time):

• Course and term, e.g. "Stanford CME 295, fall 2026". The term decides which slide deck is the right one; courses revise slides every year.
• Which lecture the photos are from, and its date.
• What the student finds hard. Ask if they have not said. These become focus topics: they get worked examples, a focus callout and self-check questions. This is the most valuable thing to know.
• Familiarity, 1 to 5. Ask with the scale written out (references/levels.md): 1 new to the subject and what it builds on, 3 comfortable with the background but new to the topic, 5 a dense review. It decides how much the book teaches around each idea, including whether it opens with a Background part on the prerequisites. Ask too whether any hard topic sits lower than the rest.
• Where the materials are. Find them yourself first: search for the course website and syllabus, then the posted slides, the course's own study guide or cheatsheet, and the paper a lecture is built on. Confirm the site you found with the student ("Is this the course: cme295.stanford.edu, fall 2026?") before relying on it. Ask for a link or an upload only when nothing public turns up, or the materials sit behind a login or a paywall (Canvas, Ed, a textbook). If there are no slides at all, ask what the lecture followed (a textbook chapter, a recording) and cite that.

Record the answers once in research/COURSE.md, headed # <Course> · <School> · <term> (course, term, website, syllabus, lecture list with dates, familiarity level, focus topics, lectures done) so later lectures skip these questions (in a new conversation, use what the conversation or memory already holds about the course before asking again); ask only for the new lecture's number, what was hard in it, and whether the last book's level felt right.

1. Transcribe

Read references/transcription.md, then transcribe every photo into research/transcription/transcript.md, line by line, with page and line numbers. Keep spelling, abbreviations and struck-out words. Describe each diagram in words. Mark every reading you are not sure of, and say whether the doubt changes the meaning. Finish with a coverage table: the lecture's topics by slide range against what the notes cover.

2. Gather the sources

Get the lecture's slides for the student's term and read all of them, recording slide numbers; the slide number is the citation the student can check. Get the course study guide or cheatsheet and the paper(s) the lecture teaches. Record each with its version in research/SOURCES.md. If the shell cannot download a file, read it through the browser or a fetch tool; never fall back to memory for what the slides say.

3. The gap ledger

Write research/GAPS.md by walking the slides against the transcript (references/gaps.md): what is missing, what is thin, what needs a fix, and what you must ask because only the student knows. Tie each row to a source and to the section of the book that will handle it. Missing and thin rows become ordinary sections; fix rows become "Common mistake" callouts; ask rows are asked in the conversation and never printed.

4. The check kit

Skip this step when the lecture has nothing to compute (a history or literature lecture); its worked examples come from cited cases, documents and timelines instead.

For anything computable in the lecture (a formula, a worked example, an algorithm's steps, sizes of a model, a probability), write a small script in research/check/ that computes it and saves JSON to research/captures/. The notes bind to those numbers, so they are computed, not recalled. Toy examples are fine and often clearer than real ones; label them toy, and say how they differ from the real thing.

5. Write

Write at the reader's level (references/levels.md). At levels 1–3, first list the lecture's prerequisites in research/PREREQS.md and write the Background sections the level needs, with their own check script; at levels 1–2 every formula or formal statement gets an "In plain words" reading and every new term is defined where it first appears. Set "level" in manual.json; the build checks the rules for that level.

One file per lecture. Follow the lecture's own order and cover the whole lecture. For each topic:

1. A section with a title, an italic deck that states the point, and a lede that says what the section does.
2. The content: what the slides say, cited by slide; formulas from the slides or paper; a figure. When the student sketched a slide's figure, redraw that slide's figure with real or computed values, and cite the slide.
3. Common mistakes in a .callout.fix labelled "Common mistake": state the plausible wrong reading in general terms (never as a quote from the notes), say what the source says, and cite it. The student's own errors are the best guide to which mistakes are worth a callout.
4. Focus topics (the ones the student found hard) get a .callout.focus with the whole idea in two or three sentences, a worked example from the check kit, and a step-by-step walk-through.
5. Where this goes next: one line pointing to the later lecture that builds on the point.

Topics the student's pages skip get the same treatment as the rest, scaled to the lecture's emphasis; a long stretch of slides can be one short summary section. Never say which parts the notes covered.

6. Reference

The Reference part holds, cumulatively across lectures: self-check questions and answers, formulas and sizes, a glossary, open questions the slides leave unresolved (for office hours), sources and versions, and an index of figures. The front matter maps the lecture's slide ranges to sections.

7. Build, look, ask

Build, fix every problem, look at every page. Then send the student the PDF and, in the conversation, the first open question about their pages from the ledger, one at a time. When they answer, update the transcript and the notes, and rebuild.

Next lecture

Add a part to the same book and a folder research/lecture-N/. Keep the glossary and formula sheet cumulative. When a later lecture revisits a topic, cross-reference the earlier section instead of repeating it.

Voice

Write like a good TA writing study notes for the whole class: direct, specific, warm without padding.

• No second person and no reference to the source notes: no "you wrote", "your notes", "not in the notes", "the handwritten pages". The book reads as if written from the lecture alone.
• Lead with the idea, then the mechanism, then a number. "Three quarters of the attention goes to 'cute'" beats "attention is distributed".
• Quote the slide when its wording is the point.
• When a common reading is wrong, say what the source says; do not soften it into vagueness.
• No hype, no filler, no "simply", no "obviously". A student who struggles with a topic needs the steps, not reassurance.

Grounding rules

1. The transcript records the student's words exactly. A misreading is fixed in the transcript.
2. Every expansion carries a <cite>: slide 62, L2 slide 20, Vaswani §3.2.1, cheatsheet. Every section ends with a Sources line. The build refuses a common-mistake callout without a cite.
3. The slides of the student's term outrank everything else; a paper outranks the web; the web is context only. When the slides and a paper differ, say so.
4. Computed numbers come from the check kit and are bound with data-value. Toy numbers are labelled toy.
5. What could not be checked (the recording, a paywalled book) is listed under "Not checked".

Build

cp -r "<this skill's folder>/assets/kit" kit
(cd kit && npm install)                    # pagedjs, fonts, highlight.js, playwright-core
python3 research/check/<script>.py         # captures first
python3 research/figs/make.py              # then figures from the captures
node kit/build.mjs . --png                 # -> <slug>-Course-Notes.pdf and build/pages/*.png

manual.json sets "kind": "Course Notes" so the cover and footers say so. To check the install, copy references/example beside the kit and build it: 35 pages, no problems. The build needs a Chromium; it tries Playwright's, then PLAYWRIGHT_BROWSERS_PATH, CHROME_PATH and common install paths.

The build reports missing sources, uncited common-mistake callouts, broken bindings and references, and layout problems.

Before you hand it over

• Every line of every photo appears in the transcript.
• Every uncertain reading that changes the meaning is an open question for the student, asked in the conversation.
• The book has no "you", "your", quotes from the notes, or mention of them; grep for it.
• Every focus topic has a worked example, a walk-through and a self-check question.
• The book meets its level: at 1–2, every prerequisite in research/PREREQS.md has a Background section, and no term is used before it is defined.
• Every topic in the coverage table, including those the notes skip, has a section.
• Every number is bound to a capture or cited to a slide or paper.
• You have looked at every page, and you are sending one question, not a list.