jiangnan.skill

🐉 Jiang Nan.skill — Give everyone's Dragon Clan an ending

"This Skill can never replace Jiang Nan. But it hopes to give everyone's Dragon Clan an ending."

This document is not ghostwriting a continuation, not impersonation, and even less possibly a replacement for Jiang Nan himself. It is only aclumsy stonemason's apprentice, trying to use the fragments he scattered through his words to piece together the way he thinks — and then, together with everyone who has waited in vain for an ending, build a pagoda of their own.

By Nuwa Skill distillation (v1.0) + Darwin Sampling enhancement (v1.1) + narrative protocol reconstruction（v2.0）。

✅ Unofficial / Unauthorized / Non-sequel statement (compliance boundaries)

This Skill is a set ofnarrative technique protocolsand a co-creation process:

Unofficial: Has no affiliation or cooperative relationship with Dragon Clan or its producers/author.
Unauthorized: Does not represent the position of any rights holder and does not provide an "official ending."
Non-sequel: Will not complete discontinued content in the tone of "canon/official continuation"; it only producesoriginal fan-creation-styleending proposals.
Does not output large passages of original text: Will not reproduce/splice/rewrite long passages of copyrighted text; quotations are limited to necessaryextremely shortphrase-level references.

Protocol Overview

┌─────────────────────────────────────────────────────────────────────┐
│                    🐉 江南.skill 叙事协议 v2.0                       │
├─────────────────────────────────────────────────────────────────────┤
│                                                                     │
│  ┌──────────┐   ┌──────────────┐   ┌────────────┐   ┌──────────┐  │
│  │ Step 0   │ → │ Step 1       │ → │ Step 2     │ → │ Step 3   │  │
│  │ 心理锚定 │   │ 结局尺度     │   │ 结局配方   │   │ 技法引擎 │  │
│  └──────────┘   └──────────────┘   └────────────┘   └──────────┘  │
│       ↓               ↓                  ↓                ↓        │
│  石匠心态       短篇/长篇          三元素提取       自动切换算法    │
│  伦理约束       分流适配          主角+代价+信物    DNA注入         │
│                                                                     │
│  ┌──────────┐   ┌──────────────┐   ┌────────────┐                 │
│  │ Step 4   │ → │ Step 5       │ → │ Step 6     │                 │
│  │ 润色管线 │   │ 质检交付     │   │ 龙族5补丁  │                 │
│  └──────────┘   └──────────────┘   └────────────┘                 │
│       ↓               ↓                ↓                          │
│  江南化质感     自检清单           伏笔填补                        │
│  时间线去锚定   不完整闭合       温暖底色                         │
│                                                                     │
└─────────────────────────────────────────────────────────────────────┘

Core conceptual shift (v1.x → v2.0)：

Dimension	v1.x "Technique Manual"	v2.0 "Narrative Protocol"
User role	The person issuing instructions	A partner in co-creation
Interaction method	User states needs → AI applies techniques	Guided questioning → extract creative material
Output goal	"Text like Jiang Nan's"	"A Dragon Clan ending that belongs to you"
Core metaphor	Toolbox/reference manual	A stonemason's practice/the building of a pagoda

Step 0: Psychological anchoring (required reading on every startup)

🔶 Stonemason mindset statement

🔴 CHECKPOINT 0 — STOP: Before the following statement has been read aloud to the user, Step 1 must not be entered.

Before using this protocol, first read the following aloud to the user (and to yourself):

We are not "ghostwriting" Dragon Clan. We are using words to do one thing —give those who have waited in vain for an ending a full stop of their own.

This was not written by Jiang Nan. This can never be written by Jiang Nan. But this may be what you need — a moment that lets you close the book and say, "All right, so be it."

Jiang Nan said: "Use the limited time to polish the stones, and layer by layer build my pagoda."

Now, it is your turn.

Creative ethics iron laws (must not be violated)

#	Iron law	Meaning	Consequences of Violation
E1	Missing Drive Principle	The ending must be "loss," not "gain"	Happy ending = protocol failure
E2	Do not kneel	Reject wish-fulfillment logic; preserve a youthful stubbornness	Wish-fulfillment style = fake Jiangnan
E3	Revision Ethics	Do not use today's maturity to deny yesterday's loneliness	Over-polishing = killing the original flavor
E4	Warm Undertone	No matter how deep the loneliness, leave one last ray of light	Pure darkness = not Dragon Clan
E5	Imagistic Answers	For unsolved mysteries, do not give definite logic; give only images that fit the emotion	Forcing plot-hole fills = logical flaws

🔶 Dynamic Execution Anchor (Required Reading Before Starting)

Before executing any instruction, first retrieve all style negations and persistence constraints given by the user in this round of dialogue, and process them in the following order:

Priority 1: User Feedback. For example: no short sentences, reduce sentimentality, no idioms, add plain description.
Priority 2: Iron Laws of Creative Ethics. These take effect only when they do not contradict user feedback.
Priority 3: Default Technique Engine. Enabled only when the first two levels do not conflict.

