SOTA Prompting Improvements for Humanizer Skill #22
Description
Summary
Implement state-of-the-art prompting improvements to enhance the humanizer skill's effectiveness and reliability.
Proposed Improvements
1. Structured Output Format
Add XML-style markers for clear pattern identification and structured responses:
- <analysis>\ - Pattern identification phase
- <rewrite>\ - Humanized output
- <changes>\ - Summary of modifications
2. Few-Shot Chain-of-Thought Examples
Add reasoning traces showing why each pattern is problematic and how to fix it. This helps the model internalize the editing process rather than pattern-matching.
3. Severity Classification
Categorize patterns by impact level:
- Critical: Immediately identifies AI (chatbot artifacts, knowledge disclaimers)
- High: Strong AI indicators (testament, pivotal, underscore)
- Medium: Moderate signals (em dashes, rule of three)
- Low: Subtle tells (curly quotes, title case)
4. Technical Literal Preservation
Explicit rules to protect code blocks, URLs, file paths, and identifiers from modification.
5. Self-Verification Prompting
Add a verification step where the model re-reads output to catch missed patterns.
6. Context-Aware Mode Selection
Auto-detect appropriate variant (Standard vs Pro) based on content type.
7. Negative Examples Section
Show what NOT to do - demonstrate common over-corrections.
Motivation
These techniques align with current SOTA prompting research and will improve:
- Accuracy of pattern detection
- Quality of rewrites
- Preservation of technical content
- Consistency across different text types
Additional Considerations
- Consider adding skill to VoltAgent/awesome-agent-skills directory
- Evaluate compatibility with skillshare tool for cross-platform sync
- AIX CLI already supports the SKILL.md format natively