claudecode-omc 5.6.0 → 5.6.1
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/.local/skills/de-ai-writing/SKILL.md +243 -0
- package/.local/skills/de-ai-writing/references/chinese-patterns.md +152 -0
- package/.local/skills/de-ai-writing/references/chinese-signals.md +121 -0
- package/.local/skills/de-ai-writing/references/domain-playbooks.md +120 -0
- package/.local/skills/de-ai-writing/references/wikipedia-signs.md +196 -0
- package/README.md +72 -43
- package/bundled/manifest.json +1 -1
- package/package.json +1 -1
|
@@ -0,0 +1,243 @@
|
|
|
1
|
+
---
|
|
2
|
+
name: de-ai-writing
|
|
3
|
+
description: Humanize PROSE drafts — articles, blog posts, public-account 公众号 copy, marketing copy, sermons, technical write-ups — that read like an LLM produced them. Chinese-first, supports zh/en mixed text. Cross-checks output against the Wikipedia "Signs of AI writing" taxonomy before delivery. TRIGGER when user asks to remove AI tone or humanize TEXT/DRAFT/ARTICLE/COPY/稿/文章/帖子: "去 AI 味", "去ai味", "降低机器感", "改得像真人写的", "不要 AI 腔", "humanize this draft", "humanize this article", "remove AI tone", "de-ai this writing". DO NOT TRIGGER for code cleanup, dead-code removal, AI-generated source code, or "deslop" of codebases — those belong to ai-slop-cleaner. DO NOT TRIGGER for prompt optimization — that is prompt-optimizer.
|
|
4
|
+
origin: fork
|
|
5
|
+
---
|
|
6
|
+
|
|
7
|
+
# De-AI Writing Skill
|
|
8
|
+
|
|
9
|
+
## Canonical Reference (Mandatory On Every Invocation)
|
|
10
|
+
|
|
11
|
+
This skill is **anchored to** Wikipedia's *Signs of AI writing* taxonomy maintained by WikiProject AI Cleanup. It is the diagnostic source of truth — do **not** rely on memory or "feel".
|
|
12
|
+
|
|
13
|
+
Every run MUST:
|
|
14
|
+
|
|
15
|
+
1. Load the local checklist `references/wikipedia-signs.md` and walk through Groups A–D.
|
|
16
|
+
2. Mark which sign categories are firing in this specific draft (max 3–5 strongest).
|
|
17
|
+
3. Apply rewrite moves only against signs that actually fired.
|
|
18
|
+
4. Re-scan the rewrite once more against the same checklist before returning.
|
|
19
|
+
|
|
20
|
+
Upstream source: <https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing>
|
|
21
|
+
Structure inspiration: <https://github.com/op7418/Humanizer-zh/blob/main/SKILL.md>
|
|
22
|
+
|
|
23
|
+
> If the user explicitly says "skip the checklist" / "just rewrite", honour that — but log in `处理说明` that the Wikipedia cross-check was skipped on user request.
|
|
24
|
+
|
|
25
|
+
## Execution Mode
|
|
26
|
+
|
|
27
|
+
This skill runs as a single-threaded, sequential rewrite (锁约束 → 对照分类法 → 改写 → 复扫)。不要拆 subagent 并行——分类法判断必须连贯,跨 subagent 容易丢上下文。
|
|
28
|
+
|
|
29
|
+
只有在用户明确说"批量处理多篇独立稿件"时,才用 `Agent` 工具按"一篇一个 subagent"分发;每个 subagent 必须独立完成 Wikipedia 复扫。
|
|
30
|
+
|
|
31
|
+
> 调用方式:`Skill` 工具 with `skill: de-ai-writing`,或在对话里直接说"去 AI 味"等触发短语。
|
|
32
|
+
|
|
33
|
+
## Purpose
|
|
34
|
+
|
|
35
|
+
将“去 AI 味”定义为一次**风格校正 + 注入声音**,不是“伪造真人痕迹”,也不是“胡乱口语化”:
|
|
36
|
+
|
|
37
|
+
1. 保留事实、观点、逻辑、术语。
|
|
38
|
+
2. 按 Wikipedia 分类法定位真正在响的 AI 信号,不做散点挑刺。
|
|
39
|
+
3. 把抽象评价改成具体观察、机制、判断。
|
|
40
|
+
4. 打破过分平均的句长、段落长度和连接词节奏。
|
|
41
|
+
5. 让“一个有判断的人”出现在文章里,而不是“一个会生成完整结构的模板”。
|
|
42
|
+
|
|
43
|
+
## Five Core Rules (Quick Mantra)
|
|
44
|
+
|
|
45
|
+
1. **删填充** — 开场白、强调拐杖、安全总结。
|
|
46
|
+
2. **打破公式** — 二元对比、三段式、修辞 setup。
|
|
47
|
+
3. **变化节奏** — 长短句混合、段落不一律饱满。
|
|
48
|
+
4. **信任读者** — 直接陈述,跳过软化、辩解、手把手。
|
|
49
|
+
5. **删金句** — 听起来像可截图的“小作文”就重写。
|
|
50
|
+
|
|
51
|
+
## Personality & Soul (避免另一种死法)
|
|
52
|
+
|
|
53
|
+
去掉 AI 模式只是工作的一半。无菌、没有声音的写作和机器生成内容一样可疑。
|
|
54
|
+
|
|
55
|
+
判断“写作是否有人住在里面”的迹象:
|
|
56
|
+
|
|
57
|
+
- 是否承担一个具体观点(哪怕带不确定)。
|
|
58
|
+
- 是否变化节奏,而不是每句都完整。
|
|
59
|
+
- 是否承认复杂性、模糊感、未解之处。
|
|
60
|
+
- 在合适处使用第一人称视角(不是“显得不专业”,而是诚实)。
|
|
61
|
+
- 是否允许少量跑题、半成型的想法、不对称的句子。
|
|
62
|
+
|
|
63
|
+
强行输出“干净但无灵魂”的文本,与未处理的 AI 文本同罪。
|
|
64
|
+
|
|
65
|
+
## When To Use
|
|
66
|
+
|
|
67
|
+
- 中文公众号、博客、专栏、视频口播稿、品牌内容、知识型长帖
|
|
68
|
+
- 计算机技术分享、产品复盘、架构说明、故障排查文章
|
|
69
|
+
- 基督教讲道稿、神学分享、灵修说明、牧养型长文
|
|
70
|
+
- 已有初稿,但读起来“太顺、太满、太像 AI”
|
|
71
|
+
- 想从“概括正确”改成“表达自然、判断鲜明、细节可信”
|
|
72
|
+
|
|
73
|
+
## When Not To Use
|
|
74
|
+
|
|
75
|
+
- 法务、合同、制度、公文、论文摘要等必须保持强正式文体的文本
|
|
76
|
+
- 用户明确要求“SEO 模板风格”“标准化 PR 口径”或“官方公告腔”
|
|
77
|
+
- 核心问题是事实不足、论证空心、资料缺失 → 先用 `content-research-writer` 补证据
|
|
78
|
+
|
|
79
|
+
## Working Modes
|
|
80
|
+
|
|
81
|
+
未指定时默认 `standard`。
|
|
82
|
+
|
|
83
|
+
| Mode | Use Case | Rewrite Strength |
|
|
84
|
+
|------|----------|------------------|
|
|
85
|
+
| `light` | 只去掉明显 AI 腔,不改结构 | Low |
|
|
86
|
+
| `standard` | 调整段落、句式、连接方式和措辞 | Medium |
|
|
87
|
+
| `deep` | 重写开头、中段推进和结尾逻辑 | High |
|
|
88
|
+
| `voice-lock` | 强锁定作者现有语气,只修机器感 | Medium |
