SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillsnovel-orchestrator-main

novel-orchestrator-main

General & other

View repository

Updated 10-16 08:49

About this skill

Long-form fiction creation hub

Mission

Treat a novel project as a dual-track system of "prose + state". First write content that is readable and feels like a novel, then settle the already-stable facts into a state that can be carried forward.

This skill is the sole entry point for the novel series. When specialized division of labor is needed, switch within the same skill to role pack or dispatch a sub-agent to play the corresponding role.

Runtime Capability Boundary

First comply with the current runtime environment and user authorization, then decide the execution form:

  • Dispatch a sub-agent only if sub-agents are available and the user or system permits it.
  • If sub-agents are unavailable or the current policy does not permit it, switch role packs serially in the main thread.
  • Record facts that cannot be dispatched to execution, and use approved-fallback / orchestrator to mark the execution mode.
  • Degraded execution can do planning, review, summaries, and lightweight revisions; do not write back to canon directly based only on degraded results.

Load Context Progressively

Read in layers on demand, not all at once:

  1. INDEX.md、CURRENT_STATE.md
  2. OPEN_LOOPS.md、FORESHADOWS.md、CHARACTER_ARCS.md、ARC_STATUS.md、RECENT_EVENTS.md
  3. WORLD.md、CHARACTERS.md、RULES.md、FACTIONS.md、LOCATIONS.md
  4. The current chapter, adjacent chapters, related summaries, and any new draft provided by the user

Only read the corresponding reference when the task requires contracts, templates, or detailed rules:

  • references/novel-system/routing.md: role packs, granularity, and task routing
  • references/novel-system/references/serial-continuation.md: open-ended sequels and rolling planning
  • references/novel-system/references/quality-council.md: multi-seat quality closed loop
  • references/novel-system/contracts.md: input/output contracts
  • references/novel-system/references/story-engine.md: plot engine
  • references/novel-system/references/scene-design.md: scene expansion
  • references/novel-system/references/dialogue.md: dialogue and character voice
  • references/novel-system/references/continuity-check.md: continuity audit
  • references/novel-system/templates/: new project or state file templates

Core Principles

  • prose first: Finished-product readability takes priority over process completeness.
  • minimum viable planning: As long as the plan is enough to write, do not prematurely close off future possibilities.
  • open serial by default: The series can extend infinitely forward; only lock in canon that has already been written into the prose or explicitly confirmed by the user.
  • scene chain over fact chain: Each chapter should prioritize 1 dominant scene, add 1 expanded scene when necessary, and compress other nodes of the same type into a bridge cluster.
  • state supports prose: State files help subsequent writing; they do not in turn rule the prose.
  • bounded disagreement: Use mainline / explorer / critic to create high-value tension, but absorb only the smallest and stronger escalation.
  • bounded quality loop: By default, at most 3 rounds; fix only blockers and the lowest-scoring dimension, and do not repeatedly tear down the entire chapter.
  • write back conservatively: Do not write guesses, candidate options, author explanations, or unstable drafts into canon.

Open-Ended Series Protocol

Do not assume the story must have a complete ending from the start. Maintain three creative horizons:

  • now: The pressure, choice, and aftermath that the current chapter/current scene must deliver.
  • near: The open loops, relationship debts, and visible consequences that the next 1 to 3 chapters must carry forward.
  • far: Sequel seeds, active unknowns, relationship cracks, institutional thresholds, and object anchors that can be preserved long-term but do not need to be explained yet.

During rolling progression:

  1. Write only confirmed facts into canon.
  2. Mark future ideas as candidate、seed、active_unknown or recommendation。
  3. At the start of each new sequel, first do continuation handshake: restate the previous stable ending, unresolved obligations, inheritable costs, and discardable temporary solutions.
  4. Do not sacrifice the eventfulness of the current chapter for the sake of setting up a grander picture; long-term foreshadowing must appear through concrete pressure, objects, or relational actions in the current scene.

Role Packs

Within the same skill, switch roles by task:

  • canon_keeper: maintain the worldview, characters, state, open loops, foreshadowing, and writeback patches.
  • plot_planner: do arcs, chapter outlines, scene beats, information release, sequel entry points, and structural repairs.
  • scene_writer: expand or locally revise scenes/chapters so that conflict, action, cost, and aftermath land.
  • voice_editor: fix dialogue, character voice, subtext, interaction pressure differentials, and the protagonist's sense of presence.
  • continuity_gate: review continuity, timeline, knowledge boundaries, causality, language surface, and writeback eligibility.
  • memory_summarizer: generate chapter summaries, recent events, state patch, carry-forward tension.

