opencode-skills-collection 4.0.63 → 4.0.65
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/bundled-skills/.antigravity-install-manifest.json +4 -1
- package/bundled-skills/antigravity-maintainer-batch-release/SKILL.md +1 -1
- package/bundled-skills/cv-generator/references/detailed-guide.md +0 -25
- package/bundled-skills/de-ai-writer/SKILL.md +144 -0
- package/bundled-skills/de-ai-writer/references/ai-patterns-zh.md +368 -0
- package/bundled-skills/seo-aeo-blog-writer/SKILL.md +140 -59
- package/bundled-skills/seo-aeo-content-cluster/SKILL.md +180 -61
- package/bundled-skills/seo-aeo-content-quality-auditor/SKILL.md +221 -64
- package/bundled-skills/seo-aeo-content-quality-auditor/references/seo-audit-checklist.md +63 -0
- package/bundled-skills/seo-aeo-internal-linking/SKILL.md +220 -59
- package/bundled-skills/seo-aeo-keyword-research/SKILL.md +297 -76
- package/bundled-skills/seo-aeo-landing-page-writer/SKILL.md +116 -83
- package/bundled-skills/seo-aeo-meta-description-generator/SKILL.md +316 -62
- package/bundled-skills/seo-aeo-orchestrator/SKILL.md +281 -0
- package/bundled-skills/seo-aeo-schema-generator/SKILL.md +302 -68
- package/bundled-skills/seo-aeo-schema-generator/references/aggregate-rating-schema.json +8 -0
- package/bundled-skills/seo-aeo-schema-generator/references/article-schema.json +10 -0
- package/bundled-skills/seo-aeo-schema-generator/references/breadcrumb-schema.json +5 -0
- package/bundled-skills/seo-aeo-schema-generator/references/faq-schema.json +5 -0
- package/bundled-skills/seo-aeo-schema-generator/references/howto-schema.json +7 -0
- package/bundled-skills/seo-aeo-schema-generator/references/organization-schema.json +7 -0
- package/bundled-skills/seo-aeo-schema-generator/references/product-schema.json +8 -0
- package/bundled-skills/seo-aeo-schema-generator/references/review-schema.json +8 -0
- package/bundled-skills/seo-aeo-schema-generator/references/webpage-schema.json +7 -0
- package/bundled-skills/seo-aeo-schema-generator/references/website-schema.json +7 -0
- package/bundled-skills/youtube-transcript-skills/SKILL.md +110 -0
- package/package.json +1 -1
- package/skills_index.json +101 -8
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
{
|
|
2
2
|
"schemaVersion": 1,
|
|
3
|
-
"updatedAt": "2026-09-
|
|
3
|
+
"updatedAt": "2026-09-17T02:07:42.936Z",
|
|
4
4
|
"entries": [
|
|
5
5
|
"00-andruia-consultant",
|
|
6
6
|
"007",
|
|
@@ -614,6 +614,7 @@
|
|
|
614
614
|
"ddd-context-mapping",
|
|
615
615
|
"ddd-strategic-design",
|
|
616
616
|
"ddd-tactical-patterns",
|
|
617
|
+
"de-ai-writer",
|
|
617
618
|
"debate-review",
|
|
618
619
|
"debug-buttercup",
|
|
619
620
|
"debugger",
|
|
@@ -1697,6 +1698,7 @@
|
|
|
1697
1698
|
"seo-aeo-keyword-research",
|
|
1698
1699
|
"seo-aeo-landing-page-writer",
|
|
1699
1700
|
"seo-aeo-meta-description-generator",
|
|
1701
|
+
"seo-aeo-orchestrator",
|
|
1700
1702
|
"seo-aeo-schema-generator",
|
|
1701
1703
|
"seo-audit",
|
|
1702
1704
|
"seo-authority-builder",
|
|
@@ -2114,6 +2116,7 @@
|
|
|
2114
2116
|
"youtube-seo-optimizer",
|
|
2115
2117
|
"youtube-summarizer",
|
|
2116
2118
|
"youtube-transcript",
|
|
2119
|
+
"youtube-transcript-skills",
|
|
2117
2120
|
"zapier-make-patterns",
|
|
2118
2121
|
"zcode-delegate",
|
|
2119
2122
|
"zendesk-automation",
|
|
@@ -55,7 +55,7 @@ Before changing anything:
|
|
|
55
55
|
- Use the repository validation, test, docs-security, source-credit, reference, warning-budget, and targeted app checks required by the changed files.
|
|
56
56
|
- Fix deterministic policy failures in the source; do not wait for them as if they were flaky CI.
|
|
57
57
|
- For source-only changed paths, a Git copy changes only its destination; renames change both paths. Preserve independent raw-record and blob safety checks and regressions for genuine generated-file mutations.
|
|
58
|
-
- Treat `pr-policy` fork classification from the exact protected-base implementation as an unprivileged fail-fast gate before dependent work, never as approval authority. Install and resolve every dependency used by that classifier from the same protected-base worktree; never expose it to pull-request-controlled `node_modules`. `merge:batch` must still recompute the current trusted decision before approving any fork run or merging.
|
|
58
|
+
- Treat `pr-policy` fork classification from the exact protected-base implementation as an unprivileged fail-fast gate before dependent work, never as approval authority. Install and resolve every dependency used by that classifier from the same protected-base worktree; never expose it to pull-request-controlled `node_modules`. `merge:batch` must still recompute the current trusted decision before approving any fork run or merging. The intake allowlist also covers browser source under `apps/web-app/src/**` (`.css`, `.ts`, `.tsx`); those fork runs may be approved, but every web-app source change still requires an exact-head maintainer attestation before merge.
|
|
59
59
|
- Treat `impact_profile` as shadow-only telemetry. It must not skip, downgrade, or satisfy any required check.
|
|
60
60
|
- For ordinary source PRs, require `source-validation` to generate preview state once and `artifact-preview` to verify the manifest bound to the exact head and run identity. For canonical-sync PRs, rely on `pr-policy` exact-tree reproduction, keep `source-validation` lightweight, require `artifact-preview` to confirm no drift, and retain final CI and CodeQL on the merged `main` commit.
|
|
61
61
|
- Keep timing observational and test sharding opt-in. Required CI must continue to run the full unsharded `npm run test`; deterministic local shards may be used only through `npm run test:local -- --shard-index N --shard-count M`.
|
|
@@ -2,31 +2,6 @@
|
|
|
2
2
|
|
|
3
3
|
> This file contains the detailed procedure and reference material extracted from `SKILL.md` for focused loading. The root skill defines activation, examples, safety constraints, and limitations.