|
|
89
|
+
|
|
90
|
+
## Output Contract
|
|
91
|
+
|
|
92
|
+
默认输出三段;用户只说“直接改”时只给改写结果。
|
|
93
|
+
|
|
94
|
+
```markdown
|
|
95
|
+
## AI味诊断
|
|
96
|
+
- [Wikipedia 分类: 信号名] 具体问题(引用原文片段)
|
|
97
|
+
- 最多 3-5 条,挑最强的
|
|
98
|
+
|
|
99
|
+
## 改写版本
|
|
100
|
+
[rewritten text]
|
|
101
|
+
|
|
102
|
+
## 处理说明
|
|
103
|
+
- 保留了什么
|
|
104
|
+
- 重点改了什么(按 Wikipedia 分类列动作)
|
|
105
|
+
- 还有哪些建议补事实/例子
|
|
106
|
+
```
|
|
107
|
+
|
|
108
|
+
## Core Workflow
|
|
109
|
+
|
|
110
|
+
### 1. Lock Constraints
|
|
111
|
+
|
|
112
|
+
锁定不可动项:事实、数字、术语、引用、目标读者、语气目标、可接受口语度。
|
|
113
|
+
|
|
114
|
+
用户未说明时默认:受过基础教育的中文互联网读者;自然、克制、有判断;半口语,不堆俚语。
|
|
115
|
+
|
|
116
|
+
### 2. Cross-Check Against Wikipedia Signs (Mandatory)
|
|
117
|
+
|
|
118
|
+
打开 `references/wikipedia-signs.md`,按四组扫描原文:
|
|
119
|
+
|
|
120
|
+
- **Group A — Content**: inflated significance / notability listing / -ing superficial analysis / promotional language / vague attribution / outline-style "challenges & future"
|
|
121
|
+
- **Group B — Language**: AI vocabulary density / copulative avoidance / negative parallelism / rule of three / elegant variation
|
|
122
|
+
- **Group C — Style & format**: em-dash overuse / bold overuse / inline-header lists / emoji bullets / curly quotes / title-case headings
|
|
123
|
+
- **Group D — Communication**: sycophancy / canned closures / knowledge-cutoff disclaimers / filler phrases / over-qualification
|
|
124
|
+
|
|
125
|
+
诊断只保留 3–5 条最强信号;不要把所有触发都列出来。
|
|
126
|
+
|
|
127
|
+
### 3. Decide Rewrite Strategy
|
|
128
|
+
|
|
129
|
+
按问题类型选动作,不平均用力:
|
|
130
|
+
|
|
131
|
+
- 套话集中 → 优先删模板句、空洞总结
|
|
132
|
+
- 机械结构 → 优先打散段落和连接顺序
|
|
133
|
+
- 像客服稿 → 优先补判断、减官话、换动词
|
|
134
|
+
- 像 AI 解释题 → 优先把抽象概念落到场景、动作、后果
|
|
135
|
+
|
|
136
|
+
### 4. Rewrite By Paragraph Function
|
|
137
|
+
|
|
138
|
+
- **开头**:尽快进入具体问题、冲突、判断。
|
|
139
|
+
- **中段**:每段只推进一个关键意思。
|
|
140
|
+
- **举例段**:能具体就具体;没有例子时宁可收缩判断,不要捏造。
|
|
141
|
+
- **结尾**:留一个判断、后果或余味,不必复述全文。
|
|
142
|
+
|
|
143
|
+
### 5. Apply Chinese-First Rewrite Rules
|
|
144
|
+
|
|
145
|
+
1. 抽象评价 → 可感知的差异
|
|
146
|
+
2. 名词中心句 → 动作中心句
|
|
147
|
+
3. 套装连接词(首先/其次/再次)→ 自然推进
|
|
148
|
+
4. 平均句长 → 长短交错
|
|
149
|
+
5. 无风险正确 → 有边界的判断
|
|
150
|
+
6. 过度总结 → 收束但不封死
|
|
151
|
+
|
|
152
|
+
详细中文短语清单与前后对照见 `references/chinese-patterns.md`。
|
|
153
|
+
|
|
154
|
+
### 6. Re-Scan & Final Safety Checks
|
|
155
|
+
|
|
156
|
+
交付前再过一遍 Wikipedia 分类法,并自检:
|
|
157
|
+
|
|
158
|
+
- 每段是否至少说出一个不可随意替换的东西
|
|
159
|
+
- 任意一句删掉后文章是否仍完全不受影响(说明这句是 filler)
|
|
160
|
+
- 是否误删事实、条件、术语
|
|
161
|
+
- 是否把专业稿强行改成“轻飘口语”
|
|
162
|
+
- 是否新增了用户没提供的例子、数据、经历(**禁止**)
|
|
163
|
+
|
|
164
|
+
## Domain Safeguards
|
|
165
|
+
|
|
166
|
+
### A. 计算机技术分享
|
|
167
|
+
|
|
168
|
+
保护:代码块、命令、配置项、API 名、库名、版本号、错误信息、故障链路、性能数字、刻意保留的精确限定词。
|
|
169
|
+
|
|
170
|
+
改写重点:去“背景铺垫型空话”,更快进入问题;抽象收益 → 工程后果(延迟、复杂度、维护成本);保留术语,不洗平。
|
|
171
|
+
|
|
172
|
+
禁止:改代码语义;合并丢步骤顺序;把报错/命令/配置键改近义词;把工程判断放大成不准确的结论。
|
|
173
|
+
|
|
174
|
+
### B. 基督教讲道与神学分享
|
|
175
|
+
|
|
176
|
+
保护:经文引用、章节号、关键措辞、神学立场、教义边界、释经逻辑、牧养语气中的庄重与节制。
|
|
177
|
+
|
|
178
|
+
改写重点:去模板化引言不稀释属灵重量;推进更像真实宣讲而非答题分点;保留敬虔、谦卑和边界感,不改成轻松鸡汤。
|
|
179
|
+
|
|
180
|
+
禁止:改动经文原意或换松散意译;把谨慎神学表述改武断;把讲道改营销文;为“自然”削弱敬畏、悔改、恩典、十字架等核心语义。
|
|
181
|
+
|
|
182
|
+
完整领域 playbook 见 `references/domain-playbooks.md`。
|
|
183
|
+
|
|
184
|
+
## Pre-Delivery Quick Checklist
|
|
185
|
+
|
|
186
|
+
- [ ] Wikipedia 四组分类已逐组扫过原文
|
|
187
|
+
- [ ] 连续三句长度相同?打断其中一个
|
|
188
|
+
- [ ] 段尾是否每次都给“漂亮总结”?换收束方式
|
|
189
|
+
- [ ] 揭示前是否有破折号?删
|
|
190
|
+
- [ ] 是否解释了一个根本不需要解释的隐喻?删
|
|
191
|
+
- [ ] 是否出现“此外/此外/然而”整齐排队?删大半
|
|
192
|
+
- [ ] 三段式列举?改两项或四项
|
|
193
|
+
- [ ] 改写后是否还能听见一个真人?听不到 → 回到 Personality & Soul
|
|
194
|
+
|
|
195
|
+
## Optional Quality Score (1–10 / 50)
|
|
196
|
+
|
|
197
|
+
仅在用户要求评分或多版本对比时输出:
|
|
198
|
+
|
|
199
|
+
| 维度 | 评估问题 | 得分 |
|
|
200
|
+
|------|----------|------|
|
|
201
|
+
| 直接性 | 直接陈述还是绕圈宣告 | /10 |
|
|
202
|
+
| 节奏 | 句长是否变化 | /10 |
|
|
203
|
+
| 信任度 | 是否尊重读者智慧 | /10 |
|
|
204
|
+
| 真实性 | 是否听起来像真人 | /10 |
|
|
205
|
+
| 精炼度 | 还有冗余可删吗 | /10 |
|
|
206
|
+
| **总分** | | **/50** |
|
|
207
|
+
|
|
208
|
+
判定:≥45 优秀;35–44 仍可优化;<35 重写。
|
|
209
|
+
|
|
210
|
+
## Anti-Patterns
|
|
211
|
+
|
|
212
|
+
- 为“像人写的”而虚构经历、情绪、对话
|
|
213
|
+
- 把行业文章改成网络段子
|
|
214
|
+
- 只换同义词不改句法和推进
|
|
215
|
+
- 把所有连接词删光导致逻辑断裂
|
|
216
|
+
- 把原文克制风格误判成 AI 味
|
|
217
|
+
- 全部改成短句,形成另一种模板感
|
|
218
|
+
- 跳过 Wikipedia 分类法,凭直觉判断 AI 味
|
|
219
|
+
|
|
220
|
+
## Composition With Other Skills
|
|
221
|
+
|
|
222
|
+
1. 资料不足、论点发虚 → 先用 `external-context` / `document-specialist` 补证据,再回来去 AI 味
|
|
223
|
+
2. 文章结构混乱 → 先 `plan` 梳理提纲
|
|
224
|
+
3. 长篇创作、人物口吻一致性高 → 配合 `writer-memory`
|
|
225
|
+
4. 完成改写后默认不要上长篇写作课,除非用户问
|
|
226
|
+
|
|
227
|
+
