Human Writing
Makes AI-written Chinese read like a specific person talking, fixing prose that flows but has no recognizable author.
Pure prompt files plus a local check script; no network calls, no sensitive data handling, no external side effects; explicit non-goals (no author profiles, no skill creation). Deducted because the publisher is unverified and there is no signing/checksum for provenance.
Instructions are self-consistent with staged progressive loading, and check_prose.py is well structured with readable error handling. Deducted for the absence of tests and any execution evidence; script behavior on edge inputs can only be inferred statically, and static review caps this at 10.
Scenarios, genre list, fact/fiction boundary and trigger conditions (load-per-task) are unusually explicit, Chinese-native, and free of overseas-service dependencies. Deducted for limited counter-example documentation and single-file agent integration evidence.
Clear layered docs with progressive disclosure, MIT license, version badge and changelog reference, a lite dist build, and an Issue-based maintenance path. Deducted because changelog accuracy and actual maintenance responsiveness cannot be verified statically, and copyright holder is a generic 'contributors'.
Rules for de-AI-flavored Chinese are concrete and actionable with paired examples and a runnable checker, giving plausible marginal value over manual editing. Deducted because output quality and the claimed detection improvements are unverified in a static review; capped at 7.
The repo supplies auditable primary material (full rule texts and script source), and the claimed hard-banword consistency between revision.md and the script is statically checkable. Deducted for no tests, no independent reproduction, and no third-party corroboration.
- This is a static source review; no script was executed and effectiveness claims are unverified.
- check_prose.py has no automated tests; false positives/negatives (especially variant pivot patterns and the 不只…还… heuristic) require user judgment.
- Publisher is unverified; verify repository origin and release tags before installation.
- Soft bans and context-word checks rely heavily on subjective judgment; weaker models may over-apply hard bans and produce stilted prose.
What does this skill do, and when should you use it?
A general-purpose Chinese writing and revision skill covering Zhihu answers, WeChat articles, long forum posts, fiction, speech scripts and most other Chinese writing scenarios. Before drafting it checks whether you have enough material: for factual topics, insufficient material triggers research, follow-up questions, or a shorter piece — never filler. Fiction may invent freely, but every scene needs action and change. After drafting, it enforces hard bans on colon misuse, dashes, the 'not X, but Y' reversal pattern, and common AI jargon.
On invocation it routes to task-specific reference files: forum-prose.md for long-form posts, reality.md for factual topics, fiction.md for fictional work, formats.md for special forms, with revision.md loaded only after the first draft. It enforces a material threshold (non-fiction long pieces require at least five sourced, concrete items); if short, it researches, asks at most three questions, narrows the topic, or delivers a short answer. It prescribes concrete craft rules for narrative stance, paragraph progression and sentence rhythm, then runs post-draft hard-ban checks (no reversal rhetoric, no three-item parallel structures, no dashes, no expository colons, no business jargon). Long drafts can be checked with scripts/check_prose.py until violations reach zero. A ~2,000-character distilled version, human-writing-lite.md, can be pasted directly into chat windows.
- Content writers whose AI drafts of Zhihu answers or WeChat articles read like institutional reports and want natural Chinese with personal judgment and detail.
- Writers of real-person stories, historical narratives or industry analyses who need mandatory verification of facts, quotes and data, with no invented details.
- Fiction authors using AI for drafts who require each scene to carry character goals, actions and plot progression rather than scenery padding.
- Writers of speech scripts, tutorials or reviews needing form-specific adjustments in tone and pacing.
- Writers sensitive to AI-flavored prose who want a script to automatically catch dashes, reversal sentence patterns and model-typical jargon.
What are this skill's strengths and limitations?
- Rules are highly specific and executable: material-count thresholds, on-demand reference loading, and an explicit post-draft ban list — no vague style descriptions.
- Version 1.1.0 bans rhetorical moves rather than literal strings, catching reversal patterns in new wording, with an accompanying verification script.
- Clean separation of factual vs. fictional obligations: real-world drafts must verify sources; fiction may invent but must not fake provenance.
- MIT-licensed, with a ~2,000-character distilled version usable in chat windows without any installation.
- Rules are quite strict (no colons, no dashes, no reversal patterns); writers who want those devices will feel constrained, and the script enforces only the stated hard bans.
- The material threshold for factual topics may make the AI refuse your requested word count and deliver something shorter — unfriendly when you need to hit a length.
- No cross-model test evidence is provided in the repo; the README merely invites Issue reports when behavior looks wrong.
How do you install this skill?
Send this to your Agent: 帮我安装这个skill：https://github.com/KKKKhazix/human-writing. The Agent reads the repo and completes installation; the skill shows up as「活人感写作」. If your Agent cannot install directly, download from GitHub Releases or copy the human-writing folder from the repo into your local Skills directory (~/.agents/skills/human-writing/), keeping the folder name human-writing.
How do you use this skill?
After installation, trigger it in conversation, e.g.: 使用 $human-writing，把我的材料写成一篇有活人感和中文韵律的作品. Provide your material or topic; the skill runs a material check first and may ask up to three follow-up questions. After drafting a long piece, run python scripts/check_prose.py <draft-path> to check for violations. Without installing, you can paste dist/human-writing-lite.md directly into ChatGPT, Qwen or similar chat windows.