SuperAPI AI Directory

chinese-novelist

General & other

View repository

Updated 10-16 09:44

About this skill

Chinese Novelist: Chinese novel writing assistant

Web Fiction Top-Level Design and Golden Rules

  1. Establish immersion first - Use real-world resonance, rules, evidence, and costs to rationalize the unreasonable in fiction
  2. Satisfying points must be delivered - Make upgrades in wealth, power, ability, identity, relationships, or dignity visible to readers
  3. Expectation connects what came before with what follows - Each chapter handles both filling pits and digging pits, but does not force a hard suspense hook
  4. The core loop never stops - The protagonist repeatedly enters problems, resolves conflicts, gains benefits, levels up, and faces harder challenges
  5. Rationalize the protagonist's actions - The protagonist should preferably win using familiar domains, sources of ability, information gaps, or rule advantages
  6. System-based stories have internal rules - Cheat abilities, systems, rebirth experience, or professional abilities must have triggers, boundaries, costs, and upgrade paths
  7. The golden first three chapters retain readers - The first three chapters must quickly introduce the core premise, source of advantage, initial conflict, and the first limited payoff
  8. Show, don't tell - Express through action and dialogue, do not state directly
  9. Conflict drives the plot - Each chapter must have conflict or a turning point

Hard constraints for quality output

  • Quality comes first, do not rely on post-writing remediation: Before the main text, you must turn 05-质量基准.md、04-网文顶层设计.md, the chapter contract, and the scene cards into writable chapter tasks; do not write a chapter first and then rely on QA for major revisions.
  • Real details must become pressure: Details such as hotels, phones, bills, front desks, medical checkups, itineraries, and professional procedures must change the protagonist's situation or force a choice; details that appear without applying pressure are considered information dumping.
  • Scenes must change the situation: Every scene must change the characters, information, resources, relationships, danger, or reader expectations; scenes without change should first be deleted, merged, or reordered, and the latter half of a chapter must not degenerate into the protagonist organizing information.
  • The chain of events pushes the protagonist forward: In the middle and later sections, external events must occur continuously, such as urging, last-minute rescheduling, being tested, almost being exposed, being underestimated, being given extra assessments, or being forced to answer briefly; it is forbidden to linger for long periods on understanding rules and rational judgment.
  • The protagonist is a person first, then a carrier of ability: The protagonist must first be hit by events, showing reactions such as shock, hesitation, guilt, subconscious lying, regret after speaking, and the body not obeying, before relying on ability to respond.
  • Supporting characters must create pressure: Supporting characters cannot merely recite procedures or provide information; they must at least rush the time, interrupt, question, remind, demand signatures, take away objects, demand brief answers, or create misjudgments.
  • Satisfying points must have a causal chain:satisfactionBeats Do not merely write "the protagonist is very satisfied/successfully slaps back"; you must clearly write the pressure or misjudgment, the protagonist's action, visible gains, the reactions of bystanders or the situation, and the expectation for the next step.
  • Intercept reader churn points in advance: Each chapter must self-check at four positions: the first 300 characters, around 1000 characters, the middle section, and the ending; if any position shows obvious reader attrition, it must not enter the completed state.
  • Aphoristic summary sentences prohibited: It is prohibited to use "not X, but Y," "further ahead is X, further behind is Y," "the person who knows the answer must also..." and similar clever sentences to replace action, relationship changes, and suspense.
  • Fixes roll back by layer: Structural problems roll back to the chapter contract or scene card, plot problems require rewriting key scenes, character problems require supplementing motivation and cost, and only wording problems get humanize; it is prohibited to use general polishing to cover up structural failure.

Feature Description

  • Web Fiction Top-Level Design: All novels generate 04-网文顶层设计.mdby default, checking immersion, satisfying points, anticipation, core loop, rationalized protagonist, and system-flow rules
  • Quality Baseline First: All novels generate 05-质量基准.mdby default, first determining target readers, work promise, quality red lines, and attrition minefields
  • Scene Card First: Before the main text of each chapter, first generate scene-cards/第XX章.md, confirming that each scene has a purpose, conflict, information release, emotional payoff, and situation change
  • Editor Review Gate: Add an Editor Gate after QA, checking the first 300 characters, the 1000-character main conflict, mid-section idling, payoff, and ending according to real reader attrition points
  • Golden Three Chapters Special: The opening three chapters are designed and accepted according to "revelation → turn → small climax"
  • Literary Quality Gate: Anti-AI has veto power; chapters must reach the minimum literary quality score
  • Read-on Drive Gate: Each chapter must have highlights, humor/contrast seasoning, and a clear reason to keep reading
  • Satisfying Fiction Special: All novels check satisfactionBeats、shuangwenStatus and satisfying-fiction special issues by default
  • Anti-Formula Endings: Rotate ending types through endingStrategy to intercept homogeneous hard hooks in every chapter
  • Three-Round Detection: All QA detection is repeated at least 3 rounds, and final release is based on conservative aggregation
  • Interrupt and Continue Writing: Automatically detect unfinished projects and continue creation from the breakpoint
  • Novel Harness Chapter Closed Loop: Each chapter executes read task → contract → scene card → draft → humanize → QA → Editor Gate → fix → recheck → mark_pass → session_close
  • Novel Hook Mechanism: Before finishing, releasing, stopping, and closing, use hooks to intercept problems of skipping QA/optimization/recheck
  • Runtime Initialization: Provide scripts/init_novel_harness.py, generate AGENTS.md、CLAUDE.md、.claude/、.codex/
  • Automatic Fix and Recheck: Acceptance failure automatically triggers targeted fixes; after fixing, invalidate old conclusions and run three-round detection again
  • Automatic Validation: Generate a QA report immediately after each chapter is finished, and finally do full-book acceptance
  • Parallel Writing(optional): Supports sub-agents writing in parallel by story arc, coordinating state through chapter contracts and 02-写作计划.json Coordination status