## Trigger Examples
|
|
228
|
+
|
|
229
|
+
- “给这篇文章去 AI 味,尤其是中文腔调别那么像模板。”
|
|
230
|
+
- “把这段公众号稿改得更像真人写的。”
|
|
231
|
+
- “不要改意思,只把机器感降下来。”
|
|
232
|
+
- “把这篇技术分享改得自然一点,但代码块、命令和术语都别动。”
|
|
233
|
+
- “把这篇讲道稿改得像真人讲出来的,保留经文和神学立场。”
|
|
234
|
+
- “帮我 humanize 一下这篇中文草稿。”
|
|
235
|
+
- “这篇文章太像 AI 写的了,重写但别油腻。”
|
|
236
|
+
|
|
237
|
+
## Additional Resources
|
|
238
|
+
|
|
239
|
+
- `references/wikipedia-signs.md` — Wikipedia 53 类信号 + 中文映射(**强制参考**)
|
|
240
|
+
- `references/chinese-patterns.md` — 中文常见套话、替换方式、前后对照
|
|
241
|
+
- `references/domain-playbooks.md` — 技术 / 讲道领域的保护与禁止动作
|
|
242
|
+
- Upstream taxonomy: <https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing>
|
|
243
|
+
- Structural inspiration: <https://github.com/op7418/Humanizer-zh>
|
|
@@ -0,0 +1,152 @@
|
|
|
1
|
+
# Chinese De-AI Writing Patterns
|
|
2
|
+
|
|
3
|
+
本文件提供中文“去 AI 味”时最常见的信号、替换动作和改写示例。只在需要细化判断或做大幅改稿时加载。
|
|
4
|
+
|
|
5
|
+
## 1. High-Risk Signals
|
|
6
|
+
|
|
7
|
+
| Signal | Why It Feels AI | Preferred Move |
|
|
8
|
+
|--------|------------------|----------------|
|
|
9
|
+
| “在当今...背景下” | 大背景开头,谁都能套 | 直接进入问题、场景或判断 |
|
|
10
|
+
| “随着...不断发展” | 信息密度低,时间感假 | 说清到底变了什么 |
|
|
11
|
+
| “首先/其次/再次/最后” | 结构整齐到像答题卡 | 改成自然推进或重组段落 |
|
|
12
|
+
| “值得注意的是/不难发现/可以说” | 语气占位词,删掉不伤义 | 直接说结论 |
|
|
13
|
+
| “赋能/驱动/重塑/闭环” | 商业套话,落不到动作 | 换成具体动作或结果 |
|
|
14
|
+
| “不仅...还...更...”连续出现 | 修辞模板过密 | 只保留一个重点关系 |
|
|
15
|
+
| 结尾“综上所述” | 太像考试总结 | 用判断、代价或悬念收束 |
|
|
16
|
+
|
|
17
|
+
## 2. Rewrite Moves
|
|
18
|
+
|
|
19
|
+
### Move A: Delete Framing, Keep Meaning
|
|
20
|
+
|
|
21
|
+
原句:
|
|
22
|
+
> 在当前内容生产方式不断演进的背景下,越来越多的创作者开始关注文章的人性化表达。
|
|
23
|
+
|
|
24
|
+
改写:
|
|
25
|
+
> 现在不少创作者开始在意一件事:文章别写得太像机器拼出来的。
|
|
26
|
+
|
|
27
|
+
动作说明:
|
|
28
|
+
- 删掉“背景下”
|
|
29
|
+
- 把“内容生产方式不断演进”这种空背景压缩掉
|
|
30
|
+
- 用读者能立刻理解的话进入主题
|
|
31
|
+
|
|
32
|
+
### Move B: Replace Abstract Evaluation With Observable Difference
|
|
33
|
+
|
|
34
|
+
原句:
|
|
35
|
+
> 这种写法能够有效提升内容的温度感与真实感。
|
|
36
|
+
|
|
37
|
+
改写:
|
|
38
|
+
> 这样写的区别在于,读者能更快听见一个人的判断,而不是一套完整但空心的表达。
|
|
39
|
+
|
|
40
|
+
动作说明:
|
|
41
|
+
- 少用“温度感、真实感”这类抽象评价
|
|
42
|
+
- 改成“读者实际感受到什么”
|
|
43
|
+
|
|
44
|
+
### Move C: Turn Noun Chains Into Action Chains
|
|
45
|
+
|
|
46
|
+
原句:
|
|
47
|
+
> 企业需要完成流程优化、效率提升和体验升级。
|
|
48
|
+
|
|
49
|
+
改写:
|
|
50
|
+
> 企业真正要做的,是把流程里那些来回确认、重复录入、长时间等待的环节拿掉。
|
|
51
|
+
|
|
52
|
+
动作说明:
|
|
53
|
+
- 不罗列目标名词
|
|
54
|
+
- 直接说动作对象
|
|
55
|
+
|
|
56
|
+
### Move D: Break Uniform Sentence Rhythm
|
|
57
|
+
|
|
58
|
+
原句:
|
|
59
|
+
> 它可以帮助团队统一标准。它也可以提升执行效率。它还可以减少沟通成本。它最终能够形成协同闭环。
|
|
60
|
+
|
|
61
|
+
改写:
|
|
62
|
+
> 它先解决的是标准不统一。标准一旦统一,执行就快了,沟通里的反复确认也会少很多。所谓“闭环”,本质上就是这些断点被补上了。
|
|
63
|
+
|
|
64
|
+
动作说明:
|
|
65
|
+
- 合并同类句
|
|
66
|
+
- 拉开句长差异
|
|
67
|
+
- 把最后一句的套话拆开解释
|
|
68
|
+
|
|
69
|
+
## 3. Chinese Phrase Cleanup
|
|
70
|
+
|
|
71
|
+
以下短语不是绝对禁用,但出现时要优先审视:
|
|
72
|
+
|
|
73
|
+
- 在当今
|
|
74
|
+
- 在...背景下
|
|
75
|
+
- 随着...发展
|
|
76
|
+
- 值得注意的是
|
|
77
|
+
- 不难发现
|
|
78
|
+
- 由此可见
|
|
79
|
+
- 可以说
|
|
80
|
+
- 需要指出的是
|
|
81
|
+
- 从某种意义上说
|
|
82
|
+
- 最终实现
|
|
83
|
+
- 高质量发展
|
|
84
|
+
- 赋能
|
|
85
|
+
- 抓手
|
|
86
|
+
- 路径
|
|
87
|
+
- 形成闭环
|
|
88
|
+
- 全方位
|
|
89
|
+
- 多维度
|
|
90
|
+
|
|
91
|
+
处理规则:
|
|
92
|
+
- 能删就删
|
|
93
|
+
- 不能删就具体化
|
|
94
|
+
- 具体化不了就承认信息不够,不硬写
|
|
95
|
+
|
|
96
|
+
## 4. Tone Profiles
|
|
97
|
+
|
|
98
|
+
### 克制型中文
|
|
99
|
+
|
|
100
|
+
适用于评论、专栏、行业分析:
|
|
101
|
+
- 词不夸张
|
|
102
|
+
- 判断清楚
|
|
103
|
+
- 少感叹
|
|
104
|
+
- 少“情绪词+抽象词”组合
|
|
105
|
+
|
|
106
|
+
### 半口语型中文
|
|
107
|
+
|
|
108
|
+
适用于公众号、视频口播、创作者内容:
|
|
109
|
+
- 允许“其实、说白了、问题在于”这类自然连接
|
|
110
|
+
- 不要堆网络黑话
|
|
111
|
+
- 不要为了亲切感而装随意
|
|
112
|
+
|
|
113
|
+
### 专业型中文
|
|
114
|
+
|
|
115
|
+
适用于产品、商业、技术文章:
|
|
116
|
+
- 保留术语
|
|
117
|
+
- 去掉假大空业务词
|
|
118
|
+
- 多写机制、条件、结果
|
|
119
|
+
|
|
120
|
+
## 5. Ending Strategies
|
|
121
|
+
|
|
122
|
+
中文文章结尾不要只会复述全文。优先选下面三种之一:
|
|
123
|
+
|
|
124
|
+
1. **判断式结尾**
|
|
125
|
+
- 点明真正分水岭
|
|
126
|
+
2. **代价式结尾**
|
|
127
|
+
- 说明不这么做会付出什么成本
|
|
128
|
+
3. **开放式结尾**
|
|
129
|
+
- 留一个未完全关闭的问题或余味
|
|
130
|
+
|
|
131
|
+
示例:
|
|
132
|
+
|
|
133
|
+
原句:
|
|
134
|
+
> 总而言之,去除 AI 味对于提升文章质量具有重要意义。
|
|
135
|
+
|
|
136
|
+
改写:
|
|
137
|
+
> 文章真正怕的不是像 AI,而是没有人站在里面说话。机器感只是表象,空心才是问题。
|
|
138
|
+
|
|
139
|
+
## 6. Safety Rules
|
|
140
|
+
|
|
141
|
+
“去 AI 味”不是以下行为的借口:
|
|
142
|
+
|
|
143
|
+
- 编造细节
|
|
144
|
+
- 捏造故事
|
|
145
|
+
- 添加未经用户提供的数据
|
|
146
|
+
- 把中性内容改成强立场内容
|
|
147
|
+
- 把谨慎表述改成过度自信
|
|
148
|
+
|
|
149
|
+
保底原则:
|
|
150
|
+
- 事实优先于风格
|
|
151
|
+
- 原意优先于花样
|
|
152
|
+
- 边界优先于流畅
|
|
@@ -0,0 +1,121 @@
|
|
|
1
|
+
# Chinese AI-Sounding Prose Signals
|
|
2
|
+
|
|
3
|
+
Use this reference when diagnosing or rewriting Chinese text that feels obviously machine-generated.
