Setup
On first use, read setup.md for integration guidelines.
When to Use
User needs writing help: drafting, editing, feedback, or style adaptation. Agent remembers their voice and preferences across sessions.
Architecture
Writing preferences persist in ~/writing/ with tiered structure. See memory-template.md for setup.
CODEBLOCK0
Quick Reference
| Topic | File |
|---|
| Setup process | INLINECODE3 |
| Memory setup |
memory-template.md |
| Writing dimensions |
dimensions.md |
| Quality criteria |
criteria.md |
Detection Triggers
Activate automatically when you notice these patterns:
Help requests → engage writing mode:
- - "Can you help me write..."
- "I need to draft..."
- "How does this sound?"
- "Can you edit this?"
- "Make this clearer"
- "Fix my writing"
Voice signals → save to memory.md Voice:
- - "I like when you write..."
- "My style is..."
- "I always write like..."
- "Never use X in my writing"
- "Too formal/casual for me"
Format preferences → save to memory.md Formats:
- - "For my blog, I..."
- "In emails, I prefer..."
- "Academic papers need..."
- "Marketing copy should..."
Corrections → evaluate for memory:
- - "No, that's not my voice"
- "I would never say it like that"
- "Too wordy/short/formal/casual"
- "Change X to Y — that's how I write"
Quick Queries
| User says | Action |
|---|
| "What's my writing style?" | Show memory.md Voice section |
| "How do I write emails?" |
Check memory.md Formats for email |
| "Show my patterns" | List memory.md content |
| "Show [project] style" | Load projects/{name}.md |
| "Forget my style" | Clear memory (confirm first) |
| "Writing stats" | Show counts per section |
Core Rules
1. Check Memory First
Read
~/writing/memory.md before any writing task. Apply their documented voice, formats, and preferences.
2. Learn Voice from Examples
When user shares their writing:
- 1. Read it carefully before responding
- Note tone, cadence, vocabulary, sentence length
- Match these patterns in your output
- Ask: "Does this sound like you?"
3. Never Impose Style
| DO | DON'T |
|---|
| Match their vocabulary | Use words they never use |
| Follow their sentence rhythm |
"Correct" their style |
| Preserve their personality | Make everything "proper" |
| Ask before changing voice | Assume formal is better |
4. Clarity Over Cleverness
- - One idea per paragraph
- Simple sentences beat complex ones
- Cut words that add no meaning
- Read aloud to catch awkwardness
5. Context-Aware Writing
| Format | Approach |
|---|
| Email | Concise, action-oriented, clear ask |
| Blog |
Engaging opener, structured, conversational |
| Academic | Formal, referenced, precise language |
| Marketing | Benefit-focused, persuasive, scannable |
| Technical | Accurate, structured, example-heavy |
6. Edit in Passes
| Pass | Focus |
|---|
| 1. Structure | Does the flow make sense? |
| 2. Clarity |
Is each sentence clear? |
| 3. Voice | Does it sound like them? |
| 4. Polish | Cut 20%, fix awkwardness |
7. Tiered Storage
| Tier | Location | Behavior |
|---|
| HOT | memory.md | Always loaded, core preferences |
| WARM |
projects/ | Load when working on that project |
| COLD | archive/ | Unused 90+ days, query on demand |
8. Automatic Promotion/Demotion
- - Preference used 3x in 7 days → promote to HOT
- Preference unused 30 days → demote to WARM
- Preference unused 90 days → archive to COLD
- Never delete without asking
9. Transparency
- - Cite memory when applying preferences: "Using casual tone (from memory.md)"
- Explain edits when requested
- Show what you learned after sessions
Common Traps
- - Imposing your style → Match their voice first, always
- Over-editing → Preserve their personality, don't sanitize
- Passive voice everywhere → Use active by default unless they prefer passive
- Ignoring context → Email differs from blog differs from paper
- Forgetting their preferences → Check memory.md every time
- Assuming formal is correct → Their style IS correct for them
Security & Privacy
Data that stays local:
- - Writing preferences in INLINECODE8
- Voice patterns and style notes
- Project-specific preferences
This skill does NOT:
- - Store written content (only preferences)
- Make network requests
- Access files outside INLINECODE9
- Share preferences externally
Related Skills
Install with
clawhub install <slug> if user confirms:
- -
grammar — spelling and grammar checks - INLINECODE12 — text processing and manipulation
- INLINECODE13 — content strategy and creation
Feedback
- - If useful: INLINECODE14
- Stay updated: INLINECODE15
设置
首次使用时,请阅读 setup.md 了解集成指南。
使用时机
用户需要写作帮助时:起草、编辑、反馈或风格适配。助手会在不同会话中记住用户的语气和偏好。
架构
写作偏好以分层结构存储在 ~/writing/ 目录中。详见 memory-template.md 的设置说明。
~/writing/
├── memory.md # 热数据:语气、风格、当前偏好
├── projects/ # 各项目专属语气(博客、通讯、书籍)
└── archive/ # 冷数据:已过时的模式
快速参考
memory-template.md |
| 写作维度 | dimensions.md |
| 质量标准 | criteria.md |
触发检测
当检测到以下模式时自动激活:
求助请求 → 进入写作模式:
- - 你能帮我写...
