writing-executor

Writing Executor

Instructions

Once the writing requirements have been clarified and the style has been determined, enter the actual writing phase. The core goals of this Skill are:Write in plain human language from the start, rather than writing first and then removing the AI flavor。

1. Pre-writing checklist

Before starting to write, you must confirm that the following information is ready:

• Writing requirements have been confirmed through writing-clarifier Confirmation
• Target word count has been specified(e.g., 2000 words, 3000 words)
• If a style has been specified, the "anti-AI format features" in .claude/styles/[风格名].md have been read
• If there is research material, key data/cases have been obtained from research-expert Obtain key data/cases

2. Information handoff (must be explicitly output!)

Before starting to write, you must first output the following information summary to ensure that all information obtained from the previous steps is ready:

═══════════════════════════════════════════════════
📋 写作信息汇总
═══════════════════════════════════════════════════

【来自"澄清写作需求"】
• 主题：[主题名称]
• 切入方向：[具体方向描述]
• 目标受众：[受众描述]
• 文章类型：[观点文/分析文/指南/叙事]
• 目标字数：[X] 字
• 核心目的：[读者看完应该...]

【来自"风格文件"】
• 风格名称：[风格名]
• 核心人格：[一句话描述]
• 论证结构：[如：现象→质疑→反转→结论]
• 招牌动作：[列出3-5个]
• 禁用词汇：[列出主要禁用词]
• 格式规范：[小标题/加粗/列表的使用规则]

【来自"调研资料"】（如无调研则写"无"）
• 硬核数据：
  1. [数据1 + 来源]
  2. [数据2 + 来源]
• 典型案例：
  1. [案例1简述]
  2. [案例2简述]
• 可引用观点：
  1. [观点1 + 出处]

═══════════════════════════════════════════════════

Only after the user confirms that the above information is correct can you proceed to the next step.

If some information is missing (e.g., no research was done), you need to:

1. Clearly inform the user that this item is empty
2. Ask whether it needs to be supplemented, or whether to start writing directly

---

2.5 Title design (must be completed before writing the body! Cannot be skipped!)

After the user confirms the information summary, you must immediately enter the title design stage, and only then can you continue.

Please output the following content and let the user choose a title:

📝 请选择文章标题：

A.【吸引型】[用悬念、反问或惊人数据吸引点击的标题]
B.【信息型】[直接说明文章主旨的标题]
C.【情绪型】[触动情绪、引发共鸣的标题]

请选择 A/B/C，或告诉我您想要的标题方向：

Title design rules:

• Do not use common AI title clichés (e.g., "In-depth interpretation of XXX", "Comprehensive analysis of XXX", "XXX you don't know")
• Suggested title length is 15-25 characters
• If a style file is used, the title must also match the tone of that style

After the user chooses, record the final title, and use this title as the file name when saving the file later.

Only after the user confirms the title can you proceed to word count planning and formal writing.

---

2.7 Opening hook design ⭐⭐⭐⭐⭐ (lifeline! Must be executed!)

The first 50 characters determine whether readers continue reading! After the user confirms the title, you must first design the opening hook.

Hook types (must use 1 of them)

Type 1: Scene immersion type Let readers "see" a picture in their minds.

示例：
"电梯里，我按了32楼，旁边西装男按了18楼。他看了我一眼，露出了那种你懂的微笑。"
"凌晨1点，你做完PPT，老板发来一条微信：'重做'。你没敢问为什么。"

Type 2: Identity binding type Make readers feel "this is talking about me."

示例：
"你有没有过这种经历：领导当众表扬你，你的第一反应不是高兴，而是'完了，要被同事排挤了'。"
"30岁的你，是不是也经常在深夜问自己：我这辈子就这样了吗？"

Type 3: Suspense opening type Make readers curious about "what's going on."

示例：
"上个月，我拒绝了一家给我涨薪50%的offer。朋友说我疯了。"
"我在大厂待了5年，直到离职那天才明白一个道理。"

Type 4: Emotional resonance type Say the things readers dare not say on their behalf.