|
|
4
4
|
|
|
5
|
-
## FLAW REGISTER — KNOWN ISSUES FIXED IN THIS VERSION
|
|
6
|
-
|
|
7
|
-
The following issues were identified across the two prior skill drafts and are corrected here:
|
|
8
|
-
|
|
9
|
-
| # | Flaw | Fix applied |
|
|
10
|
-
|---|------|-------------|
|
|
11
|
-
| F-01 | Output was Markdown-first, not paste-ready plain text | Final output is plain text; Markdown is internal staging only |
|
|
12
|
-
| F-02 | FlowCV/Canva field structure was never addressed | Section mapping to tool fields added (section 11c) |
|
|
13
|
-
| F-03 | Questionnaire dumped all 20 questions at once in practice | Hard rule: one question at a time, wait for answer |
|
|
14
|
-
| F-04 | Anti-hallucination rules listed but never enforced structurally | Enforcement gate added before every output (section 10) |
|
|
15
|
-
| F-05 | Cover letter was offered but never scoped for these tools | Cover letter now outputs to a separate plain-text block, not inline |
|
|
16
|
-
| F-06 | ATS check listed but had no scored output | Flaw report now scores 0–100 with per-item pass/fail |
|
|
17
|
-
| F-07 | Seniority detection was "detect or ask" with no fallback | Default is mid-level if undetectable; user is told the assumption |
|
|
18
|
-
| F-08 | No guidance on what FlowCV/Canva cannot render | Added explicit field-by-field paste map (section 11c) |
|
|
19
|
-
| F-09 | Tense rules stated but never verified in quality gate | Tense check is now a hard gate — output blocked until corrected |
|
|
20
|
-
| F-10 | "Passionate about" and similar banned phrases still appeared in examples | Phrase blocklist now machine-checkable (section 7c) |
|
|
21
|
-
| F-11 | Nepal/South Asia market conventions were present but incomplete | Confirmed and expanded (section 14) |
|
|
22
|
-
| F-12 | No explicit rule on what to do when LinkedIn scraping is blocked | Hard fallback rule: ask for PDF export immediately, do not proceed empty |
|
|
23
|
-
| F-13 | File naming convention mentioned once, never enforced | File name rule is part of the final output block (section 11) |
|
|
24
|
-
| F-14 | Skill had no version history or upgrade path | Version field added to frontmatter |
|
|
25
|
-
| F-15 | GitHub was listed as a source but extraction rules were missing | GitHub extraction rules added (section 4f) |
|
|
26
|
-
|
|
27
|
-
---
|
|
28
|
-
|
|
29
|
-
|
|
30
5
|
## 1. Invocation
|
|
31
6
|
|
|
32
7
|
```
|
|
@@ -0,0 +1,144 @@
|
|
|
1
|
+
---
|
|
2
|
+
name: de-ai-writer
|
|
3
|
+
description: "Chinese AI-smell removal engine: 35 Chinese AI-tell patterns (赋能/闭环), AI-smell scoring, de-AI rewriting, style clone. Use when a Chinese draft reads machine-written or the user asks 去AI味."
|
|
4
|
+
category: content-writing
|
|
5
|
+
risk: safe
|
|
6
|
+
source: community
|
|
7
|
+
source_repo: jiawood2006/hermes-skills
|
|
8
|
+
source_type: community
|
|
9
|
+
date_added: "2026-09-14"
|
|
10
|
+
author: jiawood2006
|
|
11
|
+
license: "MIT"
|
|
12
|
+
license_source: "https://github.com/jiawood2006/hermes-skills/blob/main/LICENSE"
|
|
13
|
+
tags: [chinese, writing, editing, humanize, de-ai, copywriting, style-clone, ai-detection]
|
|
14
|
+
tools: [claude, cursor, codex, gemini, hermes]
|
|
15
|
+
---
|
|
16
|
+
|
|
17
|
+
# De-AI Writer — Chinese AI-Smell Removal
|
|
18
|
+
|
|
19
|
+
## Overview
|
|
20
|
+
|
|
21
|
+
Chinese AI writing has its own tells, and they are not the English ones. English humanizers hunt `delve`, "it's not just X, it's Y" and em-dash overuse; a Chinese draft reads machine-written because of 赋能 / 闭环 / 抓手 / 底层逻辑 (pattern 15), 首先-其次-最后 scaffolding (pattern 30), 随着…的发展 openers (pattern 25), 拔高意义 endings (pattern 16), and 公文套话 (pattern 22). A translated English humanizer misses all of it.
|
|
22
|
+
|
|
23
|
+
This skill ships the pattern catalog plus the editing procedure: score a draft for AI smell, rewrite it against the specific patterns it hits, clone a reference style, and review the result. The full 35-pattern catalog lives in `references/ai-patterns-zh.md` and is plain Markdown — usable as a prompt by any assistant.
|
|
24
|
+
|
|
25
|
+
*Note on the engine: this bundle is documentation only. A zero-dependency local rule engine is published in the source repository (see `source_repo`) as an optional external prerequisite — it is not included here.*
|
|
26
|
+
|
|
27
|
+
## When to Use This Skill
|
|
28
|
+
|
|
29
|
+
- Use when a Chinese draft "reads like AI" and needs to sound human-authored.
|
|
30
|
+
- Use when the user asks 去AI味, 改得像人写的, or 这段是不是AI写的 (is this AI-written?).
|
|
31
|
+
- Use when editing marketing copy, WeChat articles, product listings, or social posts written in Chinese.
|
|
32
|
+
- Use when asked to imitate a reference writing style (风格克隆) or to produce A/B variants of the same copy.
|
|
33
|
+
- Use when shifting tone: casual / formal / marketing / humor / direct.
|
|
34
|
+
|
|
35
|
+
## How It Works
|
|
36
|
+
|
|
37
|
+
### Step 1: Score the draft first (diagnose before editing)
|
|
38
|
+
|
|
39
|
+
Scan the text, record which patterns hit *and where*, then compute the AI-smell index with the deterministic formula below. Do not rewrite from vibes — the same phrases recur, and you need the hit list to verify the edit afterwards.
|
|
40
|
+
|
|
41
|
+
#### The AI-smell index (deterministic — the same formula must be used before and after)
|
|
42
|
+
|
|
43
|
+
1. **Count hits per paragraph.** For each pattern, count one hit per paragraph — repeats inside the same paragraph do not inflate the score.
|
|
44
|
+
2. **Weight by evidence strength.** 强模式 = **2 points**; patterns the catalog marks 弱证据 (破折号 / 限定词 / 被动与无主语 / "的"-字堆叠 / 引号不统一) = **1 point**.
|
|
45
|
+
3. **Normalize by length.** `D = 加权总分 / max(1, 总字数 / 100)` — weighted hits per 100 characters.
|
|
46
|
+
4. **Index.** `AI味指数 = min(100, round(D × 10))`.