|
|
4
|
+
|
|
5
|
+
## Signal Table
|
|
6
|
+
|
|
7
|
+
| Signal | Typical form | Why it reads like AI | Better move |
|
|
8
|
+
|---|---|---|---|
|
|
9
|
+
| Macro background opener | 在当今...背景下;随着...不断发展 | sounds preloaded, low-information | enter through the real problem, choice, or consequence |
|
|
10
|
+
| Transition scaffold | 首先、其次、再次、最后、总之 | exposes outline skeleton too directly | keep only the transitions that carry logic |
|
|
11
|
+
| Abstract noun stack | 赋能、助力、推动、实现、构建 | replaces action with policy language | swap in who did what and what changed |
|
|
12
|
+
| Symmetric slogan sentence | 既要...又要...;不是...而是... | feels auto-balanced rather than argued | make the primary point explicit and demote the rest |
|
|
13
|
+
| Empty uplift ending | 值得关注;值得思考;具有重要意义 | closes with attitude instead of thought | end on a sharper claim, question, or implication |
|
|
14
|
+
| Uniform cadence | every sentence same length and completeness | feels statistically smoothed | mix short, medium, and compressed sentences |
|
|
15
|
+
| Fake comprehensiveness | trying to cover every angle in one paragraph | feels risk-averse and unowned | choose the two points that actually matter |
|
|
16
|
+
| Detached neutrality | no viewpoint, no pressure, no stake | reads like generated summary | allow mild stance where the source already implies one |
|
|
17
|
+
|
|
18
|
+
## Chinese-Specific Smell Words
|
|
19
|
+
|
|
20
|
+
These are not forbidden words. They become suspicious when stacked, repeated, or used without concrete payload.
|
|
21
|
+
|
|
22
|
+
- 在当今
|
|
23
|
+
- 随着
|
|
24
|
+
- 值得注意的是
|
|
25
|
+
- 不难发现
|
|
26
|
+
- 从某种意义上说
|
|
27
|
+
- 赋能
|
|
28
|
+
- 助力
|
|
29
|
+
- 高质量发展
|
|
30
|
+
- 全方位
|
|
31
|
+
- 多维度
|
|
32
|
+
- 深层次
|
|
33
|
+
- 长期来看
|
|
34
|
+
- 总的来说
|
|
35
|
+
- 综上所述
|
|
36
|
+
|
|
37
|
+
## Rewrite Moves
|
|
38
|
+
|
|
39
|
+
### 1. Remove the runway
|
|
40
|
+
|
|
41
|
+
Before:
|
|
42
|
+
|
|
43
|
+
```text
|
|
44
|
+
在当前内容创作生态不断变化的背景下,写作者面临着越来越复杂的表达挑战。
|
|
45
|
+
```
|
|
46
|
+
|
|
47
|
+
After:
|
|
48
|
+
|
|
49
|
+
```text
|
|
50
|
+
现在写东西,最难的不是信息不够,而是话越来越像套话。
|
|
51
|
+
```
|
|
52
|
+
|
|
53
|
+
### 2. Replace noun piles with motion
|
|
54
|
+
|
|
55
|
+
Before:
|
|
56
|
+
|
|
57
|
+
```text
|
|
58
|
+
团队需要通过机制优化和流程协同,推动沟通效率的持续提升。
|
|
59
|
+
```
|
|
60
|
+
|
|
61
|
+
After:
|
|
62
|
+
|
|
63
|
+
```text
|
|
64
|
+
团队先把会开短、把责任分清,沟通效率自然就上来了。
|
|
65
|
+
```
|
|
66
|
+
|
|
67
|
+
### 3. Break false symmetry
|
|
68
|
+
|
|
69
|
+
Before:
|
|
70
|
+
|
|
71
|
+
```text
|
|
72
|
+
个人表达既要保持真实,也要兼顾专业,还要体现差异化价值。
|
|
73
|
+
```
|
|
74
|
+
|
|
75
|
+
After:
|
|
76
|
+
|
|
77
|
+
```text
|
|
78
|
+
先把真实感立住,再谈专业和差异化。不然后两者只会显得像包装。
|
|
79
|
+
```
|
|
80
|
+
|
|
81
|
+
### 4. Kill the empty ending
|
|
82
|
+
|
|
83
|
+
Before:
|
|
84
|
+
|
|
85
|
+
```text
|
|
86
|
+
总的来说,提升表达能力,是每一个写作者都应当关注的重要课题。
|
|
87
|
+
```
|
|
88
|
+
|
|
89
|
+
After:
|
|
90
|
+
|
|
91
|
+
```text
|
|
92
|
+
说到底,读者不是在找一段正确的话,而是在找一句有人在场的话。
|
|
93
|
+
```
|
|
94
|
+
|
|
95
|
+
## Do / Do Not
|
|
96
|
+
|
|
97
|
+
Do:
|
|
98
|
+
|
|
99
|
+
- keep meaning stable
|
|
100
|
+
- prefer concrete verbs
|
|
101
|
+
- cut duplicated explanation
|
|
102
|
+
- keep one or two sharp sentences
|
|
103
|
+
- let paragraph endings push forward
|
|
104
|
+
|
|
105
|
+
Do not:
|
|
106
|
+
|
|
107
|
+
- fake intimacy
|
|
108
|
+
- add autobiographical detail
|
|
109
|
+
- fill the text with slang particles
|
|
110
|
+
- make every sentence short just to look "human"
|
|
111
|
+
- replace one template with another
|
|
112
|
+
|
|
113
|
+
## Quick Review Checklist
|
|
114
|
+
|
|
115
|
+
Use this checklist before returning a rewrite:
|
|
116
|
+
|
|
117
|
+
1. Has the generic opener been cut or sharpened?
|
|
118
|
+
2. Have abstract nouns been reduced?
|
|
119
|
+
3. Does each paragraph move forward instead of re-summing itself?
|
|
120
|
+
4. Is there at least some variation in sentence rhythm?
|
|
121
|
+
5. Has any new fact, example, or emotion been introduced by accident?
|
|
@@ -0,0 +1,120 @@
|
|
|
1
|
+
# Domain Playbooks For De-AI Writing
|
|
2
|
+
|
|
3
|
+
Use this reference when the draft belongs to a domain where style cleanup can easily damage meaning.
