AI SummaryThis skill is a practical AutoResearch-style loop for one narrow job: optimize Chinese communication copy until it reads more like something a real The user normally only needs to define:
Install
Copy this and paste it into Claude Code, Cursor, or any AI assistant:
I want to install the "humanize" skill in my project. Please run this command in my terminal: # Install skill into your project mkdir -p .claude/skills/humanize && curl --retry 3 --retry-delay 2 --retry-all-errors -o .claude/skills/humanize/SKILL.md "https://raw.githubusercontent.com/TianyiDataScience/humanize/main/SKILL.md" Then restart Claude Code (or reload the window in Cursor) so the skill is picked up.
Description
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 polished AI writing. 中文文案去 AI 味和人味优化 skill，适合自媒体文案、客户邮件、微信回复、售后沟通、面试跟进、上级汇报、产品宣传和社群通知。The user normally only needs to provide the task and constraints. If they also provide an original draft, the skill switches to rewrite mode automatically.
What This Skill Does
This skill is a practical AutoResearch-style loop for one narrow job: optimize Chinese communication copy until it reads more like something a real person would send. The user normally only needs to define: • task: what situation this message is for • constraints: hard limits such as length, phrases to keep, or phrases to avoid Optional: • original draft: only when the user wants rewrite mode instead of generate mode This skill then: • Bootstraps a local runtime and downloads the default local scorer model • Normalizes the user's input into a spec and session mode • Creates a run folder with the spec and drafts • Generates a baseline when the user did not provide one • In rewrite mode, adds direct-rewrite from the current main model to the candidate pool • Generates heuristic challenger drafts such as heuristic-natural and heuristic-balanced • Scores every candidate in one unified pool with the official local scorer • If the best candidate improves but fails the quality gate, repairs best-so-far with direct-repair plus heuristic repair candidates • Persists a small strategy state so the next run starts from the better policy bias • Records each round in JSON and renders a visible report so the process is inspectable V2 product rule: `text humanize is not competing against the main model's direct rewrite. It includes the main model's direct rewrite in the candidate pool, then uses the local scorer + repair loop to choose the steadier version. `
First Run
Before the first evaluation, bootstrap the local runtime: `bash cd {this_skill_dir} && python3 scripts/bootstrap_runtime.py ` This installs a dedicated venv under CoPaw's working directory and downloads the default scorer model: • BAAI/bge-reranker-v2-m3
Inputs You Need From The User
Always collect or infer these before you start iterating: • task • hard_constraints.min_chars / hard_constraints.max_chars when length matters • hard_constraints.must_include when facts must be preserved • hard_constraints.banned_phrases for phrases the user dislikes Optional: • original draft Default assumptions for V1: • goal: built in, unless the user explicitly overrides it • max_rounds: defaults to 3, and stops early when the quality gate passes • style_notes: infer from the task and constraints unless the user adds special tone requirements • session_mode: generate unless the user provides an original draft If the user does not explicitly give a spec file, create one in the run folder. If the user does not provide a custom goal, use the built-in default goal internally: `text 更像真人自然发送的中文沟通消息，减少模板腔、客服腔、公告腔和过度AI润色感。保持清楚、可信、有分寸。 `
Product Semantics
This skill follows one product rule: • the user only needs to express intent and necessary boundaries Split inputs into two layers.
Discussion
Health Signals
My Fox Den
Community Rating
Sign in to rate this booster