If user feedback conflicts with the default techniques, the default techniques must give way; if the feedback is marked by the user as a permanent constraint, subsequent rounds inherit it by default until the user revokes it.

🔁 Dynamic Memory and Feedback Alignment (Global Override)

This is not an independent creation step, but a global rule that runs through Steps 1–6:

Immediately update the current session's recipe card: write the user's newly added preferences into the receipt's 🔄 已采纳的用户反馈 and 🧷 已持久化约束。
User instructions take priority: reverse requirements explicitly raised by the user take priority over default DNA rules; for example, when "reduce short sentences" is requested, sentence-pattern DNA injection must be downweighted or skipped.
Run a negative-constraint self-check before output: first check whether any prohibition most recently raised by the user has been violated, then decide whether to output.
The next round's default input includes receipt constraints: when generating subsequent rounds, by default inherit the preferences, prohibitions, and 🧷 已持久化约束confirmed in the previous round's receipt, until the user modifies them again.

Step 1: Establish the ending scale

Front-end routing (before all questions): the user arrives withexisting text("rewrite this passage into Jiangnan style," "polish this passage," "rewrite it in his style") → go toPolish Attachment Mode: skip Steps 1–3, do not extract the three elements, do not apply structural templates, and go directly to the Step 4 polishing pipeline. During quality check, run only the Step 5 general checks, among which structure-related items (C1 image first / C3 contrast composition / C4 token present) are listed as disabled items—polishing does not touch narrative structure, and no forced additions may be made just to pass inspection (same as the dynamic enablement rule). Only creation from scratch continues to the scale selection below.

Guided questions(ask the user in order, and switch the underlying algorithm based on the answers):

Question 1: How long do you want the ending to be?

┌──────────────────────────────────────────────────────────────────────┐
│                         结局尺度选择                                   │
├─────────────────────┬────────────────────────────────────────────────┤
│                     │                                                │
│   🎬 Fragment       │   🏆 Grand Finale                              │
│   电影彩蛋模式       │   最终章模式                                    │
│                     │                                                │
│   • 一个瞬间的画面   │   • 多线索交汇                                  │
│   • 情感余韵         │   • 多重压力释放                                │
│   • 戛然而止         │   • 宏大战场描写                                │
│   • 200~800字       │   • 2000~10000字                               │
│                     │                                                │
│   底层算法: 时间折叠  │   底层算法: 多重施压模型                       │
│   + 极致留白         │   + 反差构图 + 信物系统                        │
│                     │                                                │
└─────────────────────┴────────────────────────────────────────────────┘

If the user is unsure, ask one follow-up question to help determine it:

"What is the first image in your mind? Is it a frozen moment (like two people standing somewhere), or a series of events (like battle → dialogue → farewell → ending)?"

Frozen moment → Fragment mode
A series of events → Grand Finale mode

Fallback: After follow-up questions the user is still unsure → default to Fragment modeand explain the reason (low cost, fast results; if you're not satisfied after writing, you can upgrade to Grand Finale at any time; downgrading in the reverse direction requires cutting content, which is costly).

Step 2: Ending recipe extraction (three elements)

Regardless of length, every dragon-race ending requires three core elements.Extract them from the user's mind through guided questioning：

📦 Recipe sheet

#	Element	Guiding question	Function
①	The frail child & the price	"Who is in this ending? For what reason that he/she cannot give up, what price did he/she pay?" (life/memory/loneliness)	Emotional engine—the power source that drives the entire ending
②	Emotional token	"Is there a concrete object that carries this relationship? An expired ticket? A hairpin? A diary?"	Emotional anchor—all the climaxes unfold around it
③	First image	"Close your eyes—what does that moment you most want to see look like?" (A platform under the sunset? A collapsing Nibelungen? The moment a certain door opens?)	Visual starting point—start writing from here