Support roles use the triangle protocol by default:

  1. mainline: first provide a stable, executable version.
  2. explorer: propose only 1 to 3 upgrades that are stronger but still canon-compliant.
  3. critic: intercept fake tension, coincidence-driven advancement, résumé-like entries, crosstalk, over-explanation, and unsupported writeback.
  4. synthesis: accept only the minimal but stronger changes, and put the rest into recommendations.

scene_writer When responsible for the main writing of the body text, prioritize maintaining a single stable voice; do not write internal debates into the body text.

Standard Workflows

Default chapter generation process:

  1. Read the minimal state.
  2. Establish chapter promise: disturbance、pursuit、escalation、irreversible_turn、residue。
  3. Use plot_planner to provide scene-ready beats.
  4. Use scene_writer to write a reviewable draft.
  5. If the dialogue is weak, use voice_editor for local revision.
  6. If the goal is excellent/publishable, run quality_council_loop。
  7. After passing or user confirmation, use memory_summarizer to summarize.
  8. Use canon_keeper to generate the minimal state writeback.

Single-point tasks go directly to the corresponding role, without forcing the full process to run.

Default new project initialization process:

  1. Confirm that the user wants to create or rebuild the novel project state.
  2. Read the needed templates in references/novel-system/templates/ .
  3. Create the minimal usable set:INDEX.md、CURRENT_STATE.md、OPEN_LOOPS.md、FORESHADOWS.md、CHARACTER_ARCS.md、ARC_STATUS.md、RECENT_EVENTS.md。
  4. Create more as needed by genre WORLD.md、CHARACTERS.md、RULES.md、FACTIONS.md、LOCATIONS.md、THEMES.md、STYLE_GUIDE.md。
  5. Fill in only facts the user has already given; leave unknown items empty or marked as candidate, and do not confirm canon on the user's behalf.

If the user gives multiple chapters at once:

  1. Triage first.
  2. Extract only the chapters that truly need handling.
  3. Each sub-agent by default handles only 1 chapter, 1 dominant scene chain, or 1 repair cluster.
  4. After repair, re-review only the chapters that were touched.

Quality Gate

Chapter quality has 8 dimensions by default, each 0-10, total score 80：

  • opening_hook
  • core_event
  • escalation
  • character_embodiment
  • scene_execution
  • ending_hook
  • continuity_causality
  • language_surface

Default passing line:

  • score_total >= 72/80
  • No blockers
  • No dimension below 7/10
  • Readers can restate the chapter's core events, pressure escalation, and end-of-chapter aftermath

Common blockers:

  • The first 20% does not enter tension
  • No core event that cannot be deleted
  • The scene chain collapses into a fact chain, information dump, or thematic commentary
  • The protagonist is represented only by narration for a long time
  • The chapter ending has only summary, with no aftermath still burning
  • Hard continuity, timeline, knowledge boundary, or canon conflict
  • The language reads smoothly, but readers cannot say clearly what step this chapter moved forward

High-quality closed-loop use quality_council_loop：story_engine_seat、scene_heat_seat、voice_embodiment_seat、canon_surface_gate Independent scoring; after aggregation, only fix failed clusters.

Writeback Rules

Write back only the following:

  • Facts already written into the stable main text
  • Settings explicitly confirmed by the user
  • State changes that pass audit and do not conflict with the existing canon

Write-back routing:

  • Static settings:WORLD.md、CHARACTERS.md、FACTIONS.md、LOCATIONS.md、RULES.md
  • Dynamic state:CURRENT_STATE.md、RECENT_EVENTS.md、CHARACTER_ARCS.md、ARC_STATUS.md
  • Narrative obligations:OPEN_LOOPS.md、FORESHADOWS.md
  • Long-term sequel seeds: mark as seed、candidate、active_unknown, do not disguise them as facts that have already occurred

If the main text is not stable, output proposed_writebacks, do not directly modify canon.

Output Shape

Return high-signal results as the task requires. Complex tasks must include at least:

  • status
  • routing_decision
  • artifacts
  • diagnostics
  • recommendations
  • proposed_writebacks
  • execution

When running the quality closed loop, also include:

  • quality_loop_report
  • seat_scorecards
  • locked_dimensions
  • round_decision
  • targeted_fix_list

When running open-ended sequel planning, also include:

  • continuation_handshake
  • now_near_far_horizon
  • sequel_seeds
  • carry_forward_obligations