Humanize-It: Iterative AI Trace Removal Rewriting
Rewrite specified documents to remove AI-generated traces. Based on the text type and content, automatically select the most suitable one from three sub-skills to start rewriting. If the rewriting result is unsatisfactory, switch to another skill and continue iterating until the result meets the standard or reaches the maximum limit of 42 iterations.
Sub-Skill Capability Matrix
| Skill | Applicable Scenarios | Core Capabilities | Rewriting Style |
|---|
| humanizer-zh | General texts, articles, blogs, copywriting | Identifies 24+ AI writing patterns, injects personality and soul, varies rhythm | Natural, warm, opinionated |
| humanize-chinese | General + academic + long texts | 20+ rule checks + statistical features + LR fusion scoring, CLI toolchain | Multiple styles available (Zhihu/Xiaohongshu/academic/literary, etc.) |
| technical-writing | Technical documents, architecture descriptions, design drafts, review documents | Removes technical jargon, prioritizes evidence, eliminates host-style tone | Plain, rigorous, verifiable |
Workflow
Step 1: Read and Analyze the Document
- Read the content of the document specified by the user
- Determine the document type:
- Technical documents (technical solutions, architecture descriptions, design documents, review drafts) → Prioritize using
- Academic papers (papers, research articles) → Prioritize using (academic mode)
- General texts (blogs, copywriting, articles) → Prioritize using
- Long texts (≥1500 words) → Prioritize using (longform mode)
Step 2: Select Rewriting Strategy
Select the first rewriting skill based on the document type:
Document type judgment:
├── Technical documents → technical-writing → humanizer-zh → humanize-chinese
├── Academic papers → humanize-chinese(academic) → humanizer-zh → technical-writing
├── General texts → humanizer-zh → humanize-chinese → technical-writing
└── Long texts(≥1500 words) → humanize-chinese(longform) → humanizer-zh → technical-writing
Step 3: Iterative Rewriting Loop
iteration = 0
MAX_ITERATIONS = 42
while iteration < MAX_ITERATIONS:
iteration += 1
# Rewrite using current skill
result = humanize(current_text, current_skill)
# Evaluate rewriting effect
score = evaluate(result)
if score >= PASS_THRESHOLD:
# Rewriting effect meets standard, output final result
output(result)
break
# Effect does not meet standard, switch to next skill
current_skill = next_skill(skill_order)
current_text = result # Continue optimizing based on last rewrite result
if all_skills_exhausted():
# All skills have been used in one round, start a new round of combination from the beginning
current_skill = first_skill(skill_order)
Step 4: Quality Evaluation
After each rewrite, evaluate the effect based on the following dimensions (100-point scale):
| Dimension | Weight | Evaluation Criteria |
|---|
| AI Trace Removal | 30% | Whether three-part structures, clichés, and mechanical conjunctions are eliminated |
| Naturalness | 25% | Whether it reads like human writing, with natural rhythm |
| Information Completeness | 20% | Whether core information is retained without losing key content |
| Style Consistency | 15% | Whether tone is consistent throughout and matches the document type |
| Readability | 10% | Whether sentences are smooth and logic is clear |
Scoring Standards:
- ≥ 80 points: Pass, output result
- 60-79 points: Acceptable, iterate one more round to see if improvement is possible
- < 60 points: Not up to standard, must continue rewriting
Step 5: Output Result
Output the final rewritten result, accompanied by:
- Complete rewritten text (written to the original file or specified output file)
- Rewrite summary (which skills were used, number of iterations, final score)
- List of main changes
Specific Execution Instructions
When Calling humanizer-zh
Use the Skill tool to call
, pass in the document content, and let it detect and rewrite according to 24 AI writing patterns. Focus on:
- Delete filler phrases
- Break formulaic structures
- Vary rhythm
- Trust readers
- Delete golden sentences
- Inject personality and soul
Preliminary evaluation is conducted using its 50-point quality scoring system (≥ 40 points is considered up to standard).
When Calling humanize-chinese
Use the Skill tool to call
, select the corresponding mode based on the document type:
- General texts: Use detect + rewrite process
- Academic papers: Use academic mode
- Long texts: Use or
If the CLI tool is available, prioritize using CLI for quantitative detection and rewriting; otherwise, manually execute according to its LLM usage guidelines.
Evaluation is conducted using its 0-100 scoring system (≤ 35 points is in the LOW range, considered up to standard).
When Calling technical-writing
Use the Skill tool to call
, mainly for rewriting technical documents. Focus on:
- Remove words from the buzzword blacklist
- Eliminate host-style and evaluative tones
- Rewrite sentences according to "Condition → Object → Judgment"
- Prioritize evidence, reduce unsupported strong judgments
- Remove transitional sentences
Iteration Strategy Details
When to Switch Skills
- No score improvement after 3 consecutive rewrites with the same skill → Switch to the next skill
- Score decreases → Roll back to the previous version and switch skill
- All skills have been used in one round but not up to standard → Restart from the first skill, using the rewritten result from the previous round as input
- Reach 42 iterations → Output the current best result
Combination Strategies
Different skills focus on different aspects of rewriting, and combining them can complement each other:
- humanizer-zh → humanize-chinese: First inject personality and soul, then perform quantitative AI pattern removal
- humanize-chinese → humanizer-zh: First perform systematic AI trace removal, then add warmth
- technical-writing → humanizer-zh: First ensure technical rigor, then increase naturalness
- humanizer-zh → technical-writing: First remove general AI traces, then adjust technical tone
User Interaction
If the user provides clear preferences (such as "maintain academic style", "be more colloquial", "technical documents should not be too casual"), follow the user's preferences during rewriting and prioritize selecting the corresponding skill.
If the user does not specify the document type, you can ask the user via AskUserQuestion:
- Document type (general / technical / academic)
- Desired rewriting style (natural and casual / plain and rigorous / academic standard)
- Whether to write to the original file or output to a new file