Course Study

Course Study v3.0

A lean four-phase workflow that turns a course into complete-coverage, Feynman-explained, exam-ready revision notes. The #1 guarantee is completeness (every topic covered, nothing silently dropped); the defining pedagogy is the Feynman concept block (plain-language capsule first + mandatory worked example).

Primary output: revision-notes.md. Optional: a one-line-per-entry quick-reference.md cheat sheet, and PDF export (CJK/bilingual aware).

This file is a thin orchestrator — load the rules/ module for each phase.

Pipeline

Phase 0 Intake (single exchange)
  ├── PDF slides → Phase 1 (extract via /pdf skill)
  ├── Topic list → Phase 1 (parse into the checklist)
  └── Course name → search standard syllabus → Phase 1
Phase 1 Cover    → extract ALL content (page-aligned) + EMIT the coverage checklist (the ledger)
Phase 2 Distill  → revision-notes.md in Feynman block order; RECONCILE against the Phase-1 checklist
Phase 3 Supplement (OPTIONAL, light) → ≤~10 sourced targets for genuine gaps / thin concepts
VERIFY/REPORT    → coverage reconciled, examples present, sources traced; emit files

Each phase ends with a one-line checkpoint; proceed on no-objection. Never spread intake across multiple messages.

• Phase 0 — Intake: load rules/phase-intake.md. One exchange: input type, rough page count → scale tier, output language, exam date, priority topics, output folder; detect web access silently.
• Phase 1 — Cover: load rules/phase-cover.md. Extract every concept page-aligned (PDF via the /pdf skill; or parse the topic list; or build the outline from a searched standard syllabus). Emit the coverage checklist enumerating every topic — the completeness ledger Phase 2 reconciles against.
• Phase 2 — Distill (main deliverable): load rules/phase-distill.md. Write revision-notes.md in backbone order, each concept in the Feynman block order below, with cross-topic bridges. Then reconcile the notes against the Phase-1 checklist — flag and fill any missing topic before finalizing.
• Phase 3 — Supplement (optional, light): load rules/phase-supplement.md only for genuine gaps / thin concepts. Cap ≤~10 targets. Dual web / no-web.

The Feynman concept block (Phase 2 — mandatory order)

Every concept is written in this exact order. The plain-language capsule comes first — never lead with the formal definition.

1. Plain-language capsule — explain it simply, as if to a smart friend with no jargon. One short paragraph. (This is FIRST, always.)
2. Intuition — why it exists, what problem it solves, an analogy if it helps.
3. Formal treatment — the precise version: LaTeX formula or code, symbols defined.
4. Worked example — a concrete, step-by-step example (numbers plugged in / algorithm traced). Mandatory for every non-trivial concept.
5. Connections + common misconception — prerequisites and what it enables; a cross-topic bridge where useful; one thing students typically get wrong.

Full template and depth calibration: rules/phase-distill.md and rules/templates.md.

Global Rules (controls)

1. PDF-only input via /pdf. ALL PDF reading — including scanned / image-only PDFs — goes through the /pdf skill. NEVER raw file I/O or Python on PDFs. Non-PDF inputs (PPTX/DOCX/images) are converted via /pdf first.
2. Completeness invariant. Phase 2 notes are reconciled against the Phase-1 coverage checklist before finalizing. Any extracted/checklist topic missing from the notes is flagged and filled — never silently dropped or skipped.
3. Worked-example invariant. Every non-trivial concept gets a concrete worked example. A pure-definition concept with no feasible example gets the plain-language capsule + a short note (e.g. "definitional — no worked example applies") — never a fabricated/forced example.
4. No fabrication. Offline supplements are marked [Standard curriculum knowledge]; ZERO invented URLs / papers / authors / slide content. Uncertain claims are omitted or flagged [Uncertain — verify before exam]. On the course-name path, never fabricate a specific lecture's slide content.
5. Source traceability. Every note traces to its source location: page (PDF, Lecture X, p. Y) or section (topic list / syllabus). Never lose it.
6. Honor the source. A slide that contradicts standard curriculum is flagged as a discrepancy (show the slide's claim + the standard view) — NOT silently "corrected" to the textbook version.
7. Scale guard. Size picks the tier — page count (PDF) or topic count (topic list / syllabus): large inputs (>400 pages, or >150 topics) → split, recommend per-module runs and batch; the checklist spans the whole course so nothing is silently dropped across batches.
8. Scope guard. Produce revision material, not answers to graded assessments. Do NOT solve / do the user's actual homework or exam questions for submission; offer how to approach them as a study topic instead.
9. Output discipline. Dense notes, no padding. quick-reference.md (if produced) is one line per entry, ordered by exam relevance — no prose. Completeness of coverage is non-negotiable — never drop a topic to be brief. If the user wants brevity, satisfy it via depth calibration (more topics at minimal capsule depth) and/or the quick-reference.md cheat sheet, never by omitting topics.
10. Track progress. Use a TodoList for which lectures/topics are processed.
11. Prioritize flagged topics. Priority topics named in Phase 0 get deeper treatment and appear first in quick-reference.md.

Reference Files

Anti-Patterns