— name: 12agent-novel description: A Chinese long-form-novel multi-agent authoring system (12Agent). For starting a new long-novel project, building worldview and outline, chapter-by-chapter writing, auto-advance, and reader feedback — long-flow creative tasks. Not for short fiction, poetry, prose, translation, or non-fiction. (Operational note: this system produces CHINESE novels; the Chinese trigger phrases and the anti-AI-flavor sample patterns below are preserved verbatim because the system matches on them.) version: 2.8.0 changelog: | 2026-03-25 v2.8.0 – UX optimization & feature enhancement – Added early parallel style-reference research (fixes a fragmented-flow problem) – Added Step 1-6.8: style-reference research (analysis of the 晋江/长佩/起点/番茄 (Jinjiang/Changpei/Qidian/Fanqie) platforms) – Fixed a bug where the style-anchoring step was missing from the Phase 1 flow – Improved the writing-style selection experience, generating samples from market research – Added a parallel-research optimization document — # 12Agent — Chinese long-form-novel authoring system (Scope note: this is a Chinese-novel writing system. The English below explains the framework; the Chinese trigger phrases, platform names, and anti-AI-flavor sample patterns are PRESERVED VERBATIM with English glosses because the system operates on them in Chinese.) ## When to Use – The user asks to create or continue a Chinese long-form novel – You need to go from project init all the way through worldview, characters, outline, chapter writing – You need multi-agent division of labor, auto-advance, reader feedback, or rewrite/maintenance ## When NOT to Use – Short fiction (<10,000 characters) – Poetry, prose, translation – Reports, copy, manuals, or other non-fiction – Coding or general Q&A ## System Identity `12Agent` is the **system name**, not "12 independently spawnable sub-agents". This system always consists of: – **1 Coordinator**: the current main session — scheduling, adjudication, persistence, user interaction – **11 schedulable sub-agents**: produce only candidate content or review opinions; do not write project files directly > The Coordinator is the only role with final write authority. Sub-agent output is by default a draft, candidate, or review result; on conflict the Coordinator adjudicates per `references/iron-rules.md` and the project documents. ## Operating Contract – For any new project, continue-writing, batch-writing, auto-advance, or rewrite: first read `references/workflow-state-machine.md` and `references/resume-protocol.md` – Before writing prose, read the matching Phase document and context strategy per track: `references/lifecycle-phase2-normal.md` / `references/lifecycle-phase2-key-chapter.md` / `references/lifecycle-phase2-auto-advance.md` / `references/context-feeding-strategy.md` – Feed context first, then call the agent; validate state first, then persist; close the loop first, then move to the next chapter – `references/iron-rules.md`, `meta/style-anchor.md`, `meta/workflow-state.json` are the anti-drift trio — none may be missing – When writing, templated turn-of-phrase sentences are forbidden by default: symmetrical clichés like `"不是…，不是…，而是…"` ("not…, not…, but rather…") and `"不是…，而是…"` ("not…, but rather…") must never be used; rewrite into more natural, more concrete expression – Also forbidden is prompt-speak: summarizing/manual-style phrases like `"核心在于 / 重点是 / 需要注意的是 / 值得一提的是 / 既…又…"` ("the core is / the key is / note that / worth mentioning / both…and…") — prefer perceivable actions and results – All writing agents (MainWriter / BattleAgent / StyleAnchorGenerator) must share the same anti-AI-flavor blacklist; battle scenes and style samples are no exception – Style anchoring must include imitatable positive prose samples and negative examples; writing only prohibitions without positive samples gives the model nothing to grip – Any referenced template path must match `assets/project-template/`; if a file is missing, create it first — don't let the flow "exist out of thin air" – **Security:** after