SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillscontent-quality-gate

content-quality-gate

Writing & research

View repository

Updated 10-15 21:55

About this skill

General Four-Layer Content Quality Inspection

Design Philosophy

Drawing on the software engineering test pyramid: progressing layer by layer from mechanical hard rules to subjective judgment. Each layer has clear pass criteria, and if any layer fails, it needs to be fixed and then rechecked.

Note: This quality inspection system is complementary to STYLE.md (global writing style specification) and SELF_REFINE_CHECK.md (lightweight self-check), and does not replace them.

  • STYLE.md defines "how to write"
  • SELF_REFINE_CHECK.md is a three-dimension quick scan (suitable for cron short content)
  • content-quality-gate is a complete four-layer quality inspection (suitable for long-form and important content)

When to Use

  • Before publishing long-form WeChat Official Account articles (required)
  • Before publishing important long X/Twitter posts
  • Before finalizing technical research reports
  • Any content that needs to go through AI quality inspection before human review

Scenarios Where It Is Not Needed: daily information feed briefings, short content automatically produced by cron (SELF_REFINE_CHECK.md is sufficient)

Four-Layer Quality Inspection Process

L1 Hard Rules (Automated Scan)

The most basic "grammar check", with zero tolerance. If any item is hit, it must be fixed.

L1-1 Banned Word Scan

Scan for prohibited vocabulary defined in STYLE.md:

  • Big-tech jargon: 赋能, 闭环, 底座, 抓手, 下钻, 落地, 对齐, 沉淀, 打通, 串联, 链路, 心智, 颗粒度, 组合拳, 护城河, 矩阵, 生态位, 赛道, 痛点, 蓝海, 红海
  • Report tone: 综上所述, 一言以蔽之, 不难看出, 显然, 众所周知, 总而言之, 直奔主题, 值得重点关注
  • Internet memes: YYDS, 绝绝子, 破防, 拿捏, 上分
  • High-frequency AI pitfalls: 说白了, 意味着什么, 这意味着, 本质上, 换句话说, 不可否认
  • Clichéd structures: 首先...其次...最后, 值得注意的是, 让我们来看看, 接下来让我们

L1-2 Empty Talk Detection

  • Piling up vague adjectives: three or more consecutive modifiers of the "very/extremely/significantly/important" type
  • Empty conclusions: paragraphs with only assertions and no evidence ("XXX is very important" with no "because..." following it)
  • Textbook openings: "In today's... era", "With the development of..."

L1-3 Factual Hard Errors

  • Technical detail errors (version numbers, API names, parameter values inconsistent with reality)
  • Data contradictions (numbers in the preceding and following text do not match)
  • Missing sources (specific data is cited but no source is given)

L1-4 Format Consistency

  • Missing spaces between Chinese and English (around technical terms)
  • Mixed punctuation (Chinese and English punctuation not unified)
  • Markdown format errors (unclosed code blocks, disordered list levels)

Pass criteria: zero hits across all four items. Fix method: replace/supplement one by one.

L2 Readability Check (Pattern Matching)

Check the reading experience of the article, similar to unit testing.

L2-1 Opening check

  • Does it start from a specific event/scene/question? (Not a grand narrative or vague background)
  • Can the first 100 characters make the reader feel "and then what?"

L2-2 Rhythm check

  • Is there an alternation of long and short sentences? (4 or more consecutive sentences with similar length = rigid rhythm)
  • Is the information density even? (No long stretches of filler and no impenetrable wall of jargon)
  • Is the paragraph length reasonable? (Paragraphs over 200 characters need to be considered for splitting)

L2-3 Structure check

  • Are there natural transitions between paragraphs? (Not hard cuts)
  • Does each section have one clear core point?
  • Is there a clear overall line of progression? (Problem → analysis → conclusion, or phenomenon → cause → solution)

L2-4 Reader perspective check

  • After reading each paragraph, can the reader answer "what is this paragraph saying?"
  • Are there logical jumps that require the reader to go back and reread to understand?
  • When technical terms appear, is there enough context for the target reader to understand them?