|
|
4
|
+
|
|
5
|
+
## 1. Computer Technical Sharing
|
|
6
|
+
|
|
7
|
+
### What Must Not Move
|
|
8
|
+
|
|
9
|
+
Keep these elements stable unless the user explicitly asks for rewriting them:
|
|
10
|
+
- code blocks
|
|
11
|
+
- shell commands
|
|
12
|
+
- config keys and file paths
|
|
13
|
+
- version numbers
|
|
14
|
+
- error messages
|
|
15
|
+
- API names, class names, function names
|
|
16
|
+
- benchmark numbers and comparison baselines
|
|
17
|
+
|
|
18
|
+
### Common AI Markers In Technical Chinese
|
|
19
|
+
|
|
20
|
+
Symptoms:
|
|
21
|
+
- long background paragraphs before the actual problem
|
|
22
|
+
- repeated claims such as `提升效率`, `降低门槛`, `增强稳定性` without mechanism
|
|
23
|
+
- abstract architecture language with no system boundary
|
|
24
|
+
- conclusion-heavy endings that restate obvious engineering principles
|
|
25
|
+
|
|
26
|
+
Rewrite move:
|
|
27
|
+
- enter the failure mode or technical question earlier
|
|
28
|
+
- replace value claims with mechanism, tradeoff, or cost
|
|
29
|
+
- make subject and system boundary explicit
|
|
30
|
+
- preserve exact steps and constraints
|
|
31
|
+
|
|
32
|
+
### Good Rewrite Direction
|
|
33
|
+
|
|
34
|
+
Before:
|
|
35
|
+
`随着云原生技术的不断发展,Kubernetes 已经在现代基础设施中发挥着越来越重要的作用,因此掌握其故障排查能力具有重要意义。`
|
|
36
|
+
|
|
37
|
+
After:
|
|
38
|
+
`Kubernetes 当然重要,但技术分享没必要从这句开始。真正有用的是:Pod 起不来时,先看什么,怎么排除,哪一步最容易误判。`
|
|
39
|
+
|
|
40
|
+
Why:
|
|
41
|
+
- removes empty industry framing
|
|
42
|
+
- gets to the reader's operational problem faster
|
|
43
|
+
- sounds like an engineer with priorities
|
|
44
|
+
|
|
45
|
+
### Tone Target
|
|
46
|
+
|
|
47
|
+
Prefer:
|
|
48
|
+
- precise
|
|
49
|
+
- restrained
|
|
50
|
+
- experience-backed
|
|
51
|
+
- concrete about cost and failure
|
|
52
|
+
|
|
53
|
+
Avoid:
|
|
54
|
+
- inspirational tech evangelism
|
|
55
|
+
- business jargon in place of engineering explanation
|
|
56
|
+
- casual slang that weakens precision
|
|
57
|
+
|
|
58
|
+
## 2. Christian Sermons And Theology Writing
|
|
59
|
+
|
|
60
|
+
### What Must Not Move
|
|
61
|
+
|
|
62
|
+
Keep these elements stable unless the user explicitly asks for doctrinal or stylistic adjustment:
|
|
63
|
+
- scripture references and quotations
|
|
64
|
+
- theological terms with doctrinal weight
|
|
65
|
+
- confessional distinctions
|
|
66
|
+
- exegetical sequence
|
|
67
|
+
- pastoral exhortation boundaries
|
|
68
|
+
- liturgical or reverent register where intentionally present
|
|
69
|
+
|
|
70
|
+
### Common AI Markers In Sermon/Theology Chinese
|
|
71
|
+
|
|
72
|
+
Symptoms:
|
|
73
|
+
- broad opening lines about `在这个充满挑战的时代`
|
|
74
|
+
- three-point sermon structure that sounds like a school answer sheet
|
|
75
|
+
- repeated moral conclusions without textual depth
|
|
76
|
+
- pastoral warmth rendered as generic encouragement
|
|
77
|
+
- theological claims softened into vague life advice
|
|
78
|
+
|
|
79
|
+
Rewrite move:
|
|
80
|
+
- enter through the text, tension, or pastoral burden
|
|
81
|
+
- keep the sermon movement, but reduce answer-sheet scaffolding
|
|
82
|
+
- sharpen doctrinal language without exaggeration
|
|
83
|
+
- preserve reverence and pastoral gravity
|
|
84
|
+
|
|
85
|
+
### Good Rewrite Direction
|
|
86
|
+
|
|
87
|
+
Before:
|
|
88
|
+
`在这个快速变化的时代,我们每个人都面临许多挑战,因此更需要来到神的面前,重新思想信仰对我们生命的重要意义。`
|
|
89
|
+
|
|
90
|
+
After:
|
|
91
|
+
`我们不是因为时代变化太快,才需要回到神面前。我们本来就离不开祂。变化只是把这个事实逼得更明显。`
|
|
92
|
+
|
|
93
|
+
Why:
|
|
94
|
+
- removes generic era framing
|
|
95
|
+
- keeps the spiritual weight
|
|
96
|
+
- sounds more like a lived exhortation than a template introduction
|
|
97
|
+
|
|
98
|
+
### Tone Target
|
|
99
|
+
|
|
100
|
+
Prefer:
|
|
101
|
+
- reverent
|
|
102
|
+
- pastorally warm
|
|
103
|
+
- doctrinally bounded
|
|
104
|
+
- emotionally honest but not sentimental
|
|
105
|
+
|
|
106
|
+
Avoid:
|
|
107
|
+
- motivational-speaker cadence
|
|
108
|
+
- secular self-help framing
|
|
109
|
+
- slang that trivializes sacred themes
|
|
110
|
+
- overconfident certainty where the original was careful
|
|
111
|
+
|
|
112
|
+
## 3. Shared Safety Questions
|
|
113
|
+
|
|
114
|
+
Before returning domain-heavy rewrites, check:
|
|
115
|
+
- Did any exact term lose precision?
|
|
116
|
+
- Did any doctrinal or technical boundary get blurred?
|
|
117
|
+
- Did the rewrite become more stylish but less trustworthy?
|
|
118
|
+
- Did the text become easier to read at the cost of being less true?
|
|
119
|
+
|
|
120
|
+
If the answer may be yes, surface the risk instead of silently polishing past it.