|
|
47
|
+
5. **Bands.** 0–20 基本像人写 · 21–45 轻度 AI 味 · 46–75 明显 AI 味 · 76–100 一眼假.
|
|
48
|
+
6. **Report three numbers, not one:** 命中处数 / 加权总分 / AI味指数. After rewriting, recompute with the same formula so the delta is comparable. If a hit cannot be attributed to a catalogued pattern, report only the observable hit count and say the index is not computed — never invent a number.
|
|
49
|
+
|
|
50
|
+
Worked example (the sample below): 8 hits in 77 characters, 6 strong (6 × 2 = 12) + 2 weak (2 × 1 = 2) → 加权总分 14 → D = 14 / 0.77 ≈ 18.2 → 指数 = min(100, 182) = **100/100**.
|
|
51
|
+
|
|
52
|
+
```
|
|
53
|
+
AI 味体检报告
|
|
54
|
+
总字数 77 | 命中 8 处 | 加权 14 | AI味指数 100/100(一眼假)
|
|
55
|
+
机械连接 ×3 官方黑话 ×2 空洞拔高 ×2 夸张词 ×1
|
|
56
|
+
```
|
|
57
|
+
|
|
58
|
+
### Step 2: Rewrite against the specific patterns, not in general
|
|
59
|
+
|
|
60
|
+
Work through the hit list one pattern at a time. Correct each pattern by its own rule (the catalog gives 识别特征 → 为什么假 → 改前/改后 for all 35). Two rules govern the whole pass:
|
|
61
|
+
|
|
62
|
+
- **A single hit is not evidence.** The catalog marks certain patterns (em-dash, hedges, passive voice, 的-stacking, quotation marks) as 弱证据 — only act when two or more appear in the same paragraph.
|
|
63
|
+
- **Delete, don't decorate.** Most patterns disappear by deleting the sentence that carries them: drop the negation half of 不是 X,而是 Y, drop the significance ending, drop the 开场铺垫, drop the assistant residue (希望对你有帮助).
|
|
64
|
+
|
|
65
|
+
**The facts come from the source, never from the pattern list.** A rewrite may delete packaging, reorder, and rephrase; it must not introduce a fact the source does not contain. If the rewrite would be clearer with a number or a specification (rotational speed, battery life, materials), and the source has none, ask the user for it — never supply one.
|
|
66
|
+
|
|
67
|
+
### Step 3: Verify the rewrite
|
|
68
|
+
|
|
69
|
+
Re-score the rewritten text with the same Step 1 formula. The index should drop and the semantic content must be preserved — report the before/after triple (命中处数 / 加权总分 / 指数), how many patterns were cleared, and confirm that no fact was added or lost. A rewrite that lowers the score by deleting facts is a failed edit; so is one that raises the count of facts in the text.
|
|
70
|
+
|
|
71
|
+
### Step 4: Optional — style clone, variants, tone, review
|
|
72
|
+
|
|
73
|
+
- **Style clone**: supply a reference sample (an old article, a novel fragment, a writer you like) and match its sentence rhythm, vocabulary and colloquial ratio.
|
|
74
|
+
- **Variants**: produce 2-6 clearly different versions (short and punchy / loose and spoken / vivid) for headline and ad-copy A/B tests.
|
|
75
|
+
- **Tone**: re-target the same content to casual, formal, marketing, humor or direct register.
|
|
76
|
+
- **Review**: score the finished text on Hook / Pacing / Emotion / AI-Smell / Clarity / Persuasion / Structure / Readability, plus three concrete improvements.
|
|
77
|
+
|
|
78
|
+
## Examples
|
|
79
|
+
|
|
80
|
+
### Example 1: Business copy with stacked AI tells (deletion-only edit)
|
|
81
|
+
|
|
82
|
+
Before:
|
|
83
|
+
|
|
84
|
+
```
|
|
85
|
+
首先,这款产品不仅性能卓越,更是彰显了品牌的匠心精神。随着智能家居时代的发展,
|
|
86
|
+
它以用户为中心,赋能每一个家庭的智慧生活,具有里程碑意义。
|
|
87
|
+
```
|
|
88
|
+
|
|
89
|
+
After — **deletion-only; every surviving phrase already appears in the source, and no specification was added**:
|
|
90
|
+
|
|
91
|
+
```
|
|
92
|
+
这款产品性能卓越,能融入家庭的智能生活。
|
|
93
|
+
```
|
|
94
|
+
|
|
95
|
+
Every change is tied to a catalogued pattern:
|
|
96
|
+
|
|
97
|
+
- 首先 → **pattern 30** (强制编号与小标题, mandatory 首先/其次 numbering)
|
|
98
|
+
- 不仅…更是 → **pattern 1** (不是 X,而是 Y — negation-half padding)
|
|
99
|
+
- 随着智能家居时代的发展 → **pattern 25** (\"随着……的发展\"式开头)
|
|
100
|
+
- 以用户为中心 / 赋能 → **pattern 15** (高频 AI 词) and **pattern 23** (互联网黑话)
|
|
101
|
+
- 彰显了品牌的匠心精神 / 具有里程碑意义 → **pattern 16** (拔高意义)
|
|
102
|
+
|
|
103
|
+
Six patterns removed, zero facts added, zero facts lost: the rewrite keeps 性能卓越 and 智能生活, the only two claims present in the source. It does **not** gain a rotation speed, a battery life or a waterproof rating — those are not in the source, and inventing them is exactly the failure mode this skill exists to prevent.
|
|
104
|
+
|
|
105
|
+
### Example 2: Assistant residue in generated output
|
|
106
|
+
|
|
107
|
+
Before:
|
|
108
|
+
|
|
109
|
+
```
|
|
110
|
+
当然可以!以下是我为你整理的三个要点,希望对你有所帮助。需要我展开说明吗?
|
|
111
|
+
```
|
|
112
|
+
|
|
113
|
+
After:
|
|
114
|
+
|
|
115
|
+
```
|
|
116
|
+
三个要点:
|
|
117
|
+
```
|
|
118
|
+
|
|
119
|
+
Pattern 31 (聊天机器人残留) — the wrapper adds nothing; the content stands alone.
|
|
120
|
+
|
|
121
|
+
## Best Practices
|
|
122
|
+
|
|
123
|
+
- ✅ Score before and after; report the triple (命中处数 / 加权总分 / 指数) so the edit is verifiable.
|
|
124
|
+
- ✅ Quote the pattern number for every change — it makes the edit reviewable and teachable.
|
|
125
|
+
- ✅ Keep every fact, number and claim from the source; only the packaging should change.
|
|
126
|
+
- ✅ Respect 弱证据 — two hits in one paragraph, not one hit anywhere.
|
|
127
|
+
- ✅ Keep the catalog in Chinese when editing Chinese; the tells are language-specific.