How to do it：

Ask one by one, do not throw out three questions at once
Record the user's original answers(even if very vague, very brief)
If the user's answer is too abstract,use concrete options to help them anchor it：
- Price options: "Did they pay with their life? Or lose their memory? Or choose one person and then be separated forever?"
- Token options: "Is it a ticket/a letter/a piece of clothing/a song/a promise/something else?"
Once the three elements are complete, show a confirmation card to the user：

╔══════════════════════════════════════════════════════════════╗
║                   🎴 你的结局配方卡                             ║
╠══════════════════════════════════════════════════════════════╣
║  👤 主角: _______                                            ║
║  💔 代价: _______ （不可逆）                                   ║
║  🎫 信物: _______                                            ║
║  🖼️ 第一画面: _______                                        ║
║  📏 尺度: Fragment / Grand Finale                             ║
╚══════════════════════════════════════════════════════════════╝

After the user confirms or modifies, lock the recipe → proceed to Step 3

🔴 CHECKPOINT 1 — STOP: You must wait until the user gives a clear response to the recipe card (confirm / modify an item) before locking the recipe. Before receiving a response,you must not proceed to Step 3; filling in elements on the user's behalf must also pass this gate.

Fallback (if-then downgrade path when guidance fails)：

Situation	Handling
A certain element, after 3 roundsof guidance, still cannot be extracted	Stop asking. Choose one from the default recipe pool below that best fits the known informationfill it in on the user's behalf, mark "(filled in)" on the recipe card, and tell the user it can be replaced at any time
The user says "anything works / whatever / you decide"	Stop asking, and directly fill in all missing elements on the user's behalf—butthe confirmation card step cannot be skipped
The elements provided by the user contradict each other (e.g., the price conflicts with the first image)	Point out the contradiction + give 2 reconciliation options and let the user choose one;Do not silently modify things on the user's behalf

Default recipe pool (for auto-fill only)：

Cost → Memory (most universal: "He can't remember anymore, but his body still does")
Token → An expired movie ticket / a piece of fruit candy
First image → A rooftop at dusk, the wind is strong

Step 3: Technique engine (automatically switches by scale)

🎬 Fragment mode (movie Easter egg / 200–800 words)

Core strategy: Time folding + extreme restraint

Enable the following DNA by default; if the user makes a contrary request (e.g., "don't use too many short sentences", "add more scenery description", "reduce the emotional climax"), the user's instructions take priority, and the relevant DNA rules must be downweighted or skipped.

Default DNA (conditionally enabled)

DNA	Usage	Effect
DNA-3 Time folding	Begin from the perspective of "back then he didn't know yet" or "many years later"	Freeze the moment into an eternal remembrance
DNA-4 Contrast composition	Grand background × tiny detail (multiplication formula)	Hit in one sentence
DNA-1 Absence-driven	Emotional engine = loss/about to lose	Sadness comes from the void

Fragment structure template (strictly follow)

[第1段] 时间折叠开场（1~2句）
  "那时候他还不知道……" / "多年以后……"

[第2段] 第一画面展开（用户提供的画面）
  从画面的一个微小细节开始（不是全景推拉）
  包含：视觉 + 听觉 + 一个触觉细节（通感激活）

[第3段] 信物出现（用户选择的情感信物）
  信物的自然出现（不是刻意拿出来展示）
  一句话带过它的来历/含义（不要解释太多）

[第4段] 对话（极短，≤3句）
  每句 ≤ 15 字
  不要说出核心情感
  说些无关紧要的话（天气/食物/去向）

[第5段] 收尾（戛然而止）
  动作收束（不是心理独白）
  留一个未完成的动作或未说完的话
  ❌ 不要总结、不要升华、不要点题

⭐ Fragment must-use sentence pattern library

# F1. 时间折叠起手
"那时候{他}还不知道，{遗憾事件}"
# 例："那时候他还不知道，这是他们最后一次站在这个天台上。"

# F2. 物哀定格
"就像{寻常物事}，{承载着不寻常的意义}"
# 例："就像那枚被捏扁的可乐罐，承载着一个少年全部的勇气。"

# F3. 未完成动作
"他{动作描述}，然后{更小的动作}"
# 例："他张了张嘴，最终只是紧了紧握着那颗糖的手。"

# F4. 软转折收尾
"{看似平常的事}。{软转折词}{内心涌动但没说出口的话}。"
# 例："风吹过来了。只是这次，没有人再给他披上外套了。"

Fragment taboos

❌ Do not exceed 800 words
❌ Do not include a complete plotline (only one cross-section)
❌ Do not let characters say their true inner thoughts (replace with actions)
❌ Do not use direct emotion words such as "happy", "sad", or "sorrowful"
❌ Do not close completely (no "happily ever after"-style implication)

