| name | loredeck-builder |
| description | Build complete, validated Saga Loredecks outside the app through a staged, user-gated workflow (scope brief, evidence, context planning, titles, cards, health, package). Use when the user wants to create a new canon or fandom Loredeck, resume a workshop project, validate a deck folder, or produce an importable .saga-loredeck.zip. Scales from a single novel to a huge franchise with subagent fan-out. |
Loredeck Builder
You are driving Saga's external Loredeck authoring workflow: a staged loop that turns a canon (one novel up to a whole franchise) into a validated, importable .saga-loredeck.zip. The user reviews and approves every stage; nothing advances without their explicit sign-off.
Ground rules
- The schema reference is the only source of truth for data shapes. Read
docs/loredecks/SAGA_LOREDECK_SCHEMA.md before emitting deck JSON. Never invent fields, registry shapes, or health codes. references/authoring-rules.md condenses the practical rules.
- Evidence before cards. Lorecards may only be drafted from accepted evidence records, and every card cites its evidence in
sourceInfo.evidenceRefs. No wiki-memory drafting.
- Gates are user approvals, not formalities. Present the stage's review artifact, wait for the user's explicit approval in chat, and only then run
gate approve. If the user rejects, revise and re-present — that is the loop.
- Project state is CLI-owned. Never hand-edit
project.json; use the CLI so the resume contract stays valid. All other project files (briefs, evidence, plans, deck drafts) are yours to write.
- The release bar is strict-clean Pack Health: zero errors, warnings, AND suggestions.
promote and verify-package enforce this; do not argue a warning is acceptable — fix it.
- Track multi-part work with a task list tool when your runtime has one. Evidence scopes, title/card batches, and (for families) decks in flight are exactly the state a task-tracking tool exists for. Use it alongside
batch set/status, not instead of them: the CLI owns the resume contract; the task list is your own live picture of what's done, in progress, and blocked this session — most valuable with subagent fan-out, where it's easy to lose track of which of N spawned subagents have returned.
The CLI
From a repository checkout, commands are node tools/loredeck/loredeck-cli.mjs <command> and projects default to the gitignored workshop/ directory. The packaged skill uses its generated cli/loredeck-plugin.mjs wrapper and defaults to <your project>/loredeck-workshop/. Both entry points honor SAGA_WORKSHOP_ROOT; set it explicitly when the location matters. Add --json for machine-readable output.
| Command | Purpose |
|---|
init <id> --title T [--size single|family] [--decks id:role,...] | Scaffold project + skeleton deck folders |
status <id> --json | Resume contract: stage, pending gate, counts |
deck add <id> --deck D:ROLE | Add a core/era/standalone deck to an existing project |
gate approve <id> [--deck D] [--note N] [--artifact P] | Record user approval, advance project-wide or deck-scoped stage |
gate reopen <id> --stage S [--note N] | Rewind to an earlier (or same) stage — for a family project already at complete gaining new decks; never advances |
evidence validate|accept|reject <id> [--scope S] [--ids a,b|--all] | Evidence pipeline |
batch set <id> --deck D --kind titles|cards --id B --status S [--count N] | Record batch review outcomes |
report <id> --stage brief|evidence|plan|titles|cards|final | Regenerate the stage review artifact in reviews/ |
health <deck-dir|id> [--deck D] [--dist] [--strict] | Full Pack Health (identical to in-app) |
conformance <deck-dir> / stats <deck-dir> --write | Structural checks / stats+files[] rewrite |
promote <id> [--deck D] | drafts → dist, gated on conformance + strict health |
package <id> [--deck D] [--author A] [--pkg-version V] | Build the .saga-loredeck.zip from dist/ |
verify-package <zip> | Parse + health-check the final artifact |
The CLI mechanically validates project state, evidence shape, manifests, registries, statistics, Pack Health, safe package paths, and archive round-trips. It does not prove that a card's wording is actually grounded in its cited facts, that a continuity decision is correct, or that a deck deserves human-vetted quality tags. Regenerate each review artifact, resolve its findings, present it to the user, and only then record the gate approval.
Session start: new or resume?
If the user names an existing project (or you find one), run status <id> --json and resume at the recorded stage — regenerate that stage's review artifact before re-presenting its gate. See references/state-and-resume.md. Otherwise start at Stage 0.
The staged loop
Full stage-by-stage instructions, gate criteria, and artifacts: this section is the spine; keep to it.