|
|
128
|
+
- ❌ Don't invent facts to make the rewrite concrete — if the source has no number, ask for one.
|
|
129
|
+
- ❌ Don't swap one AI word for another AI word (赋能 → 助力 solves nothing).
|
|
130
|
+
- ❌ Don't "polish" a draft into formal register — that usually adds AI smell rather than removing it.
|
|
131
|
+
- ❌ Don't strip caveats and qualifiers that carry real meaning (legal disclaimers, safety warnings).
|
|
132
|
+
|
|
133
|
+
## Limitations
|
|
134
|
+
|
|
135
|
+
- **Chinese-centric by design.** The pattern catalog targets Chinese AI tells; it will not fix English AI smell (use an English humanizer for that).
|
|
136
|
+
- **Documentation only.** No script and no engine is bundled in this skill. The optional local rule engine is published in the source repository; deep rewriting, style cloning, variants and scoring require an LLM of your choice.
|
|
137
|
+
- **Rule-based detection, not a detector model.** The index is a heuristic over known patterns, computed by the formula above. It cannot prove authorship and must not be used as evidence that a text "was" or "was not" AI-written.
|
|
138
|
+
- **Unlisted patterns are out of scope.** Tells that are not in the 35-pattern catalog pass through untouched; the catalog is the ceiling of what this skill sees.
|
|
139
|
+
- **Keep-conditions need human confirmation.** Some removal is context-dependent — a safety caveat such as "do not soak for long periods" may be a legal requirement; when a pattern overlaps with a claim that must stay, ask before deleting.
|
|
140
|
+
|
|
141
|
+
## Reference
|
|
142
|
+
|
|
143
|
+
- [`references/ai-patterns-zh.md`](references/ai-patterns-zh.md) — the full 35-pattern Chinese AI-smell catalog, grouped into 摆姿势 / 机械节奏 / 注水借势 / 格式装饰 / 助手残留, each entry giving 识别特征 → 为什么假 → 改前/改后.
|
|
144
|
+
- Source repository (MIT, optional runnable engine): https://github.com/jiawood2006/hermes-skills
|
|
@@ -0,0 +1,368 @@
|
|
|
1
|
+
# 中文 AI 腔模式清单 / Chinese AI-Smell Pattern Catalog
|
|
2
|
+
|
|
3
|
+
> **怎么用**:改写前先扫一遍这 35 条,标出命中的位置,再动手。清单给的是"识别特征 → 为什么假 → 改前/改后",不是禁用词表。
|
|
4
|
+
>
|
|
5
|
+
> **⚠️ 最重要的一条原则**:**单条命中不算证据。** 好的写手会故意用其中任何一条。标注了 `弱证据` 的模式必须**同段落出现 2 条以上**才动手。引文、标题、专有名词里出现的词一律不动。
|
|
6
|
+
>
|
|
7
|
+
> **来源**:中文版由 [De-AI Writer](https://github.com/jiawood2006/hermes-skills) 团队整理,模式框架参考 Wikipedia [Signs of AI writing](https://en.wikipedia.org/wiki/Wikipedia:Signs_of_AI_writing)(WikiProject AI Cleanup 维护)与 blader/humanizer(MIT)。中文特有腔调(第 22–25、29–30、35 条等)为本项目补充。