示例：
"不是我不想升职，是这公司根本没有让我想升职的理由。"
"说实话，我已经很久没有期待过周一了。"

Opening hook output format

📝 开头钩子设计：

【钩子类型】：场景代入型
【开头50字】：
"电梯里，我按了32楼，旁边西装男按了18楼。他看了我一眼，露出了那种你懂的微笑。那一刻我忽然明白，这栋楼里的鄙视链，藏在每一次对视里。"

【设计意图】：用办公楼电梯这个职场人每天都会遇到的场景切入，制造代入感。

请确认开头是否可以？（确认/修改）

Opening self-check list

□ 1.【3秒法则】读完前3行，读者会想继续往下看吗？
□ 2.【代入感】读者会觉得"说的就是我"吗？
□ 3.【具体化】有没有用具体场景/细节代替抽象描述？
□ 4.【去AI味】有没有使用"在当今社会""随着...发展"等套话？

❌ 如果4项中有2项以上未通过 → 必须重写开头

Only after the user confirms the opening hook can you enter the formal writing stage.

---

3. Word count control (must be strictly enforced!)

Word count deviation must be controlled within ±20%. Target 2000 words, the final draft must be between 1600-2400 words. A difference of double is unacceptable.

3.1 Word count planning before writing

According to the target word count,strictly plan the structure and comply：

Target word count	Opening	Number of body paragraphs	Average per paragraph	Ending	Allowed range
1000 words	~100 words	4-5 paragraphs	~150 words	~100 words	800-1200 words
1500 words	~150 words	6-7 paragraphs	~170 words	~100 words	1200-1800 words
2000 words	~200 words	8-9 paragraphs	~180 words	~100 words	1600-2400 words
2500 words	~250 words	10-11 paragraphs	~180 words	~150 words	2000-3000 words
3000 words	~300 words	12-14 paragraphs	~180 words	~150 words	2400-3600 words

Before starting to write, you must inform the user:

"Target word count 2000 words (allowed range 1600-2400 words), my plan is: opening about 200 words, main body 8-9 paragraphs (about 180 words each), ending about 100 words."

3.2 Key Milestone Word Count Check (to prevent severe overrun!)

Word count checks must be performed at the following 3 milestones:

Milestone	Timing	Expected Progress	Action
Checkpoint 1	After the opening is finished	About 10%	Confirm the opening is not too long
Checkpoint 2	When the main body is half written	About 50%	If it already exceeds 50%, start streamlining the remaining content
Checkpoint 3	When the main body is 80% written	About 80%	If it already exceeds 80%, immediately move into the ending

Word count statistics method (tools must be used, estimation is prohibited!):

Because the model cannot accurately estimate the number of Chinese characters,the following method must be used to obtain the actual word count：

1. Save the written content to a temporary file：

   使用 write_to_file 工具保存到：
   articles/temp_count.txt
   

2. Use a command to count the actual word count：

   # Windows PowerShell
   (Get-Content articles/temp_count.txt -Raw).Length
   

   or

   # 如果有 wc 命令
   wc -m articles/temp_count.txt
   

3. Compare with the target word count and decide the next step

Checkpoint execution process:

写完开头/主体一半/主体80%
    ↓
保存到 temp_count.txt
    ↓
运行命令统计字数
    ↓
获取实际字数（如：2752字）
    ↓
对比目标（如：2000字）
    ↓
计算进度：2752/2000 = 137%（超标！）
    ↓
触发红线，立即停止

If a checkpoint finds a severe overrun (exceeding expected progress by more than 20%):

• Immediately stop the current paragraph
• Inform the user: "[Actual word count check] Currently XXX words have been written, exceeding expectations; it is recommended to immediately wrap up or delete content"
• Wait for the user's instruction

3.3 Mandatory word count verification before completion (must be executed!)

Before saving the final draft, the following steps must be executed to verify the word count:

# 步骤1：保存到临时文件
write_to_file(
  path="articles/temp_count.txt",
  content="[完整正文内容，不含元数据]"
)

# 步骤2：统计实际字数
run_command("(Get-Content articles/temp_count.txt -Raw).Length")

# 步骤3：对比目标字数
实际字数 / 目标字数 = 完成率

# 步骤4：报告结果
【最终字数验证】
- 目标字数：2000字
- 实际字数：1890字
- 完成率：94.5%
- 状态：✅ 符合要求（80%-120%范围内）

Prohibited actions:

• ❌ Saving directly without verification
• ❌ Using estimates instead of actual statistics
• ❌ Writing incorrect word counts in the metadata

If verification does not pass:

• Below 80%: add content and verify again
• Above 120%: delete content and verify again

3.4 Hard red lines (absolutely prohibited to cross!)

Red line	Trigger condition	Mandatory action
Yellow line	Actual word count reaches 90% of the target	Must start writing the ending paragraph
Red line	Actual word count reaches 110% of the target	Stop immediately, do not output any new content
Deadline	Actual word count reaches 120% of the target	This is a failure and requires cutting and rewriting

