humanize
Important: all
scripts/paths are relative to this skill directory. Preferred entrypoint:cd {this_skill_dir} && python3 humanize.py --text "{entire_user_request}" --output-root ./runsLower-level scripts still work via:cd {this_skill_dir} && python3 scripts/...The runtime and model cache live under${COPAW_WORKING_DIR:-~/.copaw}/models/humanize/. In agent shell calls, prefer one-line commands. Do not paste backslash-continued multi-line commands. Always set the shell tool timeout to at least120seconds forpython3 humanize.py. 用户可见输出硬规则：如果python3 humanize.py ...打印=== HUMANIZE_FINAL_RESPONSE_BEGIN ===/=== HUMANIZE_FINAL_RESPONSE_END ===，这两个标记中间的 markdown 就是最终答案。必须原样粘贴给用户，不能总结、不能改写、不能只给最终文案、不能写“已经帮你优化完成了”。用户要看的就是完整过程。 Final relay rule: ifpython3 humanize.py ...prints=== HUMANIZE_FINAL_RESPONSE_BEGIN ===/=== HUMANIZE_FINAL_RESPONSE_END ===, return exactly the markdown between them as the final answer. Do not summarize it, do not paraphrase it, and do not add any explanation. Fallback relay rule: if the shell output is truncated or the final response block is not visible, open the latestuser-visible.mdin the run directory and return that markdown exactly. Invocation rule: do not build helper JSON or temporary Python snippets to call this skill. Invokepython3 humanize.pydirectly. Forbidden invocation: do not callcopaw skills run humanize,python -m skills.humanize..., or any package-style wrapper. Do not pass--mode; rewrite mode is inferred automatically from the full--textrequest or--original. They are not the canonical entrypoint for this skill. Preservation rule: pass the user's full request verbatim via--textby default. If the request contains原文,原稿,正文,draft, or a long draft body, never reinterpret it into separate--task/--constraintsarguments and never drop the original draft. Compatibility rule: this skill is not CoPaw-only. CoPaw / OpenClaw / Hermes style agents should use the sameSKILL.md + python3 humanize.py --text ...flow. The CoPaw installer is only a convenience sync script for the currently known CoPaw workspace path, not a separate skill protocol. Claude Code and other local coding agents can invoke the same CLI after reading thisSKILL.md.