|
|
8
|
+
|
|
9
|
+
---
|
|
10
|
+
|
|
11
|
+
## A. 只摆姿势,不给信息 / Staging instead of stating
|
|
12
|
+
|
|
13
|
+
这一组是当前最强烈的破绽。**命中一条即可动手。**
|
|
14
|
+
|
|
15
|
+
### 1. 不是 X,而是 Y
|
|
16
|
+
|
|
17
|
+
**识别**:不是……而是……;不仅是……更是……;与其说……不如说……;并非……而是……;拆成两句的版本("这不意味着 X。它意味着 Y。");句尾的否定小尾巴("……,不是猜的")。
|
|
18
|
+
**病因**:否定的一半在打倒一个没人主张的观点,让肯定的一半显得更大。它增加的是分量,不是信息。**除非读者真的相信被否定的那一半,否则直接说要说的。**
|
|
19
|
+
**改前**:
|
|
20
|
+
> 这不仅是节奏与编曲的技术问题,更是整个乐队表达态度的方式。它不只是一首歌,它是一个宣言。
|
|
21
|
+
**改后**:
|
|
22
|
+
> 沉重的鼓点让整首歌听起来更凶。
|
|
23
|
+
|
|
24
|
+
### 2. 一句收尾 + 戏剧化短句
|
|
25
|
+
|
|
26
|
+
**识别**:单独成段的短句,重复前文结论("这才是真正的关键。");"再读一遍。""细品。""你品,你细品。";每节后面都跟同一个收尾句;一连串碎句("没有审美预设。没有怀旧。");全大写或加句点的强调("每。一。天。")。
|
|
27
|
+
**病因**:这种句子在要求读者停下来,而不是补充内容。短句有价值的前提是它带来新事实。
|
|
28
|
+
**改前**:
|
|
29
|
+
> 然后 AlphaEvolve 出现了。它对对称没有偏好。没有审美预设。没有对人工设计的怀旧。旧规则消失了。
|
|
30
|
+
**改后**:
|
|
31
|
+
> AlphaEvolve 改变了搜索,因为它不偏好对称或"看起来像人做的"设计,这让一些旧假设不再适用。
|
|
32
|
+
|
|
33
|
+
### 3. 听起来很深的话
|
|
34
|
+
|
|
35
|
+
**识别**:真正的问题是;本质上;归根结底;说到底;核心在于;真正重要的是;从更深层看;X 是 Y 的镜子/试金石/底色;X 的底层逻辑;X 成为了一场修行;时代的语言;信任的货币。
|
|
36
|
+
**病因**:把一个普通观点包装成隐秘真理,包装本身不提供任何细节。
|
|
37
|
+
**改前**:
|
|
38
|
+
> 真正的问题是团队能不能适应。归根结底,核心在于组织的准备度。
|
|
39
|
+
**改后**:
|
|
40
|
+
> 问题是团队能不能适应,这主要取决于组织愿不愿意改变习惯。
|
|
41
|
+
|
|
42
|
+
### 4. 开场铺垫(把"我要说了"当成内容)
|
|
43
|
+
|
|
44
|
+
**识别**:让我们一起来看看;接下来我们深入探讨;不多说,直接上干货;给大家分享一个点;说个扎心的事实;先划重点;敲黑板;注意了;讲真;说实话(独立成句时)。
|
|
45
|
+
**病因**:作者在宣布"我要说重点"或表演一次坦诚,而不是直接说。删掉的是整个铺垫,不只是语气词。**句中的"说实话"是正常口语,独立成句的"说实话?"才是破绽。**
|
|
46
|
+
**改前**:
|
|
47
|
+
> 让我们一起来看看 Next.js 的缓存机制。这是你需要知道的重点。
|
|
48
|
+
**改后**:
|
|
49
|
+
> Next.js 在多层缓存数据:请求记忆化、数据缓存、路由缓存。
|
|
50
|
+
|
|
51
|
+
### 5. 跟空气辩论
|
|
52
|
+
|
|
53
|
+
**识别**:这不是(主要)关于……;我不是说……;澄清一下;别误会;并不是说……;有人可能会说……但;一个诱人的方案是;很容易想到;你可能以为……其实;顺带一提。
|
|
54
|
+
**病因**:文章在回应一个别处都不存在的反对意见(多半是上一稿的残留),或在否定一个从没出现的方案。
|
|
55
|
+
**改前**:
|
|
56
|
+
> 这并不是主要在讲提示词长度,我也不是在说文档不重要。你当然可以从另一个角度归类这个问题,但关键在于智能体执行时能不能用上这条指令。
|
|
57
|
+
**改后**:
|
|
58
|
+
> 关键在于智能体执行时能不能用上这条指令。
|
|
59
|
+
|
|
60
|
+
### 6. 【中】设问式过渡
|
|
61
|
+
|
|
62
|
+
**识别**:那么,问题来了——;为什么这么说?;这背后是什么?;究竟是什么让它……?;接下来会发生什么?
|
|
63
|
+
**病因**:用一个自问自答制造节奏,答案往往就是下一句的普通事实。中文营销号尤其爱用。
|
|
64
|
+
**改前**:
|
|
65
|
+
> 那么,问题来了——为什么这家店能火?答案其实很简单。
|
|
66
|
+
**改后**:
|
|
67
|
+
> 这家店火起来的原因很直接:位置在写字楼下,中午十一点半就开始排队。
|
|
68
|
+
|
|
69
|
+
### 7. 【中】假坦诚开场
|
|
70
|
+
|
|
71
|
+
**识别**:说句实话;讲真的;不吹不黑;有一说一;客观来讲;我不怕得罪人。
|
|
72
|
+
**病因**:用"我很坦诚"的开场来给后面的普通观点加权重。**句中偶尔出现是口语,独立开篇才是pose。**
|
|
73
|
+
**改前**:
|
|
74
|
+
> 有一说一,客观来讲,这个方案其实还行。
|
|
75
|
+
**改后**:
|
|
76
|
+
> 这个方案能用。
|
|
77
|
+
|
|
78
|
+
---
|
|
79
|
+
|
|
80
|
+
## B. 机械的节奏感 / Rhythm by rule
|
|
81
|
+
|
|
82
|
+
好的写手也会刻意用这一组,所以它们是 `弱证据`——**要同段落凑够 2 条以上再动手**。
|
|
83
|
+
|
|
84
|
+
### 8. 强行三段
|
|
85
|
+
|
|
86
|
+
**识别**:三个并列词("创新、灵感与洞察");三个排比的例子;三个短事实后面跟一句总结道理;中文里表现为"既要……又要……还要……"的三连。
|
|
87
|
+
**病因**:意思不一定有三部分,但为了"完整感"凑成三。检查每一项是否真的带来不同的意思。
|
|
88
|
+
**改前**:
|
|
89
|
+
> 本次活动包括主题演讲、圆桌讨论和社交环节,参会者可以期待创新、灵感与行业洞察。
|
|
90
|
+
**改后**:
|
|
91
|
+
> 活动有演讲和圆桌,中间也留了自由交流的时间。
|
|
92
|
+
|
|
93
|
+
### 9. 句首重复
|
|
94
|
+
|
|
95
|
+
**识别**:连续几句用同一个主语开头(尤其是"他""她""我"),按规则而不是按语感处理重复。
|
|
96
|
+
**病因**:真人也会故意重复句首做节奏("她来了。她看见。她征服。"),所以只在一段里连续三四句时才算。
|
|
97
|
+
**改前**:
|
|
98
|
+
> 她记下了那扇门。她记下了门上的锁。她把这些都存档了。
|
|
99
|
+
**改后**:
|
|
100
|
+
> 她记下了门和门上的锁,一起存档。
|
|
101
|
+
|
|
102
|
+
### 10. 破折号当万能连接词
|