Passing standard: L2-1 must pass, and at least 3/4 of L2-2/L2-3/L2-4 must pass.

L3 Content depth check (quality review)

Check the depth and persuasiveness of the content itself, similar to integration testing.

L3-1 Argument support

  • Does each core point have support from specific people/scenes/data/code/experiments?
  • Are there cases of "I think XXX is great" without any examples?

L3-2 Knowledge output quality

  • Does the technical explanation start from "why it exists" and then move to "how it works"?
  • Are the knowledge points brought out "conversationally in passing" or "force-fed textbook style"?
  • Are citations/references naturally integrated into the discussion?

L3-3 Originality check

  • Does this article have content where "no one else could write this angle"?
  • Or is it all common sense that is "correct but adds no information"?

L3-4 Opposing views and objectivity

  • When judgments are involved, are opposing views considered?
  • Is there one-sidedness of "only talking about the benefits and not the costs"?

L3-5 Actionability (only for tutorials/methodology)

  • Does each suggestion land on a specific action the reader "can do today"?
  • Does it honestly explain the learning cost and common failure points? (No empty promises)

Passing standard: L3-1 and L3-2 must all pass, and at least 2/3 of the others must pass (skip items that do not apply).

L4 Human feel final review (final judgment)

The most important and also the most subjective layer. Read the full text and answer the core question:

"After reading this, do I feel that an experienced person is seriously sharing, or that AI is outputting in bulk?"

L4-1 Sense of warmth

  • Is the emotional expression experiential ("I felt something was off at the time") or knowledge-based ("I felt very confused")?
  • If there are character/scene descriptions, can they make the reader "feel body warmth"?

L4-2 Uniqueness

  • Are there traces of the author's personal perspective? (Personal experience, unique judgment, distinctive expression habits)
  • Or after removing the byline, is it completely impossible to guess who wrote it?

L4-3 Stance check

  • Is the tone 'a peer sharing experience' or 'a mentor teaching a student'?
  • Does it unconsciously slip into a preaching/marketing/reporting tone?

L4-4 Flow Check

  • Read it from start to finish. Is there anywhere attention breaks and you need to go back to sort out the logic?
  • If so, that is a break point in rhythm or logic.

Pass criteria: L4-1 through L4-4 overall feel 'like it was written by a real person.' If any item feels 'too AI-ish,' it needs rework.

QC report format

After completing the four-layer QC, output:

---
## 🔍 四层质检报告

**L1 硬性规则** ✅/❌
- 禁用词：X 处命中（列表）
- 空话：X 处命中
- 事实硬伤：X 处 / 无
- 格式：X 处不一致

**L2 可读性** ✅/❌
- 开头：✅/❌
- 节奏：✅/❌（具体问题）
- 结构：✅/❌
- 读者视角：✅/❌

**L3 内容深度** ✅/❌
- 论据支撑：✅/❌（缺失位置）
- 知识输出：✅/❌
- 独创性：✅/❌
- 对立面：✅/❌/不适用
- 可执行性：✅/❌/不适用

**L4 活人感** ✅/❌
- 温度感：✅/❌
- 独特性：✅/❌
- 姿态：✅/❌
- 心流：✅/❌（断点位置）

**总评**：4 层全部通过 / L1/L2 需修复后重检 / 建议人工复审
**最需修复**：[1-3 个具体问题]

Constraints

  • QC only outputs a report,it does not automatically modify content
  • Issues must be specific to paragraph/location; vague evaluations are prohibited
  • When L4 cannot be passed, give specific guidance on 'which paragraph is heavily AI-ish,' rather than vaguely saying 'the whole thing isn't alive enough'
  • The QC report is not counted in the main text written to disk (it only appears in the Telegram output)
  • Mutually exclusive with SELF_REFINE_CHECK.md: use the four-layer QC for long-form content, use Self-Check for short content, do not use both at the same time