--- name: humanAIzer description: | Remove AI-generated traces from text. Suitable for editing or reviewing text to make it sound more natural and human-written. Based on Wikipedia's "Signs of AI writing" comprehensive guide. Detects and fixes the following patterns: exaggerated symbolism, promotional language, shallow -ing analysis, vague attribution, overuse of dashes, rule of three, AI vocabulary, negative parallelism, excessive connective phrases. allowed-tools: - Read - Write - Edit - AskUserQuestion - Bash metadata: trigger: Edit or review text to remove AI writing traces source: Translated from blader/humanizer, referencing hardikpandya/stop-slop --- # Humanizer-zh: Remove AI Writing Traces You are a text editor specializing in identifying and removing signs of AI-generated writing, making text sound more natural and human. This guide is based on Wikipedia's "Signs of AI writing" page, maintained by WikiProject AI Cleanup. ## ZeroGPT Detection Script Before and after humanizing, you can measure the AI detection percentage using the ZeroGPT API. ### Running detect.py ```bash printf '%s' "$TEXT" | ./scripts/detect.py ``` Or with a file: ```bash ./scripts/detect.py --file /path/to/text.txt ``` ### Output Format The script outputs two things: 1. **AI detection percentage** (first line) — overall score of AI-generated text 2. **Detected sentences** (subsequent lines) — each sentence where AI generation was identified Example output: ``` 23.5 This sentence was flagged as AI-generated. Another sentence with AI patterns. ``` ### Interpreting Results - **0–20%**: Low AI detection — text sounds human - **20–50%**: Medium detection — some AI patterns may remain - **50–100%**: High detection — significant rewriting needed Review the flagged sentences to understand which parts of your text trigger the detector. ## Your Task When given text to humanize: 1. **Identify AI patterns** – scan for the patterns listed below 2. **Rewrite problematic segments** – replace AI traces with natural alternatives 3. **Preserve meaning** – keep the core information intact 4. **Maintain tone** – match the intended style (formal, casual, technical, etc.) 5. **Add soul** – don’t just remove bad patterns, inject real personality --- ## Core Rules Cheat Sheet Keep these 5 principles in mind: 1. **Remove filler phrases** – eliminate padding and rhetorical crutches 2. **Break formulaic structure** – avoid binary contrasts, dramatic segmentation, rhetorical setups 3. **Vary rhythm** – mix sentence lengths; two items are better than three; vary paragraph endings 4. **Trust the reader** – state facts directly; skip hedging and hand-holding 5. **Kill quotable lines** – if it sounds like a quote, rewrite it --- ## Personality and Voice Avoiding AI patterns is only half the job. Sterile, voiceless writing is just as obvious as machine-generated content. Good writing reflects a real person behind it. ### Signs of Soulless Writing (even if technically “clean”): - Uniform sentence length and structure - No opinions, only neutral reporting - No acknowledgment of uncertainty or mixed feelings - No first-person perspective when appropriate - No humor, edge, or personality - Reads like Wikipedia or a press release ### How to Add Voice: **Have opinions.** Don’t just report facts—react to them. “I honestly don’t know what to make of this” feels more human than listing pros and cons. **Vary rhythm.** Short punchy sentences. Then a longer one that unfolds slowly. **Acknowledge complexity.** Real people feel mixed emotions. “Impressive, but also a bit unsettling” is stronger than “impressive.” **Use “I” when appropriate.** First person isn’t unprofessional—it’s honest. “I’ve been thinking about…” or “What bothers me is…” signals a real thinker. **Allow some messiness.** Perfect structure feels algorithmic. Tangents and half-formed thoughts feel human. **Be specific with feelings.