The map is the moat. Discovery follows the frame.
nlm CLI) · ZAI GLM-5.2 (architect) · Drive + audio + video artifacts · graduates to a global skillA functional walkthrough of the map-then-discover flow and the universal artifact generation payoff. (Generated by ClearSpring Studio.)
A research agent without a frame produces a bibliography. With a map, a literature.
The workflow is two steps with a deliberate model split. Step one is pure reasoning — an LLM acts as a Research Architect and returns a MECE knowledge map (Core Taxonomy, Interdependencies, Blind Spots, a Source-Acquisition Checklist). That runs on cheap ZAI/GLM, not the Gemini API, because it's inference, not retrieval. Step two is retrieval — NotebookLM discovers ~10 sources against the map's checklist, auto-imports them, and synthesizes with every claim cited.
The insight: the map is what makes retrieval auditable. You can see exactly which document type each branch of the taxonomy required, and the imported library reflects that structure instead of being a generic blog roundup.
// the two steps, by model and job# step 1 — reasoning (cheap, ZAI/GLM)
research_architect(topic) → {taxonomy, interdeps, blind_spots, source_checklist}
# step 2 — retrieval (NotebookLM, map-driven)
notebooklm_curate(map) → ~10 sources auto-imported + cited synthesis
# the map becomes a source too — the notebook is grounded on frame + literature
# then: artifacts — Drive docs, audio briefings, video explainers
// structural decisions worth knowing
- map before searchThe whole project exists because of one failure mode: asking NotebookLM to "research X" yields a generic source dump with no frame. Building the knowledge map first turns retrieval from a keyword exercise into a coverage exercise. You discover what you're missing because the map tells you which document types each branch needs.
- universal artifact generationThe system doesn't stop at a synthesis. The same notebook becomes a Drive document, an audio briefing (NotebookLM's two-voice podcast), or a video explainer — one research run, multiple output modalities. The descriptor "universal artifact generation" is literal: the research graduates into whatever artifact the consumer needs.
- driving a service with no APINotebookLM has no public API. The system drives it through the
nlmCLI, which talks to NotebookLM via an authenticated Google session cookie — and seeds/validates that auth check-first, so a freshnlm loginis never clobbered. Working around a closed interface without pretending it's open. - forge-proof verificationA deterministic verifier runs the demo and prints a
DEMO_VERIFY: PASStoken only if it actually executed end-to-end. No mocks. If the NotebookLM cookie is expired, the demo says so and prints the prepared steps rather than inventing a library. The honest-failure path is built in, not bolted on. - skill as the durable artifactThe output isn't a one-off report; it's a graduated skill that runs the same workflow on any topic. The research technique becomes a reusable capability in every session, not a script you re-run. This is the same "skill is the universal durable artifact" pattern the Demo Factory is built around.
- OK... go ahead and ask...How much better informed could I get? Is this advantage legal? Do people know this is possible?