|
103
|
+
|
|
104
|
+
**规则**:终稿不得出现 `——`(含 `--` 冒充的),除非作者样本里本来就用。替成句号、逗号、冒号或括号,或重写句子。**代码块、命令行、路径、URL 里的连字符不动。**
|
|
105
|
+
**病因**:破折号让作者跳过"两个分句到底是什么关系"的选择,所以模型到处用。**注意:单个破折号是弱证据**(很多记者也用),一段里出现三四个才是破绽。
|
|
106
|
+
**改前**:
|
|
107
|
+
> 新政策——在毫无预警的情况下宣布——影响了数千名员工。这些变化 -- 批评者称早就该来 -- 将立即生效。
|
|
108
|
+
**改后**:
|
|
109
|
+
> 新政策在毫无预警的情况下宣布,影响数千名员工。批评者称这些变化早就该来,将立即生效。
|
|
110
|
+
|
|
111
|
+
### 11. 限定词堆叠
|
|
112
|
+
|
|
113
|
+
**识别**:客观来说;一定程度上;在某种情况下;可能存在;有可能或许;某种意义上是;据推测。
|
|
114
|
+
**病因**:反复修改时一层层加限定词修补早先的过度断言,结果每句话都发虚。**只保留信源支持、且意思需要的限定词。** 法律、安全、勘误说明要留。"也许""一般来说"是正常人的习惯,不算破绽。`弱证据`
|
|
115
|
+
**改前**:
|
|
116
|
+
> 有可能或许可以认为该政策或许会对结果产生一定影响。
|
|
117
|
+
**改后**:
|
|
118
|
+
> 该政策可能影响结果。
|
|
119
|
+
|
|
120
|
+
### 12. 被动与无主语
|
|
121
|
+
|
|
122
|
+
**识别**:"被广泛认为""得到了充分体现""被认为是……""需要注意""由此可见"。
|
|
123
|
+
**病因**:藏起了动作的发出者。能说清谁做了什么时,用主动句。`弱证据`
|
|
124
|
+
**改前**:
|
|
125
|
+
> 无需配置文件。结果会被自动保存。
|
|
126
|
+
**改后**:
|
|
127
|
+
> 你不需要配置文件。系统会自动保存结果。
|
|
128
|
+
|
|
129
|
+
### 13. 【中】排比口号体
|
|
130
|
+
|
|
131
|
+
**识别**:既要……又要……还要……;不仅……而且……更……;跨越……跨越……跨越……;四字一顿的连续排比;"以 X 为抓手,以 Y 为依托,以 Z 为目标"。
|
|
132
|
+
**病因**:用排比造气势,但每一小句都不承担具体信息。读起来像领导讲话或广告语,不是人在陈述事实。
|
|
133
|
+
**改前**:
|
|
134
|
+
> 这次升级既提升了性能,又优化了体验,更重塑了行业标准。
|
|
135
|
+
**改后**:
|
|
136
|
+
> 这次升级把首屏加载从 2.4 秒压到 0.8 秒,并把设置页从五级菜单收成一级。
|
|
137
|
+
|
|
138
|
+
### 14. 【中】"的"字堆叠
|
|
139
|
+
|
|
140
|
+
**识别**:一句里四个以上"的";"……的……的……的实现";长定语套长定语。
|
|
141
|
+
**病因**:中文的定语链是模型最容易失控的地方,读起来像翻译稿。拆成短句或用动词开头。
|
|
142
|
+
**改前**:
|
|
143
|
+
> 这是一个能够满足企业级的、多租户的、可扩展的数据处理的需求的方案。
|
|
144
|
+
**改后**:
|
|
145
|
+
> 这个方案面向多租户场景,能满足企业级数据处理和扩容需求。
|
|
146
|
+
|
|
147
|
+
---
|
|
148
|
+
|
|
149
|
+
## C. 注水与借势 / Inflation and borrowed authority
|
|
150
|
+
|
|
151
|
+
事实通常没问题,要删的是外面的包装。
|
|
152
|
+
|
|
153
|
+
### 15. 高频 AI 词
|
|
154
|
+
|
|
155
|
+
**识别(中文)**:赋能、闭环、抓手、颗粒度、对齐、心智、生态位、打法、组合拳、方法论、底层逻辑、痛感、卡点、复盘、拉通、沉淀、复用、画像、链路、锚点、势能、降本增效、提质增效、数字化转型。
|
|
156
|
+
**识别(英文,若产出英文内容)**:delve, crucial, robust, pivotal, tapestry, testament, underscore, showcase, landscape, intricate, meticulous, vibrant, foster, enhance。
|
|
157
|
+
**病因**:模型使用这些词的频率远高于真人,且爱成串出现。**这是本清单唯一的词表**——不在表里的正式词汇本身不是破绽。
|
|
158
|
+
**改前**:
|
|
159
|
+
> 此外,一个显著的特点是,该方案通过底层逻辑的打通,实现了链路的闭环,成为数字化转型的有力抓手。
|
|
160
|
+
**改后**:
|
|
161
|
+
> 该方案把原来分属三个系统的订单数据合到一处,客服不用再切系统查单。
|
|
162
|
+
|
|
163
|
+
### 16. 拔高意义
|
|
164
|
+
|
|
165
|
+
**识别**:具有里程碑意义;标志着……的到来;起到了关键作用;奠定了坚实基础;具有深远的历史意义;引领行业迈向新阶段;开启了新篇章;交出了一份亮眼的答卷;尽管面临挑战……依然蓬勃发展;展望未来;未来可期。
|
|
166
|
+
**病因**:一个普通事实被说成"标志了变化、证明了遗产、承诺了未来"。三个尺度上都出现:一个短语、一节"挑战与展望"、一段送别式结尾。**保留事实,删掉意义。**
|
|
167
|
+
**改前**:
|
|
168
|
+
> 该公司成立于 1989 年,标志着区域统计事业发展的重要里程碑。这一举措是西班牙行政权力下放浪潮的重要组成部分。
|
|
169
|
+
**改后**:
|
|
170
|
+
> 该公司成立于 1989 年,属于西班牙行政职能下放的一部分。
|
|
171
|
+
**改前(展望式结尾)**:
|
|
172
|
+
> 公司前景一片光明。在追求卓越的道路上,激动人心的时刻即将到来。
|
|
173
|
+
**改后**:
|
|
174
|
+
> (直接删掉。停在最后一个具体事实上。)
|
|
175
|
+
|
|
176
|
+
### 17. 模糊关联
|
|
177
|
+
|
|
178
|
+
**识别**:与……有关;在……的背景下;伴随着……;与……紧密相连;涉及;关联到。
|
|
179
|
+
**病因**:说了两件事有关,却不说是什么关系。"他与该公司有关联"隐藏了他是 CEO、董事还是顾问。**信源没说的就别编,保留模糊表述也比编一个身份好。**
|
|
180
|
+
**改前**:
|
|
181
|
+
> 他与该乐团有关联,该乐团由他创办并指挥。音乐会与巴基斯坦建国五十周年庆祝活动有关。
|
|
182
|
+
**改后**:
|
|
183
|
+
> 他创办并指挥该乐团。音乐会是巴基斯坦建国五十周年庆祝活动的一部分。
|
|
184
|
+
|
|
185
|
+
### 18. 浅薄的"着/了"字小尾巴
|
|
186
|
+
|
|
187
|
+
**识别**:彰显着、体现着、标志着、反映着、象征着、助推着、赋能着、为……注入了新动能、进一步提升了……;英文对应 highlighting / underscoring / reflecting。
|