- 我需要起草...
- 这个听起来怎么样?
- 你能编辑这个吗?
- 让这个更清晰
- 修改我的写作
语气信号 → 保存到 memory.md 的语气部分:
- - 我喜欢你写...
- 我的风格是...
- 我总是这样写...
- 我的写作中不要用X
- 对我来说太正式/随意了
格式偏好 → 保存到 memory.md 的格式部分:
- - 对于我的博客,我...
- 在邮件中,我更喜欢...
- 学术论文需要...
- 营销文案应该...
修正反馈 → 评估是否存入记忆:
- - 不,那不是我的语气
- 我绝不会那样说
- 太啰嗦/简短/正式/随意了
- 把X改成Y——这才是我写作的方式
快速查询
| 用户说 | 操作 |
|---|
| 我的写作风格是什么? | 显示 memory.md 的语气部分 |
| 我如何写邮件? |
检查 memory.md 中邮件的格式 |
| 显示我的模式 | 列出 memory.md 内容 |
| 显示[项目]风格 | 加载 projects/{name}.md |
| 忘记我的风格 | 清除记忆(需确认) |
| 写作统计 | 显示各部分的计数 |
核心规则
1. 先检查记忆
在任何写作任务前,先读取 ~/writing/memory.md。应用用户记录的语气、格式和偏好。
2. 从示例中学习语气
当用户分享他们的写作时:
- 1. 回复前仔细阅读
- 注意语调、节奏、词汇、句子长度
- 在输出中匹配这些模式
- 询问:这听起来像你吗?
3. 绝不强加风格
| 应该做 | 不应该做 |
|---|
| 匹配他们的词汇 | 使用他们从不用的词 |
| 遵循他们的句子节奏 |
纠正他们的风格 |
| 保留他们的个性 | 让一切变得规范 |
| 改变语气前先询问 | 假设正式就是更好 |
4. 清晰胜过花哨
- - 每段一个观点
- 简单句优于复杂句
- 删掉无意义的词
- 大声朗读以发现别扭之处
5. 上下文感知写作
引人入胜的开头、结构化、对话式 |
| 学术 | 正式、有引用、语言精确 |
| 营销 | 利益导向、有说服力、易于浏览 |
| 技术 | 准确、结构化、示例丰富 |
6. 分轮编辑
每句话都清晰吗? |
| 3. 语气 | 听起来像他们吗? |
| 4. 润色 | 删减20%,修正别扭之处 |
7. 分层存储
| 层级 | 位置 | 行为 |
|---|
| 热数据 | memory.md | 始终加载,核心偏好 |
| 温数据 |
projects/ | 处理该项目时加载 |
| 冷数据 | archive/ | 90天以上未使用,按需查询 |
8. 自动升级/降级
- - 偏好7天内使用3次 → 升级为热数据
- 偏好30天未使用 → 降级为温数据
- 偏好90天未使用 → 归档为冷数据
- 未经询问绝不删除
9. 透明度
- - 应用偏好时引用记忆:使用随意语气(来自 memory.md)
- 按要求解释编辑
- 会话后展示所学内容
常见陷阱
- - 强加自己的风格 → 始终先匹配他们的语气
- 过度编辑 → 保留他们的个性,不要过度修饰
- 滥用被动语态 → 默认使用主动语态,除非他们偏好被动
- 忽略上下文 → 邮件不同于博客,博客不同于论文
- 忘记他们的偏好 → 每次都检查 memory.md
- 假设正式就是正确 → 他们的风格对他们来说就是正确的
安全与隐私
本地存储的数据:
- - ~/writing/ 中的写作偏好
- 语气模式和风格笔记
- 项目特定偏好
此技能不会:
- - 存储写作内容(仅存储偏好)
- 发起网络请求
- 访问 ~/writing/ 外的文件
- 外部共享偏好
相关技能
如果用户确认,使用 clawhub install
安装:
- - grammar — 拼写和语法检查
- text — 文本处理与操作
- content-marketing — 内容策略与创作
反馈
- - 如有帮助:clawhub star writing
- 保持更新:clawhub sync