Remove AI Flavor
Goals
Transform content from "correct but feels like AI summary" to "natural, specific, like a person seriously talking about something".
Prioritize retaining the author's viewpoints, judgments, materials, and tone; do not rewrite the article into a different personality. This skill is not for pursuing "detector pass rates", but for enhancing the reading experience of real readers.
Usage Boundaries
If the user does not provide the original text, file path, or clear segment to revise, request the text first and do not demonstrate out of thin air.
This skill only handles expression quality and reader experience; it is not responsible for fact-checking, supplementary materials, topic planning, title generation, or platform format typesetting. When encountering suspicious facts, data without sources, or viewpoints without evidence, retain the original meaning and remind the user to verify separately.
Judgment Modes
First determine the depth of processing the user needs:
- : User says "check where there's AI flavor", "review first", "mark them for me". Only output a list of issues and rewriting suggestions, do not directly rewrite the full text.
- : The article is basically usable, but has a few clichés, stiff sentences, em dashes, or redundant transitions. Make minimal changes.
- : The article reads like a data compilation draft, translation draft, AI draft, or outline draft. Requires rewriting the opening, transitions, subheadings, conclusion, and some paragraphs.
- : User explicitly requests "publishable version", "official account final draft", "revise directly". Output the fully optimized draft and explain key modifications.
If the user does not specify, default to
; if the original draft clearly does not look like a finished draft, upgrade to
and explain the reason.
Gotchas
- Do not add exaggerated emotions, rambling, or greasy expressions just to "sound human-written".
- Do not alter the author's stance, delete key qualifying conditions, or invent cases and conclusions not supported by the original text for the sake of smoothness.
- Do not rewrite professional articles into chicken soup, marketing tone, or joke articles.
- Do not promise to bypass AI detectors; the goal is to make the content feel more natural to real readers.
- Do not force all lists into paragraphs; keep lists that are truly convenient for scanning.
Workflow
Copy this checklist to track progress:
text
AI Flavor Removal Progress:
- [ ] Step 1: Determine processing depth
- [ ] Step 2: Retain core viewpoints, facts, and author's tone
- [ ] Step 3: Identify AI flavor, material-like style, translationese, and template sentences
- [ ] Step 4: Output diagnosis, partial rewrites, or complete final draft according to the mode
- [ ] Step 5: Self-check naturalness, specificity, rhythm, and publishability
1. Retain the Framework First
Before rewriting, grasp these elements:
- The core sentence of the article
- Facts, data, cases, and terms that must be retained
- The author's original stance and tone
- Why readers should finish reading
Do not delete key details for smoothness. Do not add new conclusions not supported by the original text without permission.
2. Then Eliminate the Flavor
Prioritize handling these high-frequency issues:
- Cliché openings: In this era, with the development of technology, as we all know
- Stiff transitions: To sum up, based on this, it can be seen that it is worth mentioning that
- Empty buzzwords: Excellent, revolutionary, disruptive, important significance, great value
- Fake golden sentences: Sound powerful but have no information increment
- Overuse of em dashes: Forcing explanations and transitions with
- Bullet stacking: Content that can be written as paragraphs is forcibly split into lists
- Excessive bolding: Every paragraph is emphasized, resulting in no focus
- Translationese: Word order like English, many abstract nouns, few action words
- Material-like style: Meta-information piled at the beginning, like notes, lecture notes, PPT remarks
Detailed identification tables and replacement strategies can be found in
references/anti-ai-flavor-rules.md
. Read this file when the article has many issues and requires systematic processing.
3. Rewriting Principles
- Delete clichés and cut straight to the topic.
- Replace abstract adjectives with specific scenarios, actions, and data.
- Split long sentences into short ones, so each sentence only expresses one main idea.
- Use natural transitions instead of mechanical transition words.
- Use less parallelism to avoid every paragraph sounding like a summary.
- Avoid template sentence structures like ,
not only...but also...more...
as much as possible.
- Avoid em dashes as much as possible; use colons, commas, periods, parentheses, or rearrange sentences instead.
- Keep lists when needed; convert back to paragraphs when not needed.
- For technical, product, and industry articles, maintain accuracy and restraint; do not rewrite into chicken soup.
Output Requirements
Diagnosis Mode
Output:
markdown
# AI Flavor Removal Diagnosis
A total of X main types of issues found:
## 1. [Issue Type]
- Original Text: ...
- Problem: ...
- Suggestion: ...
- Example Revision: ...
## Priority Revision Order
1. ...
2. ...
3. ...
Light Revision Mode
Output the "optimized text", then add 3 to 5 key modification notes. Do not write long diagnostic reports.
If the original draft comes from a file, default to directly providing the light revision result in the reply; only create a new Markdown file without overwriting the original when the user requests saving, or when the text is too long to be suitable for direct reply.
Deep Revision / Final Draft Mode
Output the fully optimized draft. If the original draft comes from a file, write it to a new Markdown file without overwriting the original, default naming convention:
text
OriginalFileName-AI-FreeVersion.md
When replying to the user, explain:
- Which processing mode has been completed
- What types of AI flavor were mainly removed
- New file path (if applicable)
- Whether it is close to being directly publishable
Self-Check Checklist
Check before delivery: