Reference books: fetch the book, read the answer

There is a library of technical books in the public Google Drive folder 1j7l7CVK46aIYnuaXL5iRS6FU53TCwSTR (the "Aris bookshelf"). When the user asks something a reference settles — a definition, a formula, a recipe, a design pattern, a best practice — pull the right book from the shelf and answer from its text, not from memory. The shelf is live: names and file ids are parsed from the folder page at call time, so a new file appears without a code change.

When to use this

Use this when the question is slow-moving knowledge a book would contain:

• The user names a book or a topic on the shelf ("по книге по pandas", "SQL", "временные ряды", "квантовые вычисления", "графика на Python").
• The answer is a settled recipe: how SQL joins work, how to fit a pandas DataFrame, how to structure a Python project, how an algorithm is defined.
• The user asks about an OOP-by-example book (Lott), a beginner tutorial (Васильев), a handbook (Шпаргалка тестировщика), or a "unlock the secrets" performance/security/causal-inference title.

Do not use this for:

• Questions about a codebase the user is working in — see api-doc-recall.
• A synthesis across several sources — see research-synthesis.
• Anything where the up-to-date truth lives in docs, not a book.

Rules

Ordered by how often they go wrong.

1. Match the ask from the catalog, not from guessing

The catalog in references/catalog.md is the semantic index of the shelf: every book with its topic, language, what it covers, and a "use when". Read it first — do not scan the drive or guess a title from memory. Find the row whose topic and "use when" fit the ask, then download exactly that file.

The topic map at a glance:

If the topic is not in the catalog, say so plainly instead of inventing a match; the catalog also lists what the shelf conspicuously lacks.

2. Download with the script, one file at a time

• node scripts/fetch-book.mjs --list — see every title (use before a fuzzy name).
• node scripts/fetch-book.mjs --list-ids — see name :: fileID pairs, for when a title matches several copies.
• node scripts/fetch-book.mjs "<name>" --out <dir> — download. Exit codes: 0 matched and saved, 1 no match, 2 several matches (list them and re-run with a name unique enough).

The script checks the %PDF- magic bytes after download, so a "virus scan" HTML page is never silently kept as a PDF. Save into a downloads folder and remember the path; do not re-download a file already on disk.

3. Read the relevant part, not the volume

Books here are 4–50 MB. Do not read the whole PDF for one question. Use the pdf skill to extract text, then locate the section that answers the ask: chapter titles, table of contents, an index. Quote the page or section when you answer, so the user can check.

4. Answer from the book, not from memory

The point of the fetch is that the book is the source. When the book is internalized (e.g., the user's own handbook), an answer can use it directly; when it is a third-party reference, answer with the book's wording and give the page. Do not let a download replace a question — if the ask stays unclear after the book is read, say what the book covers and re-ask.

5. Keep the catalog current

The catalog is the lookup table the agent reads first, so it must reflect the live shelf. Run node scripts/fetch-book.mjs --verify-catalog; it prints every live file missing from references/catalog.md and exits 1 if any. For each, add a row (topic, language, one-line covers, use when) and update the SKILL.md topic map. Treat (N)-suffixed names as duplicates of their base book, not as new books. Commit the update so the next session does not have to re-derive it.

If the topic is missing after that, do not guess a book exists. Name the gap: "this shelf does not carry <topic>" and offer the closest neighbor.

Reference

• references/catalog.md — the full semantic index of the shelf: every book with topic, language, what it covers, use-when, plus the list of topics the shelf lacks. Read this first to pick a book; it is also where a newly discovered file is registered.
• references/shelf.md — the live listing commands and the exit-code contract of fetch-book.mjs; read when a download returns something other than a %PDF- file.
• scripts/fetch-book.mjs — the downloader. It resolves the folder page's data-id/data-tooltip at runtime and fetches https://drive.google.com/uc?export=download&id=<FILE_ID> for the match. Run directly on any machine with Node 18+.

Checklist

• Scanned the shelf and picked the single most relevant book by topic
• Downloaded with fetch-book.mjs (exit 0) into a known folder; not re-downloading an existing file
• Extracted the answer with the pdf skill from the right section, and the answer cites the book/chapter
• Topic not on the shelf reached the user as an explicit gap, not a guess