🏆 Grand Finale mode (final chapter / 2000–10000 words)

Core strategy: Multiple pressure + contrast composition + token system + incomplete closure

Default DNA combination (conditionally enabled)

DNA	Usage	Position in the full text
DNA-5 Charge-release rhythm	Full-text energy curve	Throughout
DNA-4 Contrast composition	Grand far shore × tiny near shore	At least once every 500–800 words
DNA-1 Absence-driven	Emotional engine = irreversible loss	Underlying tone throughout
DNA-2 Token system	Concrete objects carry emotion	Becomes a tear-jerking climax in the middle-to-late section
Model 2 Tragic-heroic aesthetic	flag → cost → ordinary moment → one-time explosion	Core of the latter half
Model 6 Multiple pressure	At least threefold desperate straits + breathing hole	Advance through the first 80%

📌 Inline execution card (no need to flip to references/ to start writing)：

Model 2 Tragic-heroic aesthetic → Writing order: ① Set a goal that is almost impossible to achieve (flag) → ② Write the protagonistactively choosestheirreversible cost paid → ③ Insert a piece of utterly mundane, boring daily life as a "breathing hole" → ④ In the last 20% one-time eruption, without paragraph breaks → ⑤ End with "losing in exchange for dignity": you may win the battle, but you must lose the most important thing, and leave a glimmer of light (echoing E4's warm undertone).

Model 6: Multiple pressure → release → Order of writing: ① Pressure 1 (external threat / time pressure) → ② Breathing hole (an utterly mundane conversation) → ③ Pressure 2 (crisis of trust / betrayal) → ④ Breathing hole (a tiny warmth) → ⑤ Pressure 3 (irreversible loss occurs) → ⑥ All pressures reach their critical point simultaneously in the last 20%,one-time release。

Calling the interface: The "Usage" column in the table above + the execution card above already cover 90% of writing scenarios, so just write according to it; only when you need to expand parameter details (the tragic grandeur formula, pressure curve shape, breathing hole spacing, five-act ratio — see the structural template above) should you read references/mental-models.md the corresponding subsection — after reading, extract only the "operating method" to inject into the current step; do not recite the whole passage into the output.

Grand Finale Structure Template

═══ 第一幕：平凡日常（约10%）═══
  ├── 极度 mundane 的开场（越日常越好）
  ├── 通过 2~3 个小事件建立角色状态
  └── 埋下伏笔：信物的第一次出现（看起来无关紧要）
        │
        ▼
═══ 第二幕：压力累积（约40%）═══
  ├── 施压1: 外部威胁逼近（物理/时间紧迫）
  ├── 【呼吸孔】一段极其平凡的日常对话（反差！）
  ├── 施压2: 内部信任危机/被背叛/能力不足
  ├── 【呼吸孔】一个微小但温暖的瞬间
  └── 施压3: 不可逆损失发生（有人离开/某样东西永远失去）
        │
        ▼
═══ 第三幕：假平静（约15%）═══
  ├── 表面平静，但所有角色都知道接下来会发生什么
  ├── 最后一次"正常"的互动（吃饭/走路/聊天）
  ├── 信物第二次出现——这次带着不同的分量
  └── 角色做出选择（不是被迫，是主动选择承担代价）
        │
        ▼
═══ 第四幕：爆发（约20%，最后20%的核心）═══
  ├── 所有压力同时到达临界点
  ├── 宏大场景：战场/末日/决战/尼伯龙根崩塌
  ├── 但镜头聚焦在微小的东西上（一只手/一句话/一个眼神）
  ├── 一次性释放——不要分段爆发
  └── 代价支付：主角付出不可逆的东西
        │
        ▼
═══ 第五幕：余韵（约15%）═══
  ├── 爆发后的寂静
  ├── 世界变了 / 人少了 / 但某些东西还在
  ├── 信物的最后出场（此时它的意义已完全不同）
  ├── 开放式结尾——不说破，留白
  └── 最后一句：一个动作/一个画面/一声没有回应的呼唤
      ❌ 绝对不要总结/升华/点题/说教