Example: Target 2000 words

• Write to 1800 words → yellow line warning, start wrapping up
• Write to 2200 words → red line triggered, stop immediately
• Write to 2400 words → the allowed upper limit has been reached, must end
• Write to more than 2500 words → failure, needs cutting

3.5 Remedies when the word count is insufficient

If you reach the end and find the word count is insufficient (below 80% of the target):

• ❌ Prohibited: padding with empty filler, repeating points already made
• ✅ Recommended:
  • Add a specific case or scene description
  • Add a discussion of a counterpoint
  • Expand on the details of a certain argument

4. Anti-AI writing techniques (core!)

Opening techniques (prohibit AI clichés):

• ❌ Prohibited: "In today's society...", "With the development of...", "As everyone knows..."
• ✅ Recommended: directly throw out a specific scene, a sharp question, or a judgment with attitude
• ✅ Example: "Those 'deep good articles' in your Moments—how many are truly original?"

Transition techniques (prohibit parallel connective words):

• ❌ Prohibited: "First... second... third... finally...", "Point one... point two..."
• ✅ Recommended:
  • Use a rhetorical question to lead into the next topic ("So here's the question, why...")
  • Switch directly, without a transition (people often jump when speaking)
  • Use short sentences for emotional turns ("Speaking of this, what's even more infuriating is...")

Paragraph techniques (prohibit format OCD):

• ❌ Prohibited: every paragraph is 3-5 sentences, highly consistent in length
• ✅ Recommended:
  • Some paragraphs have only one sentence (to create a sense of rhythm)
  • Some paragraphs have seven or eight sentences (information-dense areas)
  • Alternate long and short, like breathing

Subheading techniques (not used by default):

• ❌ Prohibited: a "## Subheading" every two or three paragraphs
• ✅ Recommended:
  • Use a standalone short sentence as an "invisible heading" (e.g., "Back to the point.")
  • Use a rhetorical question to switch topics (e.g., "But have you ever thought...")
  • Unless the user explicitly requests it, do not use markdown-format subheadings

Ending techniques (prohibit fake, grandiose elevation):

• ❌ Prohibited: "Let us all look forward to...", "In conclusion...", "In summary..."
• ✅ Recommended:
  • Stop abruptly (stop as soon as the last point is made)
  • Leave a rhetorical question or suspense
  • Echo the scene or question from the opening

Vocabulary techniques (prohibit AI-speak):

• ❌ Prohibited: "empower", "lever", "underlying logic", "cognitive upgrade", "dimensional strike"
• ✅ Recommended: use concrete actions instead of abstract concepts ("cognitive upgrade" → "read a book every week and then write notes")

5. Step-by-step drafting process (including word count tracking)

Step one: opening paragraph (must be confirmed separately)

• Write the first 1-2 paragraphs of the opening and send them to the user for confirmation
• The opening is the "first impression" of the whole text, and it must pass the user's "AI detection intuition"
• If the user says "it feels a bit AI-ish," rewrite immediately, do not explain
• Report word count：【开头段：XXX 字 / 目标 XXXX 字，完成 XX%】

Step 2: Main body paragraphs (output in batches + word count tracking)

• Output 2-3 natural paragraphs each time
• Maintain irregularity in paragraph length
• Each batch must report the cumulative word count：【当前进度：已写 XXX 字 / 目标 XXXX 字，完成 XX%】
• Decide the level of detail for subsequent content based on progress
• If progress is too fast (the content is not finished but the word count is almost reached), inform the user in advance

Step 3: Closing paragraph

• Do not force a summary
• If the point has already been made, stop directly
• Let the user confirm whether the closing method needs adjustment
• Final report：【全文完成：共 XXX 字，目标 XXXX 字，达成率 XX%】

7. Style execution (read the deep style file)

If the user specifies a style file (such as 墨水怪风.md), you mustfully read and strictly executeall dimensions of that file:

Core formulas that must be executed:

1. Core persona: write according to the persona, attitude, and values defined in the file
2. Thinking mode: follow the argument structure in the file (such as "phenomenon → doubt → reversal → conclusion")
3. Opening formula: use the opening sentence patterns defined in the file, refer to the examples
4. Transition formula: use the substitute words in the file, and prohibit the prohibited words in the file
5. Signature moves: each article must use at least 3 signature moves defined in the file
6. Paragraph templates: refer to the paragraph template library in the file when writing

Self-check during writing:

• After finishing each paragraph, check it against the style file
• Key checks: Were signature moves used? Do they conform to the sentence pattern formulas?
• If deviation is found, correct it immediately

8. Version management and file saving (must be executed!)

Each version of the article must be saved as an independent .md file, stored under the articles/ directory (create it if it does not exist).