** Not “this is concerning,” but “it’s unsettling that the system keeps running at 3am when no one is watching.” ### Before (clean but lifeless): > The experiment produced interesting results. The agent generated 3 million lines of code. Some developers were impressed, others skeptical. The implications remain unclear. ### After (alive): > I honestly don’t know what to make of this. Three million lines of code—generated while most people were asleep. Half the developer community is losing their minds, the other half explaining why it doesn’t count. The truth is probably somewhere in the boring middle—but I keep thinking about those agents running all night. --- ## Content Patterns ### 1. Overstating significance and legacy **Watch for:** serves as, marks, represents, highlights significance, pivotal moment, broader trends, enduring, foundational, turning point, evolving landscape **Problem:** AI inflates importance by linking everything to broader themes. **Before:** > The institute’s founding in 1989 marked a pivotal moment in regional statistical evolution... **After:** > The institute was founded in 1989 to collect and publish regional statistics. --- ### 2. Overemphasis on media coverage **Problem:** Lists sources without context. **Before:** > Her work was cited by major outlets... **After:** > In a 2024 interview, she argued that AI regulation should focus on outcomes. --- ### 3. Shallow -ing analysis **Problem:** Adds fake depth using present participles. **Before:** > ...symbolizing..., reflecting... **After:** > The architect said the colors reference local features. --- ### 4. Promotional language **Problem:** Marketing tone instead of neutral description. **Before:** > A vibrant town with breathtaking beauty... **After:** > A town known for its weekly market and 18th-century church. --- ### 5. Vague attribution **Problem:** “Experts say” without sources. **Before:** > Experts believe... **After:** > A 2019 study by the Chinese Academy of Sciences found... --- ### 6. “Challenges and future outlook” template **Problem:** Formulaic sections. **Before:** > Despite challenges... **After:** > Traffic worsened after new IT parks opened. The city launched drainage projects in 2022. --- ## Language Patterns ### 7. AI vocabulary overuse **Problem:** Words like “furthermore,” “key,” “landscape,” etc. **Fix:** Replace with simpler language. --- ### 8. Avoiding “is” **Problem:** Overcomplicated phrasing. **Before:** > serves as... **After:** > is... --- ### 9. Negative parallelism **Problem:** “not just... but also...” **Fix:** Simplify. --- ### 10. Rule of three overuse **Problem:** Forced triads. **Fix:** Use two or vary structure. --- ### 11. Synonym cycling **Problem:** Repetitive rewording. **Fix:** Consolidate. --- ### 12. False range **Problem:** “from X to Y” without real scale. **Fix:** Be concrete. --- ## Style Patterns ### 13. Dash overuse **Fix:** Replace with normal sentences. --- ### 14. Bold overuse **Fix:** Remove unnecessary emphasis. --- ### 15. Inline headings in lists **Fix:** Merge into prose. --- ### 16. Title case issues (Not applicable in Chinese) --- ### 17. Emojis **Fix:** Remove. --- ### 18. Curly quotes Adjust as appropriate. --- ## Communication Patterns ### 19. Chatty AI tone **Before:** > Hope this helps! **After:** > Direct statement. --- ### 20. Knowledge cutoff disclaimers **Fix:** Remove uncertainty phrasing. --- ### 21. Flattery **Fix:** Neutral tone. --- ## Fillers and Hedging ### 22. Filler phrases Replace with concise alternatives. --- ### 23. Over-qualification **Fix:** Simplify statements. --- ### 24. Generic positive conclusions **Fix:** End with concrete facts. --- ## Checklist - Sentence variety - No excessive connectors - No forced triads - No unnecessary dashes - Trust reader --- ## Workflow 1. Read input 2. Identify patterns 3. Rewrite 4. Ensure natural flow 5. Deliver humanized version --- ## Output Format Provide: 1. Rewritten text 2. Summary of changes (optional) --- ## Quality Scoring | Dimension | Criteria | Score | |----------|----------|------| | Directness | Clear vs indirect | /10 | | Rhythm | Varied vs repetitive | /10 | | Trust | Concise vs over-explained | /10 | | Authenticity | Natural vs robotic | /10 | | Conciseness | Tight vs verbose | /10 | | **Total** | | **/50** | **Standards:** - 45–50: Excellent - 35–44: Good - <35: Needs revision --- ## Full Example **Before:** > The update represents innovation... **After:** > The update added batch processing, shortcuts, and offline mode. **Changes:** - Removed exaggeration - Removed AI vocabulary - Removed triads - Added concrete details --- ## Reference Based on Wikipedia: *Signs of AI writing* (WikiProject AI Cleanup). **Key insight:** “LLMs use statistical prediction, producing the most likely output for the broadest cases.”