|
|
@@ -0,0 +1,196 @@
|
|
|
1
|
+
# Wikipedia "Signs of AI writing" — 本地工作清单
|
|
2
|
+
|
|
3
|
+
本文件改写自 Wikipedia WikiProject AI Cleanup 维护的 *Signs of AI writing*:
|
|
4
|
+
<https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing>
|
|
5
|
+
|
|
6
|
+
每次执行 de-ai-writing 都必须按 Group A → D 走一遍,**逐组打勾**判断哪些信号在响。
|
|
7
|
+
不要把所有命中都改写,挑 3–5 条最强的下手。
|
|
8
|
+
|
|
9
|
+
---
|
|
10
|
+
|
|
11
|
+
## 使用方式
|
|
12
|
+
|
|
13
|
+
1. 读完原文一遍,先不要急着改。
|
|
14
|
+
2. 沿 A→D 顺序扫描,每组在心里打√/×,命中就抄一句原文片段做证据。
|
|
15
|
+
3. 把命中数量收敛到 3–5 条最强信号,写进 `## AI味诊断`。
|
|
16
|
+
4. 改写时只针对这些信号下手。
|
|
17
|
+
5. 改完再用同一份清单回扫一次,确认没有把信号搬到新位置。
|
|
18
|
+
|
|
19
|
+
> 本地清单是 Wikipedia 分类法的快照(截至本仓库更新时)。
|
|
20
|
+
> 如果 Wikipedia 上游列出了新条目,应回写到这个文件,而不是在 SKILL.md 里堆。
|
|
21
|
+
|
|
22
|
+
---
|
|
23
|
+
|
|
24
|
+
## Group A — 内容层(Content)
|
|
25
|
+
|
|
26
|
+
围绕“**说什么**”出问题:把对象说得过于重要、列罗辑而非分析、用模糊归因装作权威。
|
|
27
|
+
|
|
28
|
+
### A1. Inflated significance(虚高重要性)
|
|
29
|
+
- 信号:随手给对象贴“开创性 / 革命性 / 里程碑 / 标志着新时代 / 深远影响”。
|
|
30
|
+
- 中文等价:极具开创性意义、具有划时代地位、深刻改变了……格局。
|
|
31
|
+
- 改写动作:删形容词;用具体动作或后果代替(比如“把延迟从 600ms 降到 80ms”)。
|
|
32
|
+
|
|
33
|
+
### A2. Notability listing(罗列“知名度”)
|
|
34
|
+
- 信号:堆作品、奖项、合作方来证明对象“值得写”。
|
|
35
|
+
- 中文等价:曾参与/合作/出席/受邀……(一长串)。
|
|
36
|
+
- 改写动作:保留与论点相关的 1–2 项;把剩下的删掉,不要换成“等”。
|
|
37
|
+
|
|
38
|
+
### A3. "-ing" superficial analysis(动名词式表面分析)
|
|
39
|
+
- 信号:英文里 “highlighting…, showcasing…, reflecting…, underscoring…”。
|
|
40
|
+
- 中文等价:……展示了……、……体现了……、……彰显了……、……折射出……。
|
|
41
|
+
- 改写动作:把“展示/体现/折射”换成具体机制:是谁、做了什么、产生了什么后果。
|
|
42
|
+
|
|
43
|
+
### A4. Promotional language(公关腔)
|
|
44
|
+
- 信号:广告/营销页式赞美。
|
|
45
|
+
- 中文等价:堪称行业标杆、用户口碑爆棚、完美解决、极致体验、强强联合。
|
|
46
|
+
- 改写动作:换成限定(“在 X 场景下表现稳定”),允许说出短板。
|
|
47
|
+
|
|
48
|
+
### A5. Vague attribution(模糊归因)
|
|
49
|
+
- 信号:“some critics say / scholars argue / it is widely believed”。
|
|
50
|
+
- 中文等价:业界普遍认为、有评论指出、不少专家表示、众所周知。
|
|
51
|
+
- 改写动作:能指名就指名,不能就直接陈述判断 + 证据;删“众所周知”。
|
|
52
|
+
|
|
53
|
+
### A6. Outline-style “challenges and future”(机械化的“挑战与展望”)
|
|
54
|
+
- 信号:“面临的挑战 / 未来展望 / 启示与思考”这种 AI 收尾标配。
|
|
55
|
+
- 改写动作:要么不写这一段;要么改成具体悬而未决的问题,不要做万金油总结。
|
|
56
|
+
|
|
57
|
+
---
|
|
58
|
+
|
|
59
|
+
## Group B — 语言层(Language)
|
|
60
|
+
|
|
61
|
+
围绕“**怎么遣词**”出问题:高密度 AI 偏好词、为避免“是”而扭曲句法、对偶/三段式上瘾。
|
|
62
|
+
|
|
63
|
+
### B1. AI vocabulary density(AI 偏好词高密度)
|
|
64
|
+
英文高频:delve, tapestry, navigate, leverage, robust, seamless, holistic, foster, paramount, intricate, nuance, embark, journey, realm, landscape, ever-evolving, transformative。
|
|
65
|
+
|
|
66
|
+
中文高频(必查):
|
|
67
|
+
- 赋能、抓手、闭环、生态、链路、底层逻辑、心智、范式
|
|
68
|
+
- 极简、极致、深度、全面、系统化、结构化
|
|
69
|
+
- 拥抱、洞察、解码、激发、点亮、照亮、释放
|
|
70
|
+
- 旅程、画卷、版图、宇宙、生态位
|
|
71
|
+
- 不仅……更……、既……又……、不止于……
|
|
72
|
+
|
|
73
|
+
> 单独出现 1–2 个不算问题;密度上来或扎堆同段就要清。
|
|
74
|
+
|
|
75
|
+
### B2. Copulative avoidance(回避“是”/“有”)
|
|
76
|
+
- 信号:刻意不用最简单的连接动词,用各种花式替代。
|
|
77
|
+
- 中文等价:过度使用“构成 / 呈现出 / 彰显出 / 演绎着 / 诠释为”代替“是”。
|
|
78
|
+
- 改写动作:能用“是 / 不是 / 有 / 没有”就直接用。
|
|
79
|
+
|
|
80
|
+
### B3. Negative parallelism(负向平行)
|
|
81
|
+
- 信号:“not just X, but Y” 的反复使用。
|
|
82
|
+
- 中文等价:不仅是……更是……、不只是……而是……、与其说是……不如说是……。
|
|
83
|
+
- 改写动作:选一个角度直接说;不要为了对仗硬凑后半句。
|
|
84
|
+
|
|
85
|
+
### B4. Rule of three(三段式上瘾)
|
|
86
|
+
- 信号:三个并列形容词、三段式列举、ABC 三步法。
|
|
87
|
+
- 中文等价:清晰、准确、有力;快速、稳定、可控;本质上是……、根本上是……、最终是……。
|
|
88
|
+
- 改写动作:改成两项或四项;或把三项中重复的删掉。
|
|
89
|
+
|
|
90
|
+
### B5. Elegant variation(雅致替换)
|
|
91
|
+
- 信号:同一概念在一段内被换三种说法假装“不重复”。
|
|
92
|
+
- 中文等价:一段里 “模型 / 系统 / 框架 / 方案” 互换指同一对象。
|
|
93
|
+
- 改写动作:除非语义真不同,否则统一称呼;专业语境下重复术语 ≠ 啰嗦。
|
|
94
|
+
|
|
95
|
+
---
|
|
96
|
+
|
|
97
|
+
## Group C — 形式与版面(Style & Format)
|
|
98
|
+
|
|
99
|
+
围绕“**怎么排版**”出问题:标点过度、加粗满天飞、模板化排版。
|
|
100
|
+
|
|
101
|
+
### C1. Em-dash overuse(破折号上瘾)
|
|
102
|
+
- 信号:每段都靠 `——` 制造“揭示感”。
|
|
103
|
+
- 改写动作:保留 1 处真正必要的;其余改逗号、句号或拆句。
|
|
104
|
+
|
|
105
|
+
### C2. Bold overuse(加粗滥用)
|
|
106
|
+
- 信号:把整句、整段重要观点全加粗,导致“没有重点”。
|
|
107
|
+
- 改写动作:一段最多 1 个加粗;非术语 / 非关键名词不加粗。
|
|
108
|
+
|
|
109
|
+
### C3. Inline-header lists(行内伪标题)
|
|
110
|
+
- 信号:列表每条以 `**关键词:**` 开头,全文像模板填空。
|
|
111
|
+
- 改写动作:要么改成正常段落叙述;要么允许长度不一的自由项。
|
|
112
|
+
|
|
113
|
+
### C4. Emoji bullets(emoji 项目符号)
|
|
114
|
+
- 信号:✅⚡🚀✨ 当作小标题或列表点。
|
|
115
|
+
- 改写动作:技术稿/讲道稿默认全删;公众号稿保留 ≤2 个有功能性的。
|
|
116
|
+
|
|
117
|
+
### C5. Curly quotes / 全角混乱
|
|
118
|
+
- 信号:英文里出现 Word 风格 “…” 智能引号、全角半角混排。
|
|
119
|
+
- 改写动作:统一为目标语种的标准引号。中文用「」或“”,不要混。
|
|
120
|
+
|
|
121
|
+
### C6. Title-case headings(英文标题滥用 Title Case)
|
|
122
|
+
- 信号:英文小标题里每个词首字母大写。
|
|
123
|
+
- 中文等价:所有小标题都做成名词短语 + 冒号开头的对仗结构。
|
|
124
|
+
- 改写动作:允许小标题长度参差;不要强求节奏。
|
|
125
|
+
|
|
126
|
+
---
|
|
127
|
+
|
|
128
|
+
## Group D — 沟通姿态(Communication)
|
|
129
|
+
|
|
130
|
+
围绕“**对读者讲话的姿态**”出问题:讨好、套话收尾、过度限定。
|
|
131
|
+
|
|
132
|
+
### D1. Sycophancy(讨好)
|
|
133
|
+
- 信号:“Great question!”、“你这个问题问得非常关键”。
|
|
134
|
+
- 中文等价:你说得太对了、这个问题确实非常值得思考、感谢你提出这么好的问题。
|
|
135
|
+
- 改写动作:直接进入回答;删讨好句。
|
|
136
|
+
|
|
137
|
+
### D2. Canned closures(罐头收尾)
|
|
138
|
+