|
188
|
+
**病因**:在一个简单事实后面接一个短语让它显得更深。**挂在权威名字后面("某专家强调了……")也不会让它变真。**
|
|
189
|
+
**改前**:
|
|
190
|
+
> 寺庙的蓝绿金配色与该地区的自然风光相呼应,象征着德州蓝帽花与墨西哥湾,体现着当地社区与土地的深厚联结。
|
|
191
|
+
**改后**:
|
|
192
|
+
> 寺庙外墙刷成蓝绿金,用意是呼应德州的蓝帽花和墨西哥湾。
|
|
193
|
+
|
|
194
|
+
### 19. 广告腔
|
|
195
|
+
|
|
196
|
+
**识别**:坐拥、毗邻、宛如、尽享、匠心、臻选、解锁、焕新、一站式、硬核、天花板、yyds、宝藏、绝绝子、不容错过。
|
|
197
|
+
**病因**:文字读起来像广告,尤其是写地点、文化、产品、机构时。**直接说这是什么。**
|
|
198
|
+
**改前**:
|
|
199
|
+
> 坐落于风景如画的戈德尔地区,阿拉马塔是一座拥有深厚文化底蕴与绝美自然风光的活力小镇。
|
|
200
|
+
**改后**:
|
|
201
|
+
> 阿拉马塔是埃塞俄比亚戈德尔地区的一个镇。
|
|
202
|
+
|
|
203
|
+
### 20. 借来的权威
|
|
204
|
+
|
|
205
|
+
**识别**:专家指出;业内人士认为;有分析认为;据媒体报道;多个信源显示;在某某平台拥有 N 万粉丝;被 XX、XX、XX 等媒体报道。
|
|
206
|
+
**病因**:用一个名字或一个不明身份的权威代替"到底说了什么"。**信源给了真实出处就写上;没有就删掉这个断言或这串名单。永远不要编信源。** 单纯缺少引用不算破绽——大部分文字本来就没引用。
|
|
207
|
+
**改前**:
|
|
208
|
+
> 由于其独特性,该河流受到研究者和环保人士的关注。专家认为它在区域生态中起着关键作用。
|
|
209
|
+
**改后**:
|
|
210
|
+
> 研究者和环保人士在研究这条河,因为它有不同寻常的特征。
|
|
211
|
+
|
|
212
|
+
### 21. 绕开"是 / 有"
|
|
213
|
+
|
|
214
|
+
**识别**:作为……而存在;充当着……的角色;扮演着……的角色;堪称;不失为;具备着。
|
|
215
|
+
**病因**:把简单的"是""有"换成长短语。用"是"和"有"。
|
|
216
|
+
**改前**:
|
|
217
|
+
> 825 画廊作为 LAAA 的当代艺术展览空间而存在。该画廊拥有四个独立空间,坐拥 3000 平方英尺以上的面积。
|
|
218
|
+
**改后**:
|
|
219
|
+
> 825 画廊是 LAAA 的当代艺术展厅,有四个房间,共 3000 平方英尺。
|
|
220
|
+
|
|
221
|
+
### 22. 【中】公文套话
|
|
222
|
+
|
|
223
|
+
**识别**:高度重视、狠抓落实、进一步、切实、不断加强、深入推进、全面贯彻、统筹推进、协同发力、形成合力、压实责任、闭环管理、举一反三、提质增效。
|
|
224
|
+
**病因**:这套词来自公文模板,模型写"工作汇报式"中文时几乎必用。它传递的是态度而不是状态。**换成"谁、做了什么、结果如何"。**
|
|
225
|
+
**改前**:
|
|
226
|
+
> 下一步,我们将高度重视该问题,进一步加强统筹协调,压实各方责任,形成工作合力,确保各项举措落地见效。
|
|
227
|
+
**改后**:
|
|
228
|
+
> 下周我会把三个部门拉进一个群,订单异常由客服直接 @ 技术处理,处理结果记在同一个表格里。
|
|
229
|
+
|
|
230
|
+
### 23. 【中】互联网黑话
|
|
231
|
+
|
|
232
|
+
**识别**:对标、复盘、赋能、抓手、打法、拉齐、对齐一下、颗粒度、心智、品效合一、势能、引爆、破圈、重仓、加码、护城河。
|
|
233
|
+
**病因**:把具体动作抽象成行话,读的人要在脑子里再翻译一遍。**说动作,不说姿态。**
|
|
234
|
+
**改前**:
|
|
235
|
+
> 我们需要和各方对齐一下颗粒度,用组合拳打法赋能增长,形成方法论沉淀。
|
|
236
|
+
**改后**:
|
|
237
|
+
> 我周一和另外两个组确认了统计口径:按订单数还是按金额。先用订单数。
|
|
238
|
+
|
|
239
|
+
### 24. 【中】空转的"重要性"
|
|
240
|
+
|
|
241
|
+
**识别**:具有重要意义、意义重大、影响深远、起到了至关重要的作用、为……打下了坚实的基础、是……的重要保障、成为……的重要支撑。
|
|
242
|
+
**病因**:这类句子主语和谓语都不承担信息,删掉后读者不会少知道任何事。
|
|
243
|
+
**改前**:
|
|
244
|
+
> 该机制的建立对于保障系统稳定运行具有重要意义。
|
|
245
|
+
**改后**:
|
|
246
|
+
> 该机制在数据库连接断开后自动重连,平均 3 秒恢复。
|
|
247
|
+
|
|
248
|
+
### 25. 【中】"随着……的发展"式开头
|
|
249
|
+
|
|
250
|
+
**识别**:随着社会的发展;随着科技的进步;在当今这个……的时代;在经济全球化的背景下;近年来,随着……的不断深入;进入 21 世纪以来。
|
|
251
|
+
**病因**:用一个和内容无关的大背景开场,实际信息量是零。**直接从具体的那件事写起。**
|
|
252
|
+
**改前**:
|
|
253
|
+
> 随着人工智能技术的飞速发展,越来越多的企业开始关注智能客服。在这样的大背景下,我们公司也开始了探索。
|
|
254
|
+
**改后**:
|
|
255
|
+
> 我们去年双十一的客服排队峰值是 400 人,最长等 11 分钟,所以今年试了智能客服。
|
|
256
|
+
|
|
257
|
+
---
|
|
258
|
+
|
|
259
|
+
## D. 格式装饰 / Formatting by rule
|
|
260
|
+
|
|
261
|
+
模板和可视化编辑器也会产出干净排版,所以**只有当格式变成"每一条都装饰"时才算破绽**。
|
|
262
|
+
|
|
263
|
+
### 26. 加粗当装饰
|
|
264
|
+
|
|
265
|
+
**识别**:没有理由的加粗;每条列表项都是"**标签:**内容"的竖排结构。
|
|
266
|
+
**病因**:加粗应该只标出真正需要回看的关键词。把没有信息量的标签列表改写成正常段落。
|
|
267
|
+
**改前**:
|
|
268
|
+
> - **用户体验:** 通过新界面显著提升了用户体验。
|
|
269
|
+
> - **性能:** 通过算法优化提升了性能。
|
|
270
|
+
> - **安全:** 通过端到端加密加强了安全性。
|
|
271
|
+
**改后**:
|
|
272
|
+
> 这次更新改了界面,用新算法把加载提速,并加上了端到端加密。
|
|
273
|
+
|
|
274
|
+
### 27. 装饰性标题
|
|
275
|
+
|
|
276
|
+
**识别**:标题里每个实词首字母大写(英文);标题或列表项前面挂 emoji、箭头(→)、小图标;每节之间都有分隔线;文档开头有一个和标题重复的一级标题。
|
