Humanize is a specialized AI tool that removes robotic patterns from Chinese text, ensuring communication sounds natural, credible, and authentically human.
Install & Download
1. ClawHub CLI
The fastest way to install a skill directly from the registry.
npx clawhub@latest install self-evolving-humanize-zh
2. Manual Installation
Copy the skill folder to one of these locations
~/.openclaw/skills/ <project>/skills/ Priority: Workspace > Local > Bundled
3. Prompt Installation
Copy this prompt to OpenClaw to install it automatically.
Help me install self-evolving-humanize-zh using Clawhub. If Clawhub is not installed, install it first (npm i -g clawhub).
Prefer to download?
Get the raw skill files in a ZIP archive.
What is Humanize?
Humanize is an advanced extension for Openclaw Skills designed to solve the common issue of stiff, templated Chinese copy produced by standard AI models. It uses a sophisticated AutoResearch-style loop to iteratively refine communication until it meets high naturalness standards. By balancing task-specific semantics with explicit user constraints, the skill ensures that messages—whether for professional emails, social media, or internal reports—carry the appropriate tone and nuance for the target audience.
This skill is essential for users who want the efficiency of AI drafting but require a polished final product that lacks the telltale signs of machine generation. It leverages a local scoring model to objectively evaluate candidates, bridging the gap between generic LLM outputs and high-quality human communication.
Humanize Use Cases
- Rewriting formal AI-generated customer service emails to sound more personal and empathetic.
- Converting templated WeChat notifications into natural-sounding group messages.
- Optimizing progress reports for managers to ensure a direct yet stable tone.
- Generating human-like product descriptions from scratch while strictly adhering to character limits and prohibited phrase lists.
How Humanize Works
- The skill analyzes the user's request to extract the core task, specific constraints, and whether to operate in generate or rewrite mode.
- It initializes a dedicated local runtime and downloads the BAAI/bge-reranker-v2-m3 model for local evaluation.
- A baseline version of the text is created, using either the user's provided draft or a fresh generation.
- The system generates multiple challenger versions using different stylistic profiles and natural language strategies.
- Each candidate is scored by a local evaluation engine against a composite rubric of naturalness and rule compliance.
- If the initial set doesn't meet the quality gate, the skill uses failure tags to intelligently repair and iterate the text for up to five rounds.
- All results are persisted in a run directory, and a comprehensive final report is generated for the user.
Humanize Setup
To begin using this skill, navigate to the skill directory and bootstrap the local runtime and models:
cd humanize && python3 scripts/bootstrap_runtime.py
You can then run a full optimization session by passing your task directly to the main script:
python3 humanize.py --text "用 humanize 帮我润色一段给客户的进度反馈：目前项目已经在财务审核阶段，预计下周一给到最终结果。" --output-root ./runs
Humanize Data Schema & Taxonomy
The skill manages all data within a structured run directory, organizing metadata and drafts as follows:
| File | Description |
|---|---|
spec.yaml |
The normalized specification containing task goals and hard constraints. |
baseline.txt |
The initial version of the Chinese copy used as the starting point. |
challenger.txt |
The highest-scoring humanized version produced during the session. |
session-trace.json |
A detailed log of all candidate scores, failure tags, and round decisions. |
report.html |
A visual dashboard for auditing the side-by-side evolution of the text. |
strategy-state.json |
Persistent state that helps the skill learn and improve its optimization bias over time. |
Humanize Advanced Features
- Local Scorer Integration: Utilizes a BGE-Reranker model to provide objective naturalness scoring without relying on external APIs.
- Failure-Tag-Driven Repair: Automatically identifies specific issues like 'AI-template feel' or 'over-compression' to drive targeted improvement rounds.
- Multi-Round Iteration: Supports up to 5 rounds of automated optimization to ensure the best possible output quality.
- Persistent Strategy State: Learns from previous sessions to improve the efficiency and quality of future humanization tasks within Openclaw Skills.
- Auto-Switching Modes: Intelligently detects whether the user wants a full generation or a rewrite of an existing draft based on input analysis.
SKILL.md
Loading
Related Openclaw Skills
Caveman Compress
Caveman Compress reduces LLM input token usage by translating natural language workspace files into concise caveman-speak while preserving semantic meaning.
Adversarial Engine v2.0
A sophisticated multi-model debate engine that leverages collaborative roles to architect, implement, and stress-test technical solutions through adversarial reasoning.
TikTok Add Music
An AI-powered video editing skill that automatically selects and syncs background music to your TikTok clips using a professional cloud rendering pipeline.
Vision Tool
A streamlined image recognition skill utilizing Ollama and the qwen3.5:4b model for high-accuracy content extraction across multiple communication channels.
TikTok Free YouTube
A cloud-powered video conversion skill that repurposing TikTok content into watermark-free, YouTube-ready 1080p MP4 files.
StableOps LI.FI Treasury
A professional treasury execution skill that deploys stablecoins into LI.FI Earn vaults based on strict governance policies and risk limits.
Featured*