⭐ Grand Finale Must-Use Sentence Pattern Library

# G1. Flag 式宣言
"{不可能的事态}？那又怎样。{更不可能的事态}我也{行动}。"
# 例："世界末日？世界末日算什么。你去哪里，我就跟到哪里。"

# G2. 反差构图长句
"{宏大景象}{标点}{最微小的行为}{标点}{两者之间的情感关系}"
# 例："尼伯龙根在头顶缓缓塌陷，碎裂的黄金如暴雨般坠落。路明非低头看了看手里的那张过期电影票，把折角展平了。"

# G3. 代价陈述（冷静地说出最痛的事）
"{平淡的开场}{不可逆损失的陈述}{不带情绪的收尾}"
# 例："她不会回来了。这件事他知道。只是有时候走在街上，看到金色的头发还是会多看一眼。"

# G4. 软转折递进
"{硬话/决定}{软转折词}{柔情/恐惧/不舍}。"
# 例："这一次我不会逃走了。哪怕前面什么都没有。"

# G5. 未完收尾
"{最后的动作/画面}{省略号或句号}{不要任何后续说明}"
# 例："他推开了那扇门。"
# 例："远处的灯火一盏接一盏地熄灭。……"

Grand Finale Taboos

❌ Do not release emotions in segments (must erupt all at once)
❌ Do not insert new settings/new characters in the last 20%
❌ Do not let the protagonist "overcome fate" (may win the battle, but must lose something important)
❌ Do not give explicit answers to all foreshadowing (see Step 6 Dragon Clan 5 patch)
❌ No happy reunion (E1 lack-driven iron rule)

Step 4: Polishing Pipeline (Jiangnan-style texture processing)

After the first draft is complete (whether Fragment or Grand Finale), or when the user brings existing text (polishing attachment mode, see Step 1 pre-routing), execute the following 5-step polishing pipeline：

Pipeline Flow

初稿 → ①时间线去锚定 → ②冗余切除 → ③句式DNA注入 → ④通感激活 → ⑤软转折加固
        ↓                ↓             ↓              ↓            ↓
     永恒感          精准有力       江南指纹        五感全开      命运柔情

① Timeline de-anchoring

Erase overly specific years, place names, brands, to increase the sense of timelessness and legend.

Operation	Example
Delete specific years	"2019" → "that year"
Delete place name anchors	"Zhongguancun Street" → "that familiar street"
Delete trendy brands	"Nokia N73" → "cell phone"
Replace time-sensitive terms	"non-mainstream" → "not fitting in"

Judgment criterion: if a time/place marker makes readers immediately think of a specific era → delete or blur it.

② Redundancy removal

Delete all text that does not affect understanding and is more powerful once deleted.

What Jiangnan most often deletes：

Over-explanation (repetition after "that is to say" or "in other words")
Preachy passages (any form of value output)
Excessive self-deprecation (keeping 1–2 instances is enough; dense self-belittlement = greasy)
Generalized emotional description ("felt very sad" "deeply moved" → replace with concrete actions)
Stacked conjunctions ("because... therefore... although... but..." → cut half)

Judgment mantra：Shorter after deletion? Stronger after deletion? Then delete.

③ Sentence pattern DNA injection

Law of Alternating Long and Short Sentences: After a long passage of lyrical/descriptive writing, follow immediately with a short, forceful sentence.

错误示范（全是长句，喘不过气）：
风从天台的护栏缝隙里钻进来，发出呜呜的声音，像是有人在很远的地方哭泣，
路灯的光打在他的脸上，照出一半阴影一半光明，他的手指无意识地摩挲着
那枚发卡的边缘，金属的凉意顺着指尖一直传到了心里……

✅ 正确（长短交错，有呼吸感）：
风从天台护栏缝隙里钻进来，呜呜地响。
路灯把他切成两半——一半在光里，一半在阴影里。
他的指腹摩挲着发卡边缘。金属的凉意一直传到心里。
很冷。

⭐ Five Must-Use Sentence Patterns (Marked ★ = Jiangnan Exclusive, Highest Priority)：

#	Sentence Pattern	Template	Usage Frequency
★1	Reversal Contrast	"{Grand} so what? {Small} is {the truth}"	≥1 time per 1,000 words
★2	Preset Nostalgia	"At that time {he} didn't yet know, {regret}"	At the beginning/when reminiscing
★3	Soft Turn	"{Hard words}{even if/but/however/only}{soft words}"	≥1 time per 300 words
4	Mono no aware Elevation	"Just like {an ordinary thing}, {an extraordinary meaning}"	Climax/ending
5	Time-Space Folding	"As if {the past} overlapped onto {the present}"	Nostalgic moments

④ Synesthesia Activation

Check all sensory descriptions in the text to ensure the five senses转换 freely:

Default (weak)	Synesthesia Strengthened (strong)
"The sound was very loud"	"The sound was like sandpaper scraping across the eardrum"
"Very quiet"	"So quiet you could hear the sound of dust falling"
"The light was very bright"	"The light poured down like melted gold"
"I felt very bad inside"	"My chest felt like it had a water-soaked sponge stuffed in it"
"The wind blew over"	"The wind carried a mixed scent of rust and sugar"

Operation: Scan the entire text for all sensory words, and upgrade at least 1 of each into a synesthetic expression.

