Co-STORM for Vault Notes: Researching a New Topic

Where QSTORM retargets STORM to produce exam items, this path retargets Co-STORM to produce vault knowledge notes. Same insight: quality is bottlenecked on pre-writing — a note is only as good as the research that precedes the first sentence. One-pass drafting reliably misses unknown unknowns (facets you didn't know to ask about), and in this vault a thin note compounds downstream: notes feed flashcards, and flashcards + notes feed QSTORM distractors. Better coverage upstream pays off twice.

It is deliberately gated. Co-STORM's multi-agent roundtable is token-heavy, so it runs only for a topic that is net-new and broad — a whole surface, product, or capability with many facets (a Cowork-class launch, a new tool family). Narrow additions and incremental edits stay one-pass; see vault-sync §2.

Stage 0 — The gate

Confirm the topic is a genuine gap (grep the vault; check index.md) and broad enough to have multiple defensible perspectives. If a single definition covers it, stop and draft one-pass. Deep-note is for breadth, not for everything new.

Stage 1 — Warm start

A background research pass runs before any drafting: web-search and fetch the authoritative Anthropic docs for the topic (docs.claude.com / platform.claude.com, anthropic.com/news, anthropic.com/engineering). Collect anchor URLs and a first list of subtopics. Deliberately note retrieved-but-uncited angles — sources that came back but that no draft has used yet. That reservoir is precisely where unknown unknowns live (the Co-STORM "serendipity" idea; see co-storm).

Stage 2 — Roundtable of opposing experts

Spawn a small panel of subagents with deliberately opposing stands on the topic — for a capability note, typically:

Expert	Stand
Mechanics	"This is a concrete, well-specified thing — pin the exact identifiers, params, and loop."
Skeptic / safety	"This is experimental and a real risk surface — surface limits, failure modes, mitigations, and when not to use it."
Practitioner / architect	"When do you actually reach for this vs. alternatives, and how do you deploy it?"

Disagreement is a feature, not noise: opposing experts stress-test claims and surface facets a consensus drafter skips (compare multi-agent debate in agentic-patterns). Each expert researches docs-only and returns typed output — (claim, source-URL) pairs plus a gaps list. The grounding rule is inherited from STORM's TopicExpert: no claim without a source; if retrieval is empty, refuse rather than improvise.

Stage 3 — Moderator / completeness pass

A final agent plays Co-STORM's Moderator: given the experts' combined claims plus the Stage-1 uncited reservoir, it reports what is missing — must-have facts no expert cited, contradictions to resolve, and exact identifiers to pin down (version strings, headers, benchmark numbers). Its output is the punch-list the synthesis must close before the note is done.

Stage 4 — Synthesis (draft-then-refine)

Assemble a draft-then-refine outline (parametric structure first, then reorganized around the grounded findings), then write the note by hand — the orchestrator owns the final write so the parseable house format is exact:

vault frontmatter (title, level, domain, tags, related, updated);
definition-dense - **Term** — definition bullets, because flashcards are extracted from those lines;
a ## Key terms block and ## See also wikilinks;
every non-obvious claim traceable to a source surfaced in a run.

Stage 5 — Register and wire

Add the note to index.md (one - slug — hook line) and, if it anchors a lesson, add its slug to a track in web/scripts/tracks.config.mjs. Then the normal build pipeline carries it to the site. Keep the run trace (who claimed what, from which source) in the session record — it is the note's audit trail.

Key terms

Deep-note path — the gated Co-STORM pipeline vault-sync uses to research a broad, net-new topic before drafting; the note-authoring sibling of the QSTORM exam-item pipeline.
The gate — the rule that Co-STORM runs only when a topic is both a genuine gap and broad enough for multiple perspectives; everything else stays one-pass.
Warm start — the pre-drafting retrieval pass over authoritative docs that seeds the roundtable and stocks the uncited-angle reservoir.
Roundtable of opposing experts — subagents given deliberately conflicting stands so disagreement surfaces facets a single drafter would miss.
Retrieved-but-uncited reservoir — sources found in warm start but not yet used; the raw material for surfacing unknown unknowns.
Moderator / completeness pass — the agent that lists what the experts missed (uncited must-haves, contradictions, exact identifiers) as a punch-list for synthesis.
Draft-then-refine outline — global structure from parametric knowledge first, then reorganized around the grounded findings before writing.
Downstream compounding — the reason coverage matters here: notes feed flashcards, and notes + flashcards feed QSTORM distractors, so a thin note degrades three artifacts.