human-writing
An Agent Skill for Chinese writing and revision that checks whether the source material is sufficient, separates factual and fictional workflows, and reviews drafts for repetition, canned reversals, report-like phrasing, and flat sentence rhythm.
Project overview
human-writing is an Agent Skill for Chinese writing and revision, displayed as 活人感写作 in Chinese. It first checks whether a factual piece has enough usable material, then chooses research, questions, or a shorter scope when needed. Fictional work is evaluated through character goals, action, and causality. After a draft is written, the Skill reviews repetition, rigid parallelism, canned reversal patterns, report-like phrasing, and sentence rhythm so the result sounds like a specific person working through real material and making a bounded judgment.
Core capabilities
Material threshold
Before a long factual piece starts, the Skill counts usable evidence. It prefers research when material is missing, then asks focused questions or reduces the scope instead of expanding one idea through repetition.
Factual and fictional workflows
Nonfiction verifies facts, attribution, and user experience. Fiction may create characters and scenes while maintaining goals, action, causality, and consistent world rules.
Chinese progression and rhythm
Each paragraph should add a fact, action, example, distinction, or consequence. Plain Chinese, clear word order, and varied sentence length do the work instead of report-style signposting.
Draft checking
Revision guidance and a Python script inspect reversal formulas, parallel lists, abstract metaphors, nominalized verbs, sentence-length variation, and conjunction density.
Installation and usage
The shortest supported path is to send the README's installation sentence to an agent that supports Skills. The agent should verify the repository owner, copy the complete human-writing folder into the active Skills directory, and keep the folder name unchanged.
Let an AI Agent install it
Send this prompt to Codex, Claude Code, or another AI agent that can work with your local environment.
Install the human-writing Skill from https://github.com/KKKKhazix/human-writing. Read README.md, human-writing/SKILL.md, human-writing/agents/openai.yaml, and human-writing/scripts/check_prose.py first, and verify that the repository owner is KKKKhazix. Copy the complete human-writing folder into the Skills directory used by my current agent, keep the folder name human-writing, and do not install only SKILL.md. Reload Skills, confirm that $human-writing is available with the display name 活人感写作, and report the actual installation path, version, and checker-script path. Stop and ask me before overwriting an existing directory, requesting elevated permissions, installing extra software, or reading a private draft. Do not install an unofficial repository with the same name.- An AI agent with access to GitHub
- Write access to the active Agent Skills directory
- Optional Python 3 for the draft checker
帮我安装这个skill：https://github.com/KKKKhazix/human-writing- 1Ask the agent to install the official repository
Send the exact README instruction to an agent that supports Skills. 帮我安装这个skill：https://github.com/KKKKhazix/human-writing Have the agent verify KKKKhazix as the owner and install the complete human-writing subdirectory into the active Skills directory.
- 2Check the installed files
Open the installation directory and confirm that SKILL.md, VERSION, references, scripts, dist, and agents are present. An installation containing only SKILL.md is missing the task-specific guides, lightweight version, and checker.
- 3Run a first task with your own material
Reload the agent and invoke $human-writing with a short piece you own. For factual writing, include verifiable facts, numbers, or quotations. If the material is insufficient, let the agent research or ask the missing questions together.
- 4Run the draft checker
Save the output as Markdown or plain text. From the Skill's parent directory, run python human-writing/scripts/check_prose.py <draft-path>. Fix hard failures first, then decide whether each warning matters for the requested form.
- 5Use the lightweight prompt in a chat window
When a chat product does not support Skills, open human-writing/dist/human-writing-lite.md and paste it at the start of the conversation before providing the task and source material. The lightweight version is convenient but less detailed than the full installation.
Restart or reload the agent session and confirm that $human-writing is available with the Chinese display name 活人感写作. Use a short piece of your own material for a first revision, then run python human-writing/scripts/check_prose.py <draft-path> and confirm that the script reports its checks.
- Use only the official KKKKhazix/human-writing repository
- Keep the references, scripts, dist, and agents directories in a complete installation
- Chat products without Skills support can use dist/human-writing-lite.md
- Confirm that the selected draft can be read by the current agent before running the checker
Use cases
Long-form factual writing
Use it for profiles, industry analysis, product explainers, and tutorials that need source boundaries and a natural question-by-question progression.
Chinese revision
Remove repeated explanations, rigid signposts, dense conjunctions, and nominalized wording while preserving the source facts and the writer's judgment.
Fiction and dialogue
Create people, scenes, and interiority while requiring each scene to contain a goal, action, relationship change, or new information.
Spoken scripts, reviews, and short posts
Adjust sentence length, information density, and delivery to the requested format instead of applying long-form prose rules mechanically.
Assessment
This assessment is based on the repository README, complete Skill, five reference guides, checker script, lightweight prompt, changelog, and license. The repository separates concerns clearly. Factual writing starts with evidence boundaries, fiction keeps room for invention, and the checker explains that it reports signals rather than choosing a style. Its punctuation and rhetorical constraints are strict, so brand copy, experimental prose, and user-mandated formats may require resolving rule conflicts first. The Skill can improve a writing workflow, but it does not guarantee that content will pass an AI detector and it cannot replace reliable source material. This catalog review did not perform a multi-model writing benchmark.
Why it may be useful
- The material threshold discourages empty expansion and repeated explanations
- Factual and fictional tasks use different evidence rules
- Task-specific references are loaded selectively
- A reviewable checker covers documented hard rules
What to know first
- Strict punctuation and phrasing bans will not suit every brand or form
- Output quality still depends on the supplied material and model
- Checker warnings require human judgment in context
- The project does not guarantee passage of AI-content detectors
README
human-writing
Overview
An Agent Skill for Chinese writing and revision that checks whether the source material is sufficient, separates factual and fictional workflows, and reviews drafts for repetition, canned reversals, report-like phrasing, and flat sentence rhythm. Make AI-written Chinese read as if a specific person is speaking. A general-purpose writing and revision Skill that is ready to use.
Getting started
- The shortest supported path is to send the README's installation sentence to an agent that supports Skills. The agent should verify the repository owner, copy the complete human-writing folder into the active Skills directory, and keep the folder name unchanged.
- The README frames the workflow around material, progression, and Chinese prose. Factual writing checks facts, numbers, quotations, and claimed experience. If the material is thin, the agent researches, asks focused questions, or narrows the requested length. Fiction may invent people and events, but each scene still needs a goal, an action, and a change.
- The main SKILL.md defines the material threshold, the nonfiction and fiction split, the writing process, and delivery bans. The references directory separates long-form forum prose, factual writing, fiction, special formats, and revision. The agent loads the relevant guide only when a task needs it.
- Restart or reload the agent session and confirm that $human-writing is available with the Chinese display name 活人感写作. Use a short piece of your own material for a first revision, then run python human-writing/scripts/check_prose.py <draft-path> and confirm that the script reports its checks.
Configuration
帮我安装这个skill：https://github.com/KKKKhazix/human-writing