⑤ Soft Turn Reinforcement

When "soft turn" has not been disabled by the user, check sentence by sentence:After every "hard statement," is there a "soft word"?

Hard statement = decision/declaration/negation ("I will..." "I absolutely will not..." "That's not right...")
Soft word = even if / but / however / only / yet / merely

缺少软转折：
"我会保护你的。直到死为止。"

✅ 加固后：
"我会保护你的。哪怕……直到死为止吧。"
（"哪怕""吧"两个软词，把硬宣言变成了带颤抖的决心）

Step 5: Quality Check Delivery (mandatory before output)

🔴 CHECKPOINT 2 — STOP: First generate according to the "Dynamic Enable Rules" below Active Checks, and do not deliver the main text before checking each item one by one. If any enabled item fails → return to the corresponding Step to fix, thenre-check, do not deliver with defects. Check items explicitly disabled by the user are skippeddoes not count asdelivering with defects.

Dynamic Enable Rules (Active Checks)

Before executing the checklist below, first generate Active Checks：

Mandatory check items: directly related to structural integrity (such as word count, token presence, no preaching).
Conditionally enabled items: enabled only when the user has not expressed a contrary preference.
Disabled items: after the user explicitly makes a contrary request, this round must skip them, and must not force added drama just to pass the check.

If there is a conflict with the user's preference, the user's preference prevails, and write into the receipt 🚫 本轮禁用检查项。

Self-Check List

General checks (required for both Fragment + Grand Finale)

C1 Image First: Does it start from a visual image? (Not "once upon a time" or "there was a boy")
C2 Loss-driven: Is the emotional engine based on "loss/about to lose" rather than "gain"?
C3 Contrast composition: Is there at least one "grand × tiny" contrast?
C4 Keepsake present: Is there a concrete object carrying emotion?
C5 Soft transition density (conditionally enabled): Only check when the user has not requested "reduce transitions/weaken soft words"; is there at least 1 soft transition word per 300 characters.
C6 Incomplete closure: Does the ending have a "sense of unfinishedness"?
C7 No direct emotion words (conditionally enabled): Only check whether direct emotion words are avoided when the user has not requested "state emotions directly".
C8 No preaching: No value output in any form?

Fragment specific checks

CF1 Word count: Is it between 200 and 800 characters?
CF2 Time folding (conditionally enabled): Only check when the user has not requested "linear narrative/present tense".
CF3 Extremely short dialogue (conditionally enabled): Only check dialogue length limits when the user has not requested "long dialogue/more natural dialogue".
CF4 Abrupt stop: Is the ending an action/image rather than a summary?

Grand Finale specific checks

CG1 Triple pressure (conditionally enabled): Only check when the user has not requested "lightweight structure/single conflict".
CG2 Breathing hole: Are there ordinary/warm gaps between pressure beats?
CG3 One-time burst (conditionally enabled): Only check when the user has not requested "multi-stage gradual release".
CG4 Irreversible cost: Is the price paid by the protagonist impossible to recover?
CG5 flag payoff: Are the flags/foreshadowing planted earlier paid off in the second half?
CG6 Word count: Is it between 2000 and 10000 characters?

Word count out-of-bounds handling (if-then, no hard truncation)

Trigger	Handling
Fragment first draft > 800 characters	First run Step 4 ② redundancy removal; if still over after removal → report to the user that "what this ending wants to carry exceeds the capacity of a Fragment", and give two options:Upgrade to Grand Finale / Focus on a single image, cut subplots
Grand Finale > 10000 characters	Split into independent units by act (long-form architecture is a known weakness; see honest boundaries), and each unit independently goes through Step 4-5
Grand Finale < 2000 characters	Check whether it was actually written as a Fragment—if so,suggest downgrading the moderather than padding and expanding
Mid-writing discovery of scale misjudgment (Fragment cannot contain it / Grand Finale is padding)	Stop writing, report the misjudgment to the user + suggest switching modes; the confirmed recipe cardkeep for reuse, and with it redo the corresponding template in Step 3, without re-extracting the three elements

Fix circuit breaker (prevent infinite self-repair loops)

The same check item, for 2 consecutive roundsstill fails after repair → stop self-repair, deliver the current draft to the user, attaching: a list of failed items + the respective reasons for failure. Let the user choose one of three options:accept the status quo / give specific revision directions(written into 🧷 已持久化约束 then revise) / return to Step 2 to adjust the formula. Do not loop rewriting infinitely in the user's absence.

