Written after running the Resurrection dive (the first flagship novel-dive). The handoff asked: build a new novel-dive skill, or a --novel mode of corpus-dive? Recommendation: a --novel MODE of corpus-dive, not a standalone skill — it shares all the machinery (extraction, verify_quotes, the visuals sweep, render, the verifier, the handoff) and differs only in spine weighting and two extra sections. One maintained pipeline; the hard gates stay identical.
Update 2026-06-09 — BUILT. Both gates cleared (Gate 1 schema = resolved; Gate 2 = mode, confirmed with the reader). The
--novelmode is implemented in.claude/skills/corpus-dive/— a lean overlayreferences/novel-mode.mdplus hooks inSKILL.md(argument + triggers, a Novel mode section, a Phase-2 breadcrumb, and a stale wiki-type-count fix). Verified by a dry-run against The Cossacks (an early/uncensored work with noworks/record) that confirmed the spine is followable and surfaced three late-work-calibration gaps — all since hardened: an absentworks/record now routes to a record-creating proposal; a fourth early-fiction reception shape (“acclaimed legal publication, no ban”); and marquee framing that admits interpretive/developmental claims, not only causal ones. This proposal is now the historical spec;references/novel-mode.mdis the live source of record.
This was a written proposal, not a built skill. Building it (via skill-creator) waited for a go, because it touched a schema question (below) that needed a decision first, per the project’s explain-schema-before-editing rule.
What the work-dive spine got wrong for a novel — and what the Resurrection run actually did
The work-dive spine was tuned on the non-fiction treatises, where the textual layer is exhaustive (~12–14 keystone passages, chapter by chapter). For a 500-page novel that is both wrong and impossible. The Resurrection run re-balanced it by hand; the moves that worked:
| Aspect | Work-dive (non-fiction) | What worked for the novel |
|---|---|---|
| Textual layer | Exhaustive chapter-by-chapter keystone read | Light. Extract the full novel ONCE; grep to locate pivotal scenes; close-read only those (here: communion centrepiece + 3 scenes). A thematic map carries the rest. |
| Centrepiece | One keystone among many | One close-read anchor (the communion scene), treated like the title-saga centrepiece in the Law-of-Violence dive — the scene to quote in full. |
| Genesis & composition | Standard section | Heaviest section. The composition-years diaries+letters sweep + the «История писания» commentary mine got the most agent budget. This is where a novel’s depth lives. |
| Redactions | When present | Rich. The six-redaction history + the censorship-variant text were first-class (here: the centrepiece’s late emergence in the 3rd redaction; the wholesale «Нива» cut). |
| Reception | Russian/Church first, incidental | Heaviest reception. Censorship-as-reception + the excommunication causal question carried a dedicated pass. |
| Characters / prototypes | n/a | New standing section. The real-life prototypes (Koni’s Rozalia Oni; Toporov = Pobedonostsev; Nekhlyudov’s autobiographical core) — a routing source for person entities. |
| Themes | Folded into “what it says” | Its own section (guilt, repentance, resurrection/awakening, justice, land, Church). |
Concrete changes a --novel flag would make
- Phase 1 (Sweep): shift budget to the composition-years witness sweep (diaries + letters for genesis & the people around the work). The novel’s title rarely appears in the diaries — sweep for work-mood and the named network, not a title grep.
- Phase 2 (Extract): replace “read the work’s own text deeply, chapter by chapter” with “extract the full text once; locate and close-read 3–5 pivotal scenes; map the rest thematically.” Add the censorship-variant text as a first-class extract.
- Standing sections: add “Characters & prototypes” and promote “Themes” to its own section. Keep “Genesis & composition” and “Reception & afterlife” as the two heaviest.
- Marquee-question framing: when the dive carries a contested causal claim (here, communion→excommunication), treat it as a hypothesis to test and report
confirms/complicates/extends— this surfaced a real finding (complicates, with the Pobedonostsev-caricature axis) instead of confirming a prior. - Visuals: default to a heavy four-channel fan-out for canonical novels with a rich illustration record (Pasternak here → 85 distinct items). The work-dive default is light.
- Hard gates unchanged:
--choice=reg,verify_quotes.pyexit 0, the bare voice, no vault writes, the record-creatingworkRecord, the separate-pass verifier.
A schema decision the novel-dive forces (needs a go before building)
Update 2026-06-09 — RESOLVED (gate cleared). Decided option (a) with the reader: added a
characterwiki type (minted by a tiering rule — principal, documented prototype, or recurs across works) carrying a structuredprototypes[]field (person+name+basis/certaintyevidence enums — the fiction→life edge), and agrouptype for real-world peoples/sects (Doukhobors, Shakers). The reverse direction is left to Obsidian backlinks (no field added toperson). Seewebsite/schema/wiki-schema.mdv1.4; the Resurrection and Kreutzer dossiers were re-routed off theconceptstopgap. The--novelskill should emit character rows withwikiType: character(+prototypes[]) and sects withwikiType: group. Gate (2) — standalone-skill-vs-mode — is still open and belongs to the build session.
Fictional characters (Maslova, Nekhlyudov) have no home in the 10-type wiki schema — the Resurrection dossier routed them as concept, which the verifier flagged as a stopgap that will mis-shape their eventual pages. Before the --novel mode is built (and before these entities are ingested), decide one of:
- (a) add a character wiki type (and possibly a group type — the Doukhobors are a people/sect, also routed as concept); or
- (b) fold a novel’s principal characters into the work’s own wiki page rather than giving them standalone pages.
This is the one genuinely new design question the novel format raises. Everything else is re-weighting.
Recommendation
Build corpus-dive --novel via skill-creator after: (1) a decision on the character/group schema question; (2) a go on the standalone-vs-mode call (recommend mode). The Resurrection dive is the reference implementation — the spine above is what actually worked on a real 500-page novel, not a guess.