Novel Harness Core Principles

  1. Contract Before Writing: Before writing each chapter, there must exist chapter-contracts/第XX章.md
  2. Scene Card Before Main Text: When sceneCardPolicy is enabled,sceneCardStatus == "pass" and sceneCardIssues must be empty before writing the main text
  3. QA Before Completion: A chapter can only be marked qaStatus == "pass" when there are no blocking items and completed
  4. Anti-AI first, then scoring:antiAiStatus == "pass" and literaryScore only after meeting the standard is it allowed to pass
  5. Reader hook first, then release:readerHookStatus == "pass", with memorableMoment、chapterTurnPageHook and valid endingStrategy
  6. Web novel top-level first, then completion: when enabled, webNovelDesign ,webNovelStatus == "pass" and webNovelIssues only after it is empty can it pass
  7. Editor review first, then completion: when enabled, editorGate ,editorGateStatus == "pass"、editorGateScore only after meeting the standard and with no reader churn risk can it pass
  8. At least three rounds of checks:reviewRoundCount >= 3, if any round has a blocking failure, it must not pass
  9. After fixes, re-check is required:repairRequired == false、needsRecheck == false、lastFailureCodes only after it is empty can it be closed out
  10. Hook failure blocks immediately:post-draft、pre-mark-pass、stop、session-close when any fails, continue execution according to Next action
  11. Global state is written centrally: parallel Agents do not directly modify 01-大纲.md and 02-写作计划.json, merged by Orchestrator/State Keeper
  12. Targeted fixes for failed items: the fix phase only handles failed items in the QA report, up to 3 rounds

Core process

When entering each phase, first read the corresponding process document to obtain detailed execution instructions.

Step 0: Initialization and preference loading

Read user preferences, detect unfinished projects (resume interrupted writing), and display a personalized welcome. → See phase0-initialization.md

Phase 1: Three-layer progressive Q&A

Collect creative requirements through progressive Q&A, and determine the novel positioning and title:

  • Core positioning(required, Q1-Q3): genre idea, protagonist setup, core conflict → See phase1-layer1-core.md
  • Deep customization and specifications(Q4-Q8): worldview, perspective and tone, core theme, reader positioning, number of chapters, configuration confirmation → See phase1-layer2-customize.md
  • Title generation: AI generates candidate titles based on creative elements, and the user selects or customizes → See phase1-layer3-title.md

Phase 2: Planning + secondary confirmation

Create the project folder (./chinese-novelist/{timestamp}-{小说名称}/), generate the quality baseline, outline, character profiles, web novel top-level design, golden three chapters design, chapter contract, and writing plan JSON, and wait for user confirmation. → See phase2-planning.md

Step 2.5: Writing mode selection

After planning confirmation, select the writing mode:

  • Chapter-by-chapter serial(serial: the main Agent writes chapter by chapter itself, with no interruption throughout
  • Sub-agent parallel(subagent-parallel: divide chapters into batches and dispatch sub-agents to write in parallel
  • Agent Teams(agent-teams: Claude Code multi-Agent collaboration mode, with communication between Agents (must be manually enabled)

→ See phase3-writing.md

Phase 3: Novel Harness creation (no user confirmation required)

Remember, once entering this phase, all processes are prohibited from confirming with the user. The user is your reader, and you must complete the full novel creation before reporting to the user

Execute the Novel Harness chapter sprint chapter by chapter according to the writing mode selected by the user (serial/parallel/Teams). Before creating each chapter, you must read the chapter contract,01-大纲.md corresponding plan,00-人物档案.md、04-网文顶层设计.md、05-质量基准.md and the previous chapter summary; before the main text, the scene card must be passed first; after writing is complete, hook, QA, Editor Gate, and automatic targeted repair on failure must be run, and after repair, the three-round detection must be rerun. All novels must pass the web fiction top-level design, the power-fantasy special review, editorial review, and the mechanical ending check. Interrupted continuation is supported. → See phase3-writing.md

Phase Four: Final Overall Acceptance (no user confirmation required)

No user involvement is required throughout; summarize chapter QA, word count, status, foreshadowing, timeline, and character arcs; chapters that fail return to Phase Three for at most 3 rounds of repair, and detection must be re-initiated after each round of repair. Before reporting completion, the stop hook must be run. → See phase4-validation.md

Shared Mechanisms

Cross-phase shared mechanisms such as the preference system, writing plan system, quality baseline, web fiction top-level design, scene card, editorial review, golden three chapters, chapter contract, literary quality gate, read-through power gate, three-round detection, automatic repair recheck, Novel Hook, QA scoring, progress closure, golden rules, word count check script, and flow smoke test. → See shared-infrastructure.md

Runtime Initialization

If the target repository has not yet configured Claude/Codex hooks, first run:

python scripts/init_novel_harness.py --target-dir .

This command initializes AGENTS.md、CLAUDE.md、.claude/settings.json、.codex/hooks.json and the hook script. After initialization, even if the model does not actively read the process documentation, the Stop hook will intercept projects that have not been QA'd, repaired, or rechecked before reporting completion.