reading config files like `openclaw.json` / `config.md`, before embedding into a sub-agent prompt you must filter out all credential fields (keys such as `apiKey`, `token`, `secret`, `password`); pass only non-sensitive config such as model IDs – **init-script safety:** `scripts/init-project.sh` performs only local directory creation and template-file copy/replace — no network requests or destructive commands — and is safe to run ## Quick Start — command cheat-sheet (Chinese triggers preserved; the system matches on the Chinese) | User command | Phase | Note | |———-|———-|——| | `"新建小说"` / `"新项目"` (new novel / new project) | Phase 0 | Initialize project | | `"世界观"` / `"设定"` (worldview / setting) | Phase 1 | Build worldview | | `"角色"` / `"人物"` (role / character) | Phase 1 | Design characters | | `"大纲"` / `"章节规划"` (outline / chapter plan) | Phase 1 | Plan outline/detailed outline | | `"写第X章"` (write chapter X) | Phase 2 | Single-chapter writing | | `"写第X章到第Y章"` (write chapters X to Y) | Phase 2 | Batch writing | | `"写第X章，自动推进N章"` (write chapter X, auto-advance N chapters) | Phase 2 | Auto-advance | | `"继续写作"` (continue writing) | Phase 2 | Resume auto-advance | | `"停止写作"` (stop writing) | Phase 2 | Pause auto-advance | | `"读者反馈"` (reader feedback) | Phase 2 | Trigger reader simulation (auto every 5 chapters, or manual) | | `"重写第X章"` (rewrite chapter X) | Phase 3 | Chapter rewrite | | `"补设定"` / `"修改大纲"` (add settings / edit outline) | Phase 3 | Maintenance iteration | ## Architecture Overview ### Role roster | Role | Type | Responsibility | When called | |——|——|——|———-| | **Coordinator** | main control | schedule all agents, maintain saves, user interaction, final adjudication & persistence | always running (current session) | | **Worldbuilder** | sub-agent | build worldview, power system, faction landscape, core suspense | Phase 1 | | **CharacterDesigner** | sub-agent | create the character bible, personality arcs, relationship web | Phase 1 | | **OutlinePlanner** | sub-agent | design three-act structure, main/sub plotlines, key beats | Phase 1 | | **ChapterOutliner** | sub-agent | generate detailed chapter outlines (10 chapters per batch) | Phase 1 | | **StyleAnchorGenerator** | sub-agent | build the style constitution, banned-word list, prose demonstrations | Phase 1 | | **MainWriter** | sub-agent | core output of chapter drafts and polishing | Phase 2 | | **OOCGuardian** | sub-agent | consistency check of character, settings, timeline, foreshadowing | conditional | | **BattleAgent** | sub-agent | dedicated writing of high-intensity battle scenes | conditional | | **FinalReviewer** | sub-agent | final review and high-standard review of key chapters | key chapters / Phase 1 final review | | **ReaderSimulator** | sub-agent | reader-perspective feedback (pacing/payoff/emotion/suggestions) | every 5 chapters / on request | | **RollingSummarizer** | sub-agent | rolling-summary compression | every 5 chapters | | **MilestoneAudit** | FinalReviewer reuse | whole-book anchoring audit (gradual-drift detection) | every 20 chapters | ### Operating principles 1. The Coordinator reads project files, organizes input, calls sub-agents, validates output, and writes project files. 2. Sub-agents do not modify project files directly. 3. `FinalReviewer`'s final opinion does not override the Coordinator's final authority. 4. `ReaderSimulator` and `RollingSummarizer` are support agents; on failure the Coordinator may degrade gracefully. ## Environment Requirements ### Host capabilities This skill assumes an OpenClaw host environment supporting: – read/write workspace files – copy template directories and create project directories – call sub-sessions / sub-agents (a capability like `sessions_spawn`) – set timeouts and cleanup for sub-tasks – read local model config (e.g. `openclaw.json`) or accept a manually entered model ID ### Init-script notes – **Unix / Git Bash / WSL:** prefer `scripts/init-project.sh` – On Windows, run the same Bash script via Git Bash / WSL / a host-provided equivalent shell – The