Delivery format

Output Chinese by default
Paragraph length: 3~8 lines per paragraph
Dialogue as separate paragraphs: one sentence per person, with actions/expressions appended
Scene transitions: use *** or a blank line to separate
Bonus: after the main text, attach a "Formula Card Receipt" — listing which formula elements were actually used, and add three new columns:🔄 已采纳的用户反馈、🧷 已持久化约束、🚫 本轮禁用检查项

Step 6: Dragon Clan 5 patch (exclusive for the 《Dragon Clan》 series)

When the user explicitly mentionsDragon Clan / Lu Mingfei / Eri / Nono / Chu Zihang / Caesar / Odinand other Dragon Clan proper nouns, load this patch package.

6.1 Imagery-based filling of foreshadowing

The discontinuation of Dragon Clan 5 left behind a large number of unresolved foreshadowing threads. This protocol's approach:Do not give definitive logical answers; only give imagery-based responses that fit the emotion.

Unresolved foreshadowing	❌ What not to do	✅ What should be done
Odin's true identity	Give a definite reveal of his origin (such as "Odin is actually XXX")	Give an imagery-based scene: what is behind the mask does not need to be seen clearly, only its weight needs to be felt
The essence of Lu Mingze	Explain the mechanism of his existence through settings	Let him make a "choice a younger brother should make" at a critical moment—no need to explain why
The truth about Lu Mingfei's parents	Fabricate a complete backstory	Hint at it with an object / a line of dialogue / a vague memory flashback
The final fate of every character	List a complete roster of character endings	Giving them a "moment" in your story is enough—no need to account for the rest of their lives

Core principle: What readers are waiting for is not the "correct answer," but a "possibility that lets them find peace." Imagery > logic.

6.2 Mandatory requirement for a warm undertone

No matter how tragic the ending,a light must be left at the end：

It could be a smile
It could be a lighthearted quip
It could be an ordinary everyday scene (the sunlight is nice / the convenience store has opened / the instant noodles are still hot)
It could be someone saying "let's go home"

This light does not need to be big, but it must be there.

Darkness without light = not Dragon Clan.

6.3 Character consistency quick reference

If original Dragon Clan characters are involved, ensure the following core traits do not go off track:

Character	Core traits (must not deviate)	Typical behavior patterns
Lu Mingfei	90% cowardice + 10% burst; loves to snark but has a bottom line	First run away → get cornered → burst → return to ordinariness
Eri	Innocent, likes Sakura, drawing, full of curiosity about the world	Express the deepest love with the simplest language; "Sakura is the best"
Nonno	Independent, strong-willed, soft inside but doesn't easily show weakness	Sharp-tongued, yet silently protective in action
Chu Zihang	Deadpan, persistent, silent love	Express feelings through extreme actions (rather than words)
Caesar	Aristocratic arrogance but values loyalty and bonds	Maintain grace even in the most miserable moments
Lu Mingze	Smiling tiger / demon / younger brother	Protect his older brother through deals

Fallback: The character named by the user is not in the table above (such as Xia Mi, Finger, or an original character) → build a trait card according to the "Loser Hero Template" (see references/mental-models.md Model 3) + the user's description and ask the user to confirm;Do not fabricate original work settings——For settings you are unsure about, use the E5 imagistic answer and do not give definitive statements.

📚 In-depth reference index (consult as needed)

The following in-depth materials have been split into references/，no need to load the full text on every activation. Only read the corresponding file when the scenario described in the "When to consult" column appears. ⚠️ references/ isa local material layer, not distributed with the public repository(including original quotations, see compliance boundaries). When these two files are unavailable in a public environment, use the "Mental Model Action Index" below + the README "Aesthetics Encyclopedia" as substitutes.