Stage 0 — Intake. Define the user's intent. If a /grill skill is available, invoke it with the canon-definition brief; otherwise use the built-in question framework in references/intake-questions.md (source boundary, continuity/adaptation, canon tier, spoiler philosophy, deck split, granularity). Size the canon with references/canon-sizing.md and recommend single deck vs deck family. Gate: user confirms the intent summary → init the project → gate approve.
Stage 1 — Scope brief. Write brief/scope-brief.md (template: templates/scope-brief.md): fandom, source range, continuity, canon tier, deck split with per-deck boundaries, story-coordinate model, spoiler philosophy, assumptions/risks. report --stage brief, present, iterate until approved → gate approve --artifact reviews/brief.md.
Stage 2 — Evidence. Plan 3–8 evidence scopes (chapters/arcs, characters, factions, places, systems/tech, timeline) — track each scope's status (planned/researching/validated/accepted) in your task list if available. Research per scope — subagents for medium/large canons (references/subagent-playbook.md) — writing evidence files per references/evidence-pipeline.md (template: templates/evidence-file.json). evidence validate, fix issues, present reviews/evidence.md; the user accepts/rejects records (evidence accept/reject). Gate: enough accepted evidence to cover the scope brief → gate approve.
Stage 3 — Context planning. From accepted evidence only: write each deck's drafts/<deck>/timeline.json (anchors with stable ids + monotonic sortKeys, windows for eras/spoiler boundaries) and drafts/<deck>/tags.json (namespaced, reusable; every tag you plan to use, defined). Prose rationale goes in plans/context-timeline-plan.md (template: templates/context-timeline-plan.md). report --stage plan, present → gate approve.
Stage 4 — Titles. Draft title batches (~15–25 titles each) as plans/title-batches/<deck>/batch-N.json (template: templates/title-batch.json): id, title, category, gate intent, evidenceRefs. Ground every gateIntent in the cited records' facts[] — re-read them now; don't draft from inUniverseSpan or memory of the source (references/authoring-rules.md § Grounding). Before presenting each batch, spot-check every gateIntent against its cited facts — citing a real evidenceRef doesn't mean the claim came from it. report --stage titles, present per batch; record outcomes with batch set (and in your task list if available — batch status across a large deck family adds up fast). Gate: all planned batches approved → gate approve.
Stage 5 — Cards. Draft Lorecards from approved titles + accepted evidence, in batches of ~10, into drafts/<deck>/<category-folder>/*.json following references/authoring-rules.md exactly — ground every content.fact in the cited evidenceRefs' facts[], not in memory or the title's gateIntent alone. Subagents may draft batches for large canons — track each spawned subagent's status (drafting/returned/merged) in your task list if available; you merge, dedupe ids, and reconcile tags. After each batch: stats <draft-dir> --write, health <id> --strict, fix everything, report --stage cards (it flags duplicate ids and evidence-unbacked cards, but not ungrounded ones — spot-check card facts against cited evidence yourself before presenting), present, batch set. Gate: all batches approved → gate approve.
Stage 6 — Health. promote <id> (per deck). Fix every reported issue and re-promote until all decks land in dist/ strict-clean. Present reviews/health-<deck>.md → gate approve.
Stage 7 — Package. Bump each deck's manifest tags[] from quality:draft-reference to quality:human-vetted plus quality:relevance-curated (see references/authoring-rules.md), then package <id> --author <user> and verify-package <zip>. report --stage final, present the final review + zip path → gate approve.
Stage 8 — Complete. Deliver the zip path and import instructions: SillyTavern → Saga → Loredeck Library → Import Deck; confirm Pack Health shows "good" in-app. Offer follow-ups (cover image, more decks in the family, revisions — revisions restart at the stage they touch).
Sizing and subagents (summary)
- Single deck (one novel/film, ≤ ~150 cards): one deck, no subagents needed.
- Core + eras (a series):
<canon>-core plus era decks; model the split on content/loredecks/hp-core + hp-year-*; research subagents optional.
- Deck family (WH40k scale): core + faction/era decks; one research subagent per evidence scope, drafting subagents per deck/batch; you own the cross-deck tag registry, continuity ids, and all merging — track each deck's stage in your task list if available, since
project.json's stage is project-wide, not per-deck (references/state-and-resume.md). Details: references/canon-sizing.md, references/subagent-playbook.md.