|
277
|
+
**改后**:
|
|
278
|
+
> 🚀 **启动阶段:** 产品在第三季度上线 → 产品在第三季度上线
|
|
279
|
+
> 💡 **关键洞察:** 用户偏好简洁 → 用户调研显示,用户更想要少几步操作
|
|
280
|
+
|
|
281
|
+
### 28. 引号不统一
|
|
282
|
+
|
|
283
|
+
**识别**:中文里用英文直引号 `"..."`,或该用直引号的地方出现弯引号 `"..."`;中英混排时引号、括号、逗号全角半角混乱。
|
|
284
|
+
**病因**:多数编辑器会自动弯引号,所以 `弱证据`。中文正文用「」或"",英文用 `"`,代码/路径里的引号不动。
|
|
285
|
+
|
|
286
|
+
### 29. 【中】emoji 当序号
|
|
287
|
+
|
|
288
|
+
**识别**:✅❌⚠️🔴🟢📌💡 当作列表符号或结论标记;每段开头一个 emoji。
|
|
289
|
+
**病因**:emoji 是给聊天用的,放在正式文档里会让每一条都显得在"喊"。**正文里删掉,需要保留的只在真正的提示/警告处留一个。**
|
|
290
|
+
**改前**:
|
|
291
|
+
> ✅ 已完成配置
|
|
292
|
+
> ⚠️ 注意:需要重启
|
|
293
|
+
> 💡 小技巧:可以用快捷键
|
|
294
|
+
**改后**:
|
|
295
|
+
> 配置已完成。需要重启后生效。快捷键是 ⌘K。
|
|
296
|
+
|
|
297
|
+
### 30. 【中】强制编号与小标题
|
|
298
|
+
|
|
299
|
+
**识别**:三个字的段落也硬套"一、二、三";每条内容都加一个四字小标题;"首先……其次……再次……最后……"贯穿全篇。
|
|
300
|
+
**病因**:编号应该是内容的层级,不是装饰。**如果去掉编号读者也不会迷路,就去掉。**
|
|
301
|
+
|
|
302
|
+
---
|
|
303
|
+
|
|
304
|
+
## E. 助手与草稿残留 / Leftovers from chat and draft
|
|
305
|
+
|
|
306
|
+
这一组**直接删,不需要改写**。
|
|
307
|
+
|
|
308
|
+
### 31. 聊天机器人残留
|
|
309
|
+
|
|
310
|
+
**识别**:希望对你有帮助;当然可以!;好问题!;你说得对;需要我继续吗?;要我再展开吗?;还有什么想了解的;以下是……。
|
|
311
|
+
**病因**:这是全清单里**最确定的破绽**,也最容易漏——因为它包在真实内容外面。删掉外壳,留下内容。
|
|
312
|
+
**改前**:
|
|
313
|
+
> 好问题!下面我介绍一下法国大革命。它始于 1789 年,财政危机和粮食短缺引发了大规模动荡。希望对你有帮助!需要我展开某一部分吗?
|
|
314
|
+
**改后**:
|
|
315
|
+
> 法国大革命始于 1789 年,财政危机和粮食短缺引发了大规模动荡。
|
|
316
|
+
|
|
317
|
+
### 32. 知识截止与猜测补位
|
|
318
|
+
|
|
319
|
+
**识别**:截至我的知识更新日期;根据目前公开信息;具体细节有限;没有公开资料;据了解(无出处);大概率是……;应该是……;推测其……
|
|
320
|
+
**病因**:文字在交代"模型的 knowledge 到哪结束",或在承认找不到信源后用一个听起来合理的猜测补上。**说清信源没有显示什么,或者直接删句。永远不要把猜测写成事实。**
|
|
321
|
+
**改前**:
|
|
322
|
+
> 虽然关于该公司创立的具体细节在现有公开资料中记载有限,但可以推测其成立于 1990 年代。
|
|
323
|
+
**改后**:
|
|
324
|
+
> 现有资料没有记载该公司的创立时间。(或直接删掉这句。)
|
|
325
|
+
|
|
326
|
+
### 33. 标题在首句复述
|
|
327
|
+
|
|
328
|
+
**识别**:小标题后面跟一句把标题换个说法再说一遍的话,然后才开始正文。
|
|
329
|
+
**改前**:
|
|
330
|
+
> ## 性能
|
|
331
|
+
> 速度很重要。
|
|
332
|
+
> 用户遇到慢页面就会离开。
|
|
333
|
+
**改后**:
|
|
334
|
+
> ## 性能
|
|
335
|
+
> 用户遇到慢页面就会离开。
|
|
336
|
+
|
|
337
|
+
### 34. 写"上一版"
|
|
338
|
+
|
|
339
|
+
**识别**:文档和注释在描述"这个文本替换了什么",而不是描述当前行为。
|
|
340
|
+
**病因**:上一版只在更新日志、发布说明、迁移指南里出现。
|
|
341
|
+
**改前**:
|
|
342
|
+
> 新增此函数是为了替代此前遍历所有元素的方法,后者导致 O(n²) 性能问题。
|
|
343
|
+
**改后**:
|
|
344
|
+
> 此函数用哈希表实现 O(1) 查找,避免了朴素遍历的 O(n²) 开销。
|
|
345
|
+
|
|
346
|
+
### 35. 【中】结束语套话
|
|
347
|
+
|
|
348
|
+
**识别**:综上所述;总而言之;总而言之的;总的来看;由此可见;不难看出;让我们拭目以待;期待与您携手;欢迎留言讨论。
|
|
349
|
+
**病因**:结尾再总结一遍前文,不带来新信息。**除非前面信息量大到需要收束,否则停在最后一个具体事实上更有力。**
|
|
350
|
+
**改前**:
|
|
351
|
+
> 综上所述,不难看出该方案在成本、效率、体验三方面均具有明显优势,值得我们进一步推广。
|
|
352
|
+
**改后**:
|
|
353
|
+
> 这套方案比原来省了两个人天,我们下个月在另一个仓也上。
|
|
354
|
+
|
|
355
|
+
---
|
|
356
|
+
|
|
357
|
+
## 判定规则:什么时候**不要**动手
|
|
358
|
+
|
|
359
|
+
每条模式描述的都是"默认选择",而人可能故意这么写。
|
|
360
|
+
|
|
361
|
+
- **`弱证据`的模式**(10、11、12、14、28)——只有同段落里凑够两条以上才动手。
|
|
362
|
+
- **引文、标题、专有名词、以及正在讨论这个说法本身的段落**——里面的词一律不动。
|
|
363
|
+
- **书信、评论的称呼与落款**——本来就早于聊天机器人存在。
|
|
364
|
+
- **2022 年 11 月 30 日之前写的文本**——那时候还不存在 AI 写作。
|
|
365
|
+
- **靠感觉判断**的人,准确率和瞎猜差不多;人写的文字也在不断吸收 AI 习惯。**多条同时出现才是可靠的依据。**
|
|
366
|
+
- **作者故意留下的痕迹一律保留**:具体的细节(真实地址、奇怪的引语)、矛盾的情绪("我觉得大体是好的,但它让我不安,我说不清为什么")、有年代感的梗、第一人称的私人选择、真正的自我修正("(我老想说'几乎',但这里确实是确定的。)")。
|
|
367
|
+
|
|
368
|
+
**判断顺序**:先删 E 组(助手残留)→ 再清 C 组(注水)→ 再改 A 组(摆姿势)→ 最后收 B 组(节奏)和 D 组(格式)。留到最后的,是作者的细节和人味。
|