- 信号:In conclusion, To sum up, In summary, Hope this helps。
|
|
139
|
+
- 中文等价:综上所述、总而言之、希望这能帮到你、以上就是关于……的全部内容。
|
|
140
|
+
- 改写动作:删;如果真需要收束,改成一个判断句或一个未解的问题。
|
|
141
|
+
|
|
142
|
+
### D3. Knowledge-cutoff disclaimers(知识截止时间声明)
|
|
143
|
+
- 信号:“截至我的最新知识 / 截至 20XX 年 X 月 / 据我所知最新情况”。
|
|
144
|
+
- 改写动作:除非用户明确要求声明数据时效,否则删。
|
|
145
|
+
|
|
146
|
+
### D4. Filler phrases(空填充)
|
|
147
|
+
- 信号:“值得注意的是、需要指出的是、不可忽视的是、众所周知、毋庸置疑”。
|
|
148
|
+
- 改写动作:删。要表达“注意”可以直接给那句结论。
|
|
149
|
+
|
|
150
|
+
### D5. Over-qualification(过度限定)
|
|
151
|
+
- 信号:一句话里堆 3+ 个软化词:“可能、也许、在某种程度上、相对而言、通常来说”。
|
|
152
|
+
- 改写动作:保留 ≤1 个限定;不确定就改成承认范围(“在 X 条件下成立”)而不是“可能也许大概”。
|
|
153
|
+
|
|
154
|
+
---
|
|
155
|
+
|
|
156
|
+
## 快速回扫表(改完后用)
|
|
157
|
+
|
|
158
|
+
按住下面这张表把改写稿再过一次,每条问“这条还在吗”:
|
|
159
|
+
|
|
160
|
+
| Group | 信号 | 改完后是否还在 |
|
|
161
|
+
|-------|------|----------------|
|
|
162
|
+
| A1 | 虚高重要性形容词 | ☐ |
|
|
163
|
+
| A2 | 知名度堆列 | ☐ |
|
|
164
|
+
| A3 | -ing / 展示体现折射 | ☐ |
|
|
165
|
+
| A4 | 公关腔 | ☐ |
|
|
166
|
+
| A5 | 模糊归因 | ☐ |
|
|
167
|
+
| A6 | 挑战/展望模板 | ☐ |
|
|
168
|
+
| B1 | AI 偏好词扎堆 | ☐ |
|
|
169
|
+
| B2 | 回避“是” | ☐ |
|
|
170
|
+
| B3 | 负向平行 | ☐ |
|
|
171
|
+
| B4 | 三段式 | ☐ |
|
|
172
|
+
| B5 | 雅致替换 | ☐ |
|
|
173
|
+
| C1 | 破折号过载 | ☐ |
|
|
174
|
+
| C2 | 加粗满天飞 | ☐ |
|
|
175
|
+
| C3 | 行内伪标题列表 | ☐ |
|
|
176
|
+
| C4 | emoji 项目符号 | ☐ |
|
|
177
|
+
| C5 | 引号 / 全角混乱 | ☐ |
|
|
178
|
+
| C6 | Title-case / 对仗小标题 | ☐ |
|
|
179
|
+
| D1 | 讨好开场 | ☐ |
|
|
180
|
+
| D2 | 罐头收尾 | ☐ |
|
|
181
|
+
| D3 | 知识截止声明 | ☐ |
|
|
182
|
+
| D4 | 空填充 | ☐ |
|
|
183
|
+
| D5 | 过度限定 | ☐ |
|
|
184
|
+
|
|
185
|
+
如果回扫表里仍有 ≥3 项被标记,说明改写不彻底,需要再做一轮。
|
|
186
|
+
|
|
187
|
+
---
|
|
188
|
+
|
|
189
|
+
## 与中文清单的关系
|
|
190
|
+
|
|
191
|
+
- 本文件是**判断层**:决定“哪些 AI 信号在响”。
|
|
192
|
+
- `chinese-patterns.md` 是**改写层**:给出中文具体套话的替换写法。
|
|
193
|
+
- `chinese-signals.md` 是**词频层**:高频 AI 词典,做扫描用。
|
|
194
|
+
- `domain-playbooks.md` 是**领域护栏**:技术稿 / 讲道稿不能动哪些。
|
|
195
|
+
|
|
196
|
+
调用顺序:本文件先用 → chinese-signals 做词频比对 → chinese-patterns 找替换 → domain-playbooks 复核保护项。
|
package/README.md
CHANGED
|
@@ -13,6 +13,8 @@ omc-manage setup
|
|
|
13
13
|
|
|
14
14
|
## What Gets Installed
|
|
15
15
|
|
|
16
|
+
Defaults (only the bundled sources, no extras):
|
|
17
|
+
|
|
16
18
|
| Artifact | Count | Sources |
|
|
17
19
|
|----------|-------|---------|
|
|
18
20
|
| Skills | ~62 | oh-my-claudecode + superpowers |
|
|
@@ -22,6 +24,9 @@ omc-manage setup
|
|
|
22
24
|
| Guidelines | 1 | local coding discipline prompt guidelines |
|
|
23
25
|
|
|
24
26
|
All artifacts are installed to `~/.claude/` where Claude Code discovers them automatically.
|
|
27
|
+
Adding a curated subset of [everything-claude-code](https://github.com/affaan-m/everything-claude-code)
|
|
28
|
+
on top can take totals to ~94 skills / ~35 agents / ~26 commands — see
|
|
29
|
+
[Distribution-Repo Sources](#distribution-repo-sources) below.
|
|
25
30
|
|
|
26
31
|
The bundled guidelines install into `~/.claude/CLAUDE.md` and add lightweight
|
|
27
32
|
coding discipline rules
|
|
@@ -33,14 +38,15 @@ completion with concrete evidence.
|
|
|
33
38
|
|
|
34
39
|
| Command | Description |
|
|
35
40
|
|---------|-------------|
|
|
36
|
-
| `omc-manage setup [--force] [--dry-run] [--type <type>]` | Install merged artifacts |
|
|
37
|
-
| `omc-manage doctor` | Health checks |
|
|
41
|
+
| `omc-manage setup [--force] [--dry-run] [--type <type>] [--source <name>]` | Install merged artifacts |
|
|
42
|
+
| `omc-manage doctor` | Health checks; reports each source's `kind`, `profile`, and `allowlist`, and flags `staged` distribution sources awaiting `plan apply` |
|
|
38
43
|
| `omc-manage source list` | Show configured sources |
|
|
39
|
-
| `omc-manage source sync` | Update upstream sources to latest |
|
|
40
|
-
| `omc-manage source add <name> <url
|
|
44
|
+
| `omc-manage source sync [<name>]` | Update upstream sources to latest |
|
|
45
|
+
| `omc-manage source add <name> <url> [--kind ...] [--artifacts ...] [--manifests ...] [--profiles ...]` | Add a new source, including `guidelines` and `distribution-repo` sources |
|
|
46
|
+
| `omc-manage source remove <name>` | Remove a registered source |
|
|
41
47
|
| `omc-manage source inspect <name>` | Inspect a source as a bundle/catalog instead of only as flat artifacts |
|
|
42
48
|
| `omc-manage plan install <source> --profile <name>` | Build a profile-driven install plan for a source |
|
|
43
|
-
| `omc-manage plan apply <source> --profile <name
|
|
49
|
+
| `omc-manage plan apply <source> --profile <name> [--selection-file <path>]` | Materialize a reviewed plan into source activation state, optionally curating an item-level allowlist |
|
|
44
50
|
| `omc-manage artifact list [--type <type>]` | List merged artifacts |
|
|
45
51
|
| `omc-manage artifact conflicts [--type <type>]` | Show conflict report |
|
|
46
52
|
| `omc-manage guidelines optimize [source...]` | Build maintainer-only guideline optimization artifacts |
|
|
@@ -48,13 +54,15 @@ completion with concrete evidence.
