humanize

Use this skill when the user wants to generate or optimize Chinese communication copy so it sounds more human, more natural, less templated, and less like po...

This skill appears to do what it claims: it bootstraps a local Python venv, downloads a Hugging Face reranker, invokes a local or host LLM (CoPaw bridge or a local HTTP endpoint), and iteratively optimizes Chinese copy. Before installing or running: - If you care about privacy, avoid passing secrets or sensitive data in the user prompt—the skill's preservation rule explicitly forwards the entire user request verbatim to the local generation/scoring pipeline (and to any configured HTTP endpoint). If you set HUMANIZE_LLM_BASE_URL to a remote service, your input will be sent there. - The bootstrap step downloads models and installs Python packages into a dedicated venv under your CoPaw working dir; ensure you have disk space and trust the source. The repository's default model is from Hugging Face (BAAI/bge-reranker-v2-m3). - The install_to_copaw.py script copies files into ~/.copaw skill_pool and workspace skills and can enable the skill in the workspace manifest—only run that if you want the skill persistently installed in CoPaw. - The skill reads local CoPaw logs to autodetect endpoints and will attempt to call your host active model; this is normal for local integration but means it reads files under your CoPaw working dir. If you want minimal exposure, run the CLI locally in an isolated environment (skip running install_to_copaw.py) and avoid configuring a remote HUMANIZE_LLM_BASE_URL. Review the repository (especially scripts/bootstrap_runtime.py, scripts/install_to_copaw.py, and scripts/local_generation.py) before running to verify it matches your security expectations.

Type: OpenClaw Skill Name: self-evolving-humanize-zh Version: 0.1.6 The skill is a sophisticated Chinese text optimization tool that uses a local scoring model (BGE Reranker) and an iterative refinement loop. It includes automated environment setup (venv creation) and model downloading from Hugging Face (BAAI/bge-reranker-v2-m3), which are standard for local AI tasks. The code logic in scripts like `run_from_brief.py` and `scoring_core.py` is consistent with the stated purpose of 'humanizing' AI-generated text, and no evidence of malicious behavior, data theft, or unauthorized remote access was found.