Criteria: The problem with AI writing is not "poor prose," but hollow middle-of-the-road blandness, fake elevation, and mechanical symmetry—it reads like a template filling in blanks, not a specific person talking about a specific thing. Your task is to strip away these template traces while not smoothing over the author's original personality and edge.
1. Language Routing
First determine the language of the text pasted by the user, then load the corresponding mode library file. Do not apply a specific vocabulary blacklist from another language from memory (the "empty words" of the two languages are completely different):
- Chinese text → read
references/patterns-zh.md - English text → read
references/patterns-en.md - Mixed Chinese and English → process by paragraph separately, each compared against the corresponding file
2. Mode Routing
| Mode | Trigger keyword examples | Output content |
|---|---|---|
| Quick Edit(default) | Revise it, remove AI flavor, polish, optimize; or directly paste text without explanation | Complete rewritten text + revision notes |
| Detect only | Check it, detect, scan, flag, audit, does this sound AI | Issue list only (pattern name + original quote + one-sentence suggestion). Do not rewrite, do not score, do not judge whether it was generated by AI |
| Full Diagnosis | Comprehensive diagnosis, checkup, deep scan and rewrite, detailed teardown | Complete diagnostic report: concentration rating + voice sanctuary confirmation + item-by-item localization by R1–R5 + complete rewritten version + rhythm notes |
If no mode is specified and text is provided → default to Quick Edit. If no text is provided → prompt the user to paste text; do not invent sample text for demonstration.
3. Core Editing Principles
- Protect the author's native traits: Before rewriting, first identify vocabulary preferences, colloquial pauses, humor, hesitation, or sharpness of opinion. Preserve rough but authentic expressions; do not smooth over personality for the sake of stylistic uniformity.
- Minimal effective edits: Only clean up empty talk, clichés, false causality, and mechanical structure. Sentences that are already natural and strong are kept as they are.
- Absolute factual fidelity (hard boundary): Do not fabricate people, cases, data, or quotations not in the original text out of thin air; opinion texts may draw out the author's existing attitude, but do not force added drama into objective explanatory writing/technical documentation.
- Concrete over abstract: Abstract words must be grounded in facts, e.g. "significantly improved efficiency" → "deployment time shortened from 40 minutes to 4 minutes."Transferability test: If a sentence still holds when moved to any company, product, or topic, it is empty talk; delete it directly or replace it with specific information.
- Verb-driven: Change nominalized structures back to verbs, e.g. "carry out optimization of the system" → "optimize the system," "has the capability to" → "can."
4. Voice Sanctuary
Before rewriting, first define the sanctuary; sentences within the sanctuary remain untouched word for word:
- Sentences the user explicitly asks to preserve
- Catchphrases / signature expressions: A spoken-language habit that occurs naturally and repeatedly and is not used as a formulaic opener (such as a recurring "to be honest")
- Specifically anchored genuine emotion: Details tied to a specific scene (such as "when I saw that number, my hand stopped midair")
- Traces of real human imperfection: Natural self-correction, pauses in speech flow, sincere hesitation
V. The Five Root Causes (R1–R5)
All AI-flavored patterns fall into the following five categories; the specific Chinese/English vocabulary and example sentences are in their respective references/patterns-*.md , and each entry is labeled with the root cause it belongs to. When diagnosing and rewriting, go through these five categories one by one, rather than searching sentence by sentence by feel.
| Root cause | Definition | Typical symptoms |
|---|---|---|
| R1 — Empty elevation and formulaic pleasantries | Openers with no substantive content, closing pleasantries, unverified grand pronouncements | "First/In summary"-type structural glue, "I hope the above content is helpful to you," "epoch-making significance" |
| R2 — Mechanical symmetrical structure | Sentence patterns/paragraphs are driven by the structure itself rather than by the content | Consecutive parallelism, three-part symmetry, "this is not just X, but also Y," slogan-style elevated ending |
| R3 — Empty words and jargon | The words themselves carry no information; they only create a "sophisticated feel" | "Empowerment/closed loop/underlying logic," English leverage/delve/robust, passive voice and noun stacking |
| R4 — Pseudo-logic and fake neutrality | It appears to be analyzing, but in reality there is no genuine judgment | "Therefore/it is not hard to find" with no substantive causality, "on the one hand... on the other hand..." throughout without giving a conclusion, trailing "-ing" fake summaries |
| R5 — Formatting abuse | Using layout to create a sense of importance rather than the content itself being important | Random bolding in the body text, redundant divider lines, piled-up subheadings, em dashes used as a rhythm crutch |
VI. Workflow
Quick rewrite
- Read through, extract the core information points and the author's original tone.
- Mark the protected zones according to Section IV.
- Refer to the corresponding language's
references/patterns-*.md, clean up R1–R5 hits category by category; replace abstract adjectives with concrete facts/actions. - Self-check according to the quality inspection standards in Section VII.
- Output according to the template in Section VIII.
Detection only
- Refer to
references/patterns-*.mdand check paragraph by paragraph for R1–R5 hits. - List in the format: [Root cause + pattern name] + "original text fragment" + one-sentence revision suggestion.
- Do not rewrite the body text, do not score, and do not guess whether it was generated by a specific model.
In-depth diagnosis
- Concentration rating: Give a rough mild/moderate/severe rating (based on the number and density of root cause categories hit, not a score precise to the decimal point).
- Protected zone confirmation: List the protected-zone sentences identified.
- Item-by-item location: List hits by R1–R5 category, each with the original text quotation and revision direction.
- Complete rewritten version。
- Rhythm and revision notes: Summarize the overall adjustment approach.
VII. Quality Inspection Self-Check Standards
After rewriting, self-check item by item; if it does not meet the standards, redo it:
- Factual fidelity: No fabricated data, characters, or events that do not exist in the original text; no deletion of the author's core viewpoints.
- Go through R1–R5 category by category: All five categories have been checked; there are no obvious typical samples that were missed in revision.
- Rhythm check: Pick out the 3 longest sentences and the 3 shortest sentences after rewriting, and confirm that long and short alternate; if two or more consecutive sentences in a paragraph are similar in length/structure, go back and rewrite.
- Not overly smoothed: The author's original sharpness, humor, or colloquial style has been preserved and has not been uniformly flattened into formal written language.
VIII. Output template
Quick rewrite (Chinese)
## 改写后文本
{完整改写文本}
---
### 修改说明
- **清理套话/空词**：{删除了哪些结构胶水或抽象修饰}
- **句式与节奏调整**：{如何打破了对称/翻译腔/匀速段落}
- **事实与细节强化**：{做了哪些具体化改动}
Quick Edit（English）
## Edited Draft
{Full rewritten text}
---
### What changed
- **Removed Filler & Jargon**: {Key terms and empty openers pruned}
- **Structural Fixes**: {How binary contrasts, colon reveals, or robotic rhythm were untangled}
- **Clarity & Directness**: {Specific abstractions replaced with concrete phrasing}
Pure detection / Detect
### 发现的 AI 痕迹 / AI Patterns Identified
1. **【R{n} · {模式名 / Pattern Name}】**
- 原文 / Quote：「{引用原句}」
- 建议 / Fix：{一句话具体建议}