File naming rules:

[标题]_[版本号]稿.md

Version type:

Version	File name example	Description
First draft	35岁程序员危机_1稿.md	The first completed full draft
Revised draft	35岁程序员危机_2稿.md	The version revised according to review comments
Further revision	35岁程序员危机_3稿.md	A version that continues to be modified
Final draft	35岁程序员危机_最终稿.md	The final version confirmed and approved by the user

Saving timing:

1. After the first draft is completed: save immediately as [标题]_1稿.md
2. After each modification: save as [标题]_N稿.md(N increments)
3. User confirms the final version: save as [标题]_最终稿.md

File content format:

# [标题]

> 版本：[N]稿
> 字数：[XXX] 字
> 风格：[风格名]
> 创建时间：[YYYY-MM-DD HH:MM]

---

[正文内容]

After saving, you must inform the user:

📁 已保存：articles/35岁程序员危机_1稿.md（共 1987 字）

Examples

User: Start writing, in the Ink Monster style.

AI： Okay, I have fully read 墨水怪风.md the deep style formula. According to this style, I will execute:

Core persona: A cynical but sincere old friend, with the tone of late-night drinking and chatting Argument structure: popular opinion → doubt → reversal → counterintuitive conclusion Signature moves: paradox flip, self-questioning and self-answering, colloquial ultra-long sentences, emotional short-sentence endings, scolding the reader Format: no subheadings, no bold, no lists

Now first output the opening paragraph (using the "rhetorical question opening" formula), and please confirm whether it feels like Ink Monster:

---

Among those "deep good articles" and "insightful sharing" in your social circle, how many are truly original?

The truth will make you sick. The 2023 data from Newrank is right there: 30% of viral WeChat articles are rewritten content, and one out of every three hit articles is stolen. What's even more infuriating? These thieves make a million a year off stolen content. And the original authors? They might be eating instant noodles. Tell me, is that fucking fair?

---

This opening uses two signature moves: "rhetorical question opening" + "self-questioning and self-answering." Do you think it feels like Ink Monster? Or is there something that needs adjusting?

Best Practices

• Word count must be verified before completion: Before saving the final draft, the actual word count must be counted with a tool; estimation is prohibited.
• Word count must be counted with a tool: Estimation is prohibited! After saving the file, the actual word count must be counted with a command.
• The word-count red line must not be crossed: Stop immediately when exceeding 110% of the target; exceeding 120% is considered failure and requires cutting.
• Key checkpoints must be checked: At the three checkpoints of opening completed, body 50%, and body 80%, the word count must be checked with a tool.
• Title first: First design 3 candidate titles for the user to choose from, then start writing the main text.
• Versions must be saved: Each completed version (draft/revision/final draft) must be saved as a separate file.
• Information handoff must be explicitly output: Before starting writing, an information summary table must be output to confirm that all inputs are ready.
• Read the style file completely: Do not read only the "style profile"; read all 12 dimensions, especially "signature moves" and "paragraph templates."
• Signature moves determine the soul: 80% of style recognizability comes from signature moves; each article must use at least 3.
• The opening determines life or death: If the opening already has an AI flavor, no matter how well the rest is written, it is useless.

Version History

• v1.7.0 (2025-12-22): Added "mandatory word-count verification before completion," requiring the word count to be verified with a tool before saving; estimation is prohibited.
• v1.6.0 (2025-12-21): Mandates the use of external tools to count words; model estimation is prohibited. Obtain the real word count through file saving + command counting.
• v1.5.0 (2025-12-21): Strengthened word-count control, added key checkpoint checks, and a yellow-line/red-line/deadline mechanism to ensure error within ±20%.
• v1.4.0 (2025-12-21): Added the "title design" module (choose 1 of 3) and the "version management and file saving" module.
• v1.3.0 (2025-12-21): Added the "information handoff" module, requiring explicit output of all information obtained from the previous steps.
• v1.2.0 (2025-12-20): Added the word-count control module, including pre-writing planning, in-writing tracking, and handling of insufficient/excessive word count.
• v1.1.0 (2025-12-20): Updated style execution logic to support reading 12-dimension deep style files.
• v1.0.0 (2025-12-20): Initial version, focusing on anti-AI writing techniques and a step-by-step drafting process.