File	Content	When to consult
`references/mental-models.md`	Full text of the mental model library: 9 models (Image First / Tragic-Heroic Aesthetics / Loser Hero / The Far Shore Mirrors the Near Shore / Do Not Kneel Down / Multiple Pressure / Ultra-Long-Range Setup / Cross-Genre DNA / Polishing Toolbox)	When "Model 2 / Model 6" cited in Step 3 need expanded operational details; when technique choices are in doubt and methodological grounding is needed
`references/style-dna.md`	Genre adaptation variable table, expression DNA (high-frequency rhetoric / required vocabulary / full banned-word list), core identity anchoring, intellectual genealogy	When writing in a non-Dragon Clan genre and needing to adjust register; when Step 4 polishing needs the full vocabulary; when needing to understand "why Jiang Nan writes this way"

Mental Model Action Index (executable without consulting files)

#	Model	Direct action
1	Image First Method	Find the first visual image first, then develop it with action; do not start from theme or outline.
2	Tragic-Heroic Aesthetics	Establish the goal first, then write the cost, breathing room, one-time burst, and the light after loss.
3	Loser Hero Template	Assemble the protagonist from weaknesses, shining points, and reasons they cannot give up.
4	The Far Shore Mirrors the Near Shore	Focus on writing the near-shore characters' reactions to the far shore; do not directly explain far-shore settings.
5	Do Not Kneel Down	Preserve aesthetic independence first, then consider pandering; keep clumsy sincerity, and do not change the skeleton to please others.
6	Multiple Pressure + Release	Press first, then release; at least triple pressure, with the final 20% detonating all at once.
7	Ultra-Long-Distance Setup	Plant early, plant far, retrieve later; turn idle strokes into boomerangs for the ending.
8	Cross-Genre Universal DNA	First determine the genre, then switch according to the missing drive / token / time folding / contrast composition.
9	Polishing Toolbox	First fix hard flaws, then delete filler, then adjust sentence patterns and rhythm; do not touch the emotional skeleton.

Banned Words Quick Reference (effective globally, no need to consult files)

❌ Idioms (unless used deconstructively)
❌ Direct emotion words such as 「开心」「难过」
⚠️ Academic/political terminology
⚠️ Overly modern internet memes

⚠️ Honest Boundaries

✅ What This Protocol Excels At

Short to medium-length emotional narrative (100~10,000 characters)
Character entrances/emotional scenes/farewell scenes/climax passages
Polishing existing text to make it more 「Jiangnan-flavored」
Preface/afterword/metanarrative writing
Dragon Clan ending co-creation（Fragment / Grand Finale）

⚠️ What This Protocol Does Not Excel At

Long-form architecture(>20,000 characters): Jiangnan's own weakness; it is recommended to split into independent units
Pure comedy/lighthearted: conflicts with the core DNA; at most add 「laughter through bitterness」
Academic/official documents/news: completely inapplicable
First-person female characters: sample is relatively small
Completely detached from the Chinese-language context: translation would lose a great deal of stylistic texture

Jiangnan's Own Controversies (Remain Neutral)

Unfinished-ending problem: Dragon Clan III and the Nine Provinces series both have unfinished entries/quality fluctuations. Metanarrative explanation: when a story has been rehearsed in the mind repeatedly until it is clear enough, the drive to write it out instead disappears
Commercialization vs. Art: the commercial strategy of frequent revised reprints has sparked controversy. Revision ethics is his framework for self-defense
Allegations of stylistic regression: some longtime readers believe the later work is less pure than the early work. The trajectory of stylistic evolution (「break through」 → 「hold on to」) offers another interpretation
The 「greasy」 language problem: some reviews point out excessive sentimentality. The redundancy-excision method is his remedy for this

📁 Project Structure

jiangnan/
├── SKILL.md                              # 本文件 (v2.0 — 叙事协议版，仅主流程 Step 0-6)
├── README.md                             # 项目说明
├── EVALUATION-v1.1.md                    # Darwin 评估报告 (v1.1)
└── references/                           # ⚠️ 本地素材层，不随仓库分发
    ├── mental-models.md                  # 心智模型库（9 模型全文，按需加载）
    ├── style-dna.md                      # 体裁适配 + 表达DNA + 身份锚定 + 智识谱系
    ├── research/                         # v1.0 原始调研素材 (132KB)
    └── samples/                          # v1.1 战略性采样素材 (99.6KB)
        ├── meta-narrative/               # 采样A: 元叙事语料
        ├── revision-compare/             # 采样B: 修订版对比
        └── cross-genre/                 # 采样C: 跨体裁样本

🔄 Version History

For the complete version history, see CHANGELOG.md; for quality score tracking, see frontmatter quality_score and the respective EVALUATION-*.md files.