nblm-flashcards - SKILL.md Agent Skill

name: nblm-flashcards description: "Generate and drill flashcards from NotebookLM Enterprise notebooks via Gemini (Vertex AI), with Emacs UI and babashka backfill pipeline" version: 1.1.0 trit: 1 trit_label: PLUS

NotebookLM Enterprise Flashcard Generator + Drill

Trit: +1 (PLUS — knowledge extraction and spaced repetition from notebook sources)

What This Does

Extracts knowledge from the entire plurigrid/bmorphism ecosystem (1,244 repos across 5 NBLM notebooks) into drillable flashcards.

GitHub repos (1,244 total)
    |
    v
DeepWiki URLs → NBLM Enterprise (5 notebooks × 300 sources)
    |
    v
Discovery Engine API → fetch source titles
    |
    v
Gemini 2.5 Flash (Vertex AI) → batched generation
    |  5 sources/batch, 10 cards/source, token auto-refresh
    v
flashcards.json + quiz.json (per notebook)
    |
    v
bb n/nblm_merge.bb → all-flashcards.json (deduped)
    |
    v
Emacs drill (M-x nblm-drill-all)

Interaction with notebooklm-enterprise skill

This skill is the consumer of notebooks managed by notebooklm-enterprise:

notebooklm-enterprise (ERGODIC)	nblm-flashcards (PLUS)
Creates notebooks	Reads notebook sources
Adds DeepWiki URLs as sources	Generates flashcards from sources
Manages source ingestion status	Batches Gemini calls per source
Shares notebooks with users	Merges decks across notebooks

GF(3) Triad

notebooklm-enterprise (0) + nblm-flashcards (+1) + [verifier] (-1) = 0

The verifier role is filled by the Emacs drill itself — missed cards feed back as verification signal.

Notebooks (as of 2026-03-25)

Notebook	ID	Sources	Role
Original	`9ca780dc-4e0f-4f57-9262-a6090af028e4`	300	bmorphism + plurigrid core
Backfill 1/4	`ce99b119-74f6-4e24-bd14-333aade95950`	300	Gap repos chunk 1
Backfill 2/4	(see backfill-state.json)	300	Gap repos chunk 2
Backfill 3/4	`aa69ef7a-3fb2-4ba0-863e-d74ef3b670eb`	300	Gap repos chunk 3
Backfill 4/4	`57c2d969-ebc9-4402-b8b1-c2c2d83abd40`	44	Gap repos chunk 4

Pipeline

All scripts live in ~/worlds/n/:

# Generate flashcards from a single notebook
hy n/nblm_flashcards.hy --notebook-id <UUID> --per-source 10 --batch-size 5

# Backfill: create notebooks for uncovered repos
bb n/nblm_backfill.bb              # create + add sources
bb n/nblm_backfill.bb --status     # check ingestion
bb n/nblm_backfill.bb --generate   # generate flashcards per notebook

# Merge all decks
bb n/nblm_merge.bb

# Drill in Emacs
emacs -nw --load ~/worlds/n/nblm-drill.el
# M-x nblm-drill-all    (full merged deck)
# M-x nblm-drill-repo   (filter by org/repo)

Key Parameters

Arg	Default	What
`--per-source`	10	Flashcards per notebook source
`--batch-size`	5	Sources per Gemini call (50 cards/call sweet spot)
`--difficulty`	hard	easy / medium / hard
`--model`	gemini-2.5-flash	Vertex AI model

Architecture

Token auto-refreshes every 30 min (gcloud tokens expire at 1hr)
Truncated JSON repaired by finding last complete } and closing array
Discovery Engine v1alpha for notebook source listing
Gemini via Vertex AI REST (generateContent), region us-central1
maxOutputTokens: 65536, temperature: 0.7
GCP project: merovingians (302712368086)

Ecosystem Coverage

Org	Repos	Coverage
`bmorphism/`	406	100% (across original + backfill)
`plurigrid/`	586	100% (across original + backfill)
Third-party	131	100% (backfill chunks 1-4)
Total	1,244	100%

Related Skills

notebooklm-enterprise — NBLM API (notebook CRUD, source management)
deepwiki-mcp — DeepWiki documentation for source URLs
babashka — Pipeline orchestration (backfill, merge)
gh-cli — Repo discovery for gap analysis
gh-interactome — Author/repo network for cross-repo flashcards