Study Notes

Notes exist to be studied: short, easy to scan, and faithful to the source. A wrong or invented point gets memorized, so the source decides every fact.

Inputs

Ask only for what the user did not give, all in one question round.

• Sources — PDF, .md, URL, existing Notion page. Several allowed. Read every source; unreadable → say which and stop.
• Target — Notion, Obsidian, or plain markdown.
  • Notion → page URL or name (search if name only). Notion connector missing → say so, offer markdown instead.
  • Obsidian → vault path, plus folder inside the vault (default: vault root).
  • Plain markdown → .md file or folder path, default current folder.
• Mode — boilerplate only, or full notes.
• Style reference (optional) — the user's earlier notes. Given → mirror their headings, markers, and colors over the defaults below.

Large sources

Token-heavy reads (long PDFs, big pages) → subagents, one per chapter, in parallel. Before the first one, ask once which model to use: Sonnet (recommended — cheaper, enough for digests and fact-checks) or Opus (stronger on dense or ambiguous material). Use that model for every subagent in the run. Ask each for a faithful digest: item names exact, learning objectives verbatim (code, level, exact text), ≤15-word gist per other item, every pro/con and ordered list, no outside knowledge, source oddities flagged. Short chapters → read inline.

PDF → extract text once to a scratch file (pypdf), reuse it for every later read. Pages that extract empty or near-empty, or have images (page.images) → read them with the Read tool (renders pages); still unreadable → name them in the report.

Source is also the target → snapshot it before Step 2 writes; later reads and Step 4 use the snapshot.

Step 1 — Outline

Extract from the source: chapters, sections, subsections, learning objectives (code, level, exact text), keywords, study time, exam structure. Missing info stays missing — never invent question counts, pass marks, or durations; use ?.

Step 2 — Boilerplate

• Introduction: Paper Structure (per chapter: ? Questions unless source gives counts, study time in gray), level legend with only levels the source uses, resource links the user gives.
• Per chapter: # N. Title, gray Keywords: line (source says keywords are examinable when it lists them), ## N.M Section, ### N.M.K Short title <level emoji>, then an Objective: line with the exact LO text and an empty spacer for notes.
• Headings are short topic names (e.g. "Framework Layering"), not the full LO sentence — the exact wording lives on the Objective: line.
• Source has no LOs or levels → chapters and sections only; drop Objective lines, legend, and Paper Structure.

Level emojis: K1 🟧, K2 🟩, K3 🟪, K4 🟥. Other level schemes → pick distinct square emojis and show them in the legend.

Boilerplate only → stop here and report.

Step 3 — Notes (full mode)

One chapter at a time, in order:

1. Digest the chapter (see Large sources).
2. Propose in chat, per subsection: which blocks and their key points. Wait for approval. User said to do all chapters without stopping → skip approval.
3. Add notes under that subsection's Objective: line, else under the section heading.
4. Offer the next chapter.

Blocks, chosen per content:

• Italic definition line → one-sentence meaning of the core term.
• Table → comparisons, pros/cons, "X vs Y", items with 2+ attributes.
• List → enumerations; keep the source's order when it gives one.
• Callout → key takeaway, limitation, rule the source stresses.
• Gray line → secondary detail worth keeping but not memorizing first.

Rules:

• Only what the source says. Keep the source's item names. No outside knowledge, no inferred orderings or chains the source does not state.
• Source oddity (typo, contradiction, "three" then lists four) → keep the source wording, add a short gray note. Exams follow the source wording.
• Color sparingly: green pros, red cons or critical points, bold for key terms.
• Short: 2–3 points per table cell, one line per list item.

Step 4 — Verify (full mode)

After the last chapter, a subagent (model chosen above, else ask) fact-checks this run's notes against the source. Report only: factual errors, unsupported claims, items missing from a list presented as complete. Ask for the verbatim note snippet and source location per finding.

Re-check each finding against the source yourself before fixing — subagent output is a lead, not proof. Fix confirmed ones, skip the rest with a reason. Finding in the user's existing content → report, don't fix.

Output

• Notion → follow references/notion.md.
• Obsidian or plain markdown → follow references/markdown.md. One file per course.
• Existing note → keep the user's content and edits; insert, never replace the whole note. File exists and is not the target note → ask before overwriting.

Done

Report the page link or file path, chapters covered, and what is left as ?.