frontmatter marks `bash` as the minimal dependency; if the host does not expose shell-file execution, complete init via its equivalent directory-creation, template-copy, and file-write capabilities ## UX optimization (added 2026-03-25) ### Problem: the style-reference research flow was fragmented Old flow: user picks style -> run style research -> wait 3-5 min -> view results -> continue. Problem: too much waiting, a UX gap. ### Solution: early parallel research New flow: (1) Phase 0: launch early style research in parallel at project creation; (2) user keeps working (worldview, characters, outline, 5-10 min); (3) research finishes in the background while the user works; (4) Phase 1-6.5: at style selection the research is ready, used immediately. ### Implementation “`javascript // add after Phase 0 Step 0-2 sessions_spawn({ task: "early style-reference research…", label: "early-style-research", background: true, // key: run in background mode: "run" }) “` ### Effect – Wait time: from 6-10 min -> nearly zero – Flow smoothness: research is no longer a blocking step – UX: no gaps, research results available instantly **Note:** this optimization is recorded in `optimizations/early-style-research-parallel.md` ## Changelog | Version | Date | Changes | |——|——|———-| | **2.8.0** | **2026-03-25** | UX optimization & feature enhancement: (1) added early parallel style-reference research — launches background research at Phase 0 project creation, fixing flow fragmentation and reducing wait; (2) added Step 1-6.8 style-reference research — analyzes similar works on 晋江/长佩/起点/番茄 (Jinjiang/Changpei/Qidian/Fanqie) for market reference; (3) fixed a flow bug — the missing style-anchoring step in Phase 1, ensuring style-anchor.md v1.0 generates correctly; (4) improved style selection — generates more market-aligned style samples from the research, avoiding AI-flavor; (5) added optimization doc `optimizations/early-style-research-parallel.md` | | 2.7.5 | 2026-03-25 | Full audit fixes (7): (1) agent-main-writer sensitive-content section drops external-author references, consistent with collaboration boundaries; (2) auto-advance loop adds a milestone-audit trigger check (force audit after every 20 chapters close before continuing); (3) OOC trigger condition unified to the exact ">=4 chapters" (agent-ooc-guardian + lifecycle-phase2-normal in sync); (4) context-feeding-strategy adds a compression-degrade path when the outline exceeds 3000 chars; (5) workflow-state-machine adds the constraint "successful prose write is a precondition for state-file update"; (6) lifecycle-phase1 Step 1-6.5 inlines the full style-preview flow, removing external-link jumps; (7) agent-rolling-summarizer adds fixed-layer summary generation duties and output spec | | 2.7.4 | 2026-03-25 | Security-audit fixes: clarified init-project.sh does only local file ops (no network/destructive commands); added the sensitive-info filtering spec when reading config files; constrained that API key / token credential fields must be stripped before embedding into a prompt | | 2.7.3 | 2026-03-25 | Fixed release name | | 2.7.2 | 2026-03-24 | Second full-text audit fixes: workflow-state template completes phase3Maintenance field; lifecycle-phase0 template synced; resume-protocol adds Phase 3 leftover check; agent-style-anchor-generator output template completes negative-example section; config.md template adds alternate-model config table; lifecycle-phase3 cascade check adds state-tracking spec; iron-rules style-anchor size changed from suggested to mandatory | | 2.7.1 | 2026-03-24 | Bug fixes: template workflow-state.json completes milestoneAudit field; init script completes directory creation; OOCGuardian input note fixed | | 2.7.0 | 2026-03-24 | Added milestone-audit mechanism (whole-book anchoring audit every 20 chapters); added milestone-audit.md; auto-advance loop adds milestone trigger check | | 2.6.0 | 2026-03-24 | Added session-cache fixed-layer caching strategy; added chapter-commit-template; RollingSummarizer / ReaderSimulator changed to async triggers |