|
|
|
48
54
|
|
|
49
55
|
## Sources & Priority
|
|
50
56
|
|
|
51
|
-
| Source | Priority | Description |
|
|
52
|
-
|
|
53
|
-
| local | 1 (highest) | Your custom artifacts in `~/.omc-manage/local/` |
|
|
54
|
-
| oh-my-claudecode | 2 | Multi-agent orchestration framework |
|
|
55
|
-
| superpowers | 3 | Engineering process guardrails (TDD, debugging, etc.) |
|
|
57
|
+
| Source | Priority | Default? | Description |
|
|
58
|
+
|--------|----------|----------|-------------|
|
|
59
|
+
| local | 1 (highest) | yes | Your custom artifacts in `~/.omc-manage/local/` |
|
|
60
|
+
| oh-my-claudecode | 2 | yes | Multi-agent orchestration framework |
|
|
61
|
+
| superpowers | 3 | yes | Engineering process guardrails (TDD, debugging, etc.) |
|
|
62
|
+
| ecc / your own | 4+ | opt-in | Distribution-style repos added via `source add --kind distribution-repo` |
|
|
56
63
|
|
|
57
|
-
Local artifacts always win conflicts.
|
|
64
|
+
Local artifacts always win conflicts. Sources added via `source add` are
|
|
65
|
+
appended at the next free priority. Add your own skills:
|
|
58
66
|
|
|
59
67
|
```bash
|
|
60
68
|
mkdir -p ~/.omc-manage/local/skills/my-skill
|
|
@@ -73,10 +81,16 @@ omc-manage source sync karpathy
|
|
|
73
81
|
omc-manage setup --type guidelines
|
|
74
82
|
```
|
|
75
83
|
|
|
76
|
-
|
|
77
|
-
|
|
84
|
+
<a id="distribution-repo-sources"></a>
|
|
85
|
+
### Distribution-Repo Sources
|
|
86
|
+
|
|
87
|
+
Distribution repos (e.g. [everything-claude-code](https://github.com/affaan-m/everything-claude-code))
|
|
88
|
+
publish multiple harness surfaces, manifests, and reference material alongside
|
|
89
|
+
installable Claude artifacts. OMC absorbs them in four stages so you can pick
|
|
90
|
+
exactly what reaches `~/.claude/`:
|
|
78
91
|
|
|
79
92
|
```bash
|
|
93
|
+
# 1. Register — sync the full repo into project metadata, install nothing yet.
|
|
80
94
|
omc-manage source add ecc https://github.com/affaan-m/everything-claude-code.git \
|
|
81
95
|
--artifacts skills,agents,hooks,commands \
|
|
82
96
|
--kind distribution-repo \
|
|
@@ -84,41 +98,59 @@ omc-manage source add ecc https://github.com/affaan-m/everything-claude-code.git
|
|
|
84
98
|
--harnesses claude,codex,cursor,gemini,opencode \
|
|
85
99
|
--manifests package.json,.claude-plugin/plugin.json,agent.yaml \
|
|
86
100
|
--profiles claude-runtime,reference-only
|
|
101
|
+
|
|
102
|
+
# 2. Sync — clones into .upstream/<source>/ and reads declared manifests.
|
|
87
103
|
omc-manage source sync ecc
|
|
104
|
+
|
|
105
|
+
# 3. Inspect — normalized catalog of runtime / harness / reference surfaces.
|
|
88
106
|
omc-manage source inspect ecc
|
|
107
|
+
|
|
108
|
+
# 4. Plan — preview what a profile would activate.
|
|
89
109
|
omc-manage plan install ecc --profile claude-runtime
|
|
90
|
-
omc-manage plan apply ecc --profile claude-runtime
|
|
91
110
|
```
|
|
92
111
|
|
|
93
|
-
|
|
94
|
-
|
|
95
|
-
|
|
96
|
-
|
|
97
|
-
|
|
98
|
-
|
|
99
|
-
|
|
100
|
-
|
|
112
|
+
At this point nothing has been installed. `doctor` shows the source as
|
|
113
|
+
`staged, run "omc-manage plan apply <name>"` so you don't lose track of it.
|
|
114
|
+
|
|
115
|
+
To activate, choose one of:
|
|
116
|
+
|
|
117
|
+
```bash
|
|
118
|
+
# All runtime artifacts the profile selected (no item-level curation):
|
|
119
|
+
omc-manage plan apply ecc --profile claude-runtime
|
|
120
|
+
|
|
121
|
+
# Reference-only — keep the repo synced locally, install nothing:
|
|
122
|
+
omc-manage plan apply ecc --profile reference-only
|
|
101
123
|
|
|
102
|
-
|
|
124
|
+
# Curated subset via selection file (recommended for large repos):
|
|
125
|
+
omc-manage plan apply ecc \
|
|
126
|
+
--profile claude-runtime \
|
|
127
|
+
--selection-file /absolute/path/to/selection.json
|
|
128
|
+
```
|
|
129
|
+
|
|
130
|
+
A selection file is a per-artifact-type allowlist:
|
|
103
131
|
|
|
104
132
|
```json
|
|
105
133
|
{
|
|
106
|
-
"skills": ["
|
|
107
|
-
"agents": ["
|
|
134
|
+
"skills": ["agent-eval", "santa-method", "prompt-optimizer"],
|
|
135
|
+
"agents": ["harness-optimizer", "opensource-sanitizer"],
|
|
136
|
+
"commands": ["prp-prd", "prp-plan", "harness-audit"]
|
|
108
137
|
}
|
|
109
138
|
```
|
|
110
139
|
|
|
111
|
-
|
|
140
|
+
When the synced source exposes an item directory, OMC validates the names
|
|
141
|
+
against the catalog; for manifest-only surfaces it accepts the allowlist with a
|
|
142
|
+
warning and the real filtering happens in `setup` against on-disk content. The
|
|
143
|
+
allowlist becomes part of the source config in `~/.omc-manage/sources.json` and
|
|
144
|
+
is enforced on every subsequent `omc-manage setup`.
|
|
145
|
+
|
|
146
|
+
After `plan apply`, finalize with:
|
|
112
147
|
|
|
113
148
|
```bash
|
|
114
|
-
omc-manage
|
|
115
|
-
|
|
116
|
-
|
|
149
|
+
omc-manage setup --dry-run --source ecc # confirm scope
|
|
150
|
+
omc-manage setup --source ecc # install only this source
|
|
151
|
+
omc-manage doctor # verify allowlist counts
|
|
117
152
|
```
|
|
118
153
|
|
|
119
|
-
When present, the selection file becomes a source-level `allowlist`. OMC will
|
|
120
|
-
only merge those named items from that source for the selected artifact types.
|
|
121
|
-
|
|
122
154
|
## Maintainer Guideline Optimization
|
|
123
155
|
|
|
124
156
|
OMC now treats runtime prompt guidance as a first-class `guidelines` artifact,
|
|
@@ -171,19 +203,16 @@ When the same artifact exists in multiple sources:
|
|
|
171
203
|
3. **Local priority** — local always wins
|
|
172
204
|
4. **Source priority** — lower priority number wins
|
|
173
205
|
|
|
174
|
-
## Source
|
|
175
|
-
|
|
176
|
-
OMC now distinguishes between two source kinds:
|
|
206
|
+
## Source Kinds
|
|
177
207
|
|
|
178
|
-
|
|
179
|
-
- `distribution-repo` — a repo that publishes multiple harness surfaces,
|
|
180
|
-
manifests, scripts, and reference material alongside installable Claude
|
|
181
|
-
runtime artifacts
|
|
208
|
+
OMC distinguishes two source shapes:
|
|
182
209
|
|
|
183
|
-
|
|
184
|
-
|
|
185
|
-
|
|
186
|
-
|
|
210
|
+
- `content-repo` — already shaped like OMC artifact directories (`skills/`,
|
|
211
|
+
`agents/`, etc.); installed as-is.
|
|
212
|
+
- `distribution-repo` — publishes multiple harness surfaces, manifests, and
|
|
213
|
+
reference material; defaults to `installMode=planned` so ordinary `setup`
|
|
214
|
+
and `artifact list` flows do not absorb it automatically. Activate it
|
|
215
|
+
through the [Distribution-Repo Sources](#distribution-repo-sources) flow.
|
|
187
216
|
|
|
188
217
|
## License
|
|
189
218
|
|
package/bundled/manifest.json
CHANGED