活人感写作 · Human-Feel Chinese Writing
A generative method for writing Chinese long-form prose that reads as a real person talking — a material-first gate (hold at least five concrete pieces of material that chain into an actual process before writing long, else research/ask/shorten rather than pad), a five-question speaking-position frame, first-draft craft principles, and a revision pass that bans the rhetorical moves of AI Chinese, not just their literal wording.
The thing that makes AI-written Chinese feel lifeless isn't vocabulary — it's padding three thin ideas into a thousand words and leaning on moves like 不是…而是 reversals and abstract-noun lyricism; writing that feels human starts one step earlier, by refusing to write long until you actually hold enough concrete material to say something.
Who it's for: Chinese-language writers drafting 知乎 answers, 公众号 articles, blog posts, 人物 and industry pieces, or fiction, content creators told their writing 'AI 味太重' who want a generative method rather than a cleanup checklist, bilingual teams needing a Chinese long-form writing discipline, anyone writing Chinese prose that should read as a specific person with real material behind them
Example
"帮我写一篇一千五百字的行业解读" → A material gate first: if only three abstract ideas are on hand and no concrete process, Claude researches or asks at most three questions rather than padding to length — and if it still can't reach five chainable pieces of material, it delivers an honestly shorter, denser piece instead; then it drafts from a specific speaker and real material, and strips AI-Chinese rhetorical moves (不是…而是 setups, signpost colons, abstract-noun lyricism) at revision
CLAUDE.md Template
New here? 3-minute setup guide → | Already set up? Copy the template below.
# 活人感写作 · Human-Feel Chinese Writing
通用中文创作与改稿方法。用于知乎回答、论坛长帖、公众号文章、博客、人物故事、历史叙事、新闻与行业解读、科普、教程、评测、个人叙事以及小说与剧本。默认把文章写成一个见过事、查过材料、愿意把来龙去脉讲清楚的人在说话——保留中文长回答的"活人感"和自然韵律,避免机构腔、演说腔、营销腔和模型腔。
核心次序:**活人感首先来自材料,其次来自说话位置,最后才是语气。** 不要把活人感理解成口头禅、粗口、错别字或网络梗。
## 第一关:先看作品靠什么站住(先于提纲和动笔)
用户要求的字数不能跳过这一步。现实作品靠材料,虚构作品靠事件。
**非虚构长文计划达到约一千二百字时,先在内部逐条写出至少五件具体材料**,并注明每件来自用户哪句话或哪份可靠来源。只写一个概括性的类别("用户觉得方便""可以保留状态")不算一件材料;五件材料还要能串成一条实际过程,不能是五句相邻的道理。
Get new playbooks like this one
One email a week with new Claude Code workflows. Free, like everything here.
No spam. Unsubscribe anytime.
What This Does
A generative method for writing Chinese long-form prose (知乎 answers, 论坛 long posts, 公众号 articles, blogs, 人物 and industry pieces, 科普, tutorials, personal narrative, and fiction) so that it reads as a specific, real person talking rather than a model generating. Its organizing claim is an order of operations most "sound more human" advice skips: the human feel comes first from material, second from the speaking position, and only last from tone — so it refuses to treat the problem as a surface-polish task. The load-bearing move is a material-first gate: before writing anything long, you must hold at least five concrete pieces of material that chain into an actual process, each traceable to something the user said or a verifiable source. If you can't, the method forces a real choice — research it, ask at most three pointed questions, or write something honestly shorter — rather than the default failure mode of padding three thin ideas out to the requested word count with invented examples and the same point explained five ways.
Once material is in hand, it supplies a five-question speaking-position frame (who's talking and what they actually know versus guess; what prompted them; what concrete material backs the piece; what they judge and on what basis; what the reader will ask next) and a set of first-draft craft principles (hit the subject fast, advance every paragraph with something genuinely new, let detail carry emotion instead of explaining it). Only at revision does it apply its anti-AI-slop list — and the key subtlety is that it bans the rhetorical move, not the literal phrase: the 不是…而是 reversal setup, four-part parallel restraint, abstract-noun-plus-concrete-verb lyricism, nominalized verbs, em-dashes, signpost colons, and metaphor-wrapped abstractions are all prohibited however they're reworded, because swapping the words while keeping the move leaves the same tell. This makes it a generative companion to a pure after-the-fact style audit, not a duplicate of one.
Quick Start
Step 1: Create a Project Folder
mkdir human-writing && cd human-writing
Step 2: Download the Template
Click Download above, then:
mv ~/Downloads/CLAUDE.md ./
Step 3: Write
claude
Describe what you want written in Chinese and roughly how long. Claude will first check whether it actually holds enough concrete material — researching, asking up to three questions, or proposing a shorter piece if not — then draft from a specific speaker, and strip the AI-Chinese rhetorical moves at revision rather than padding to length.
Tips & Best Practices
- Expect, and allow, the material gate to push back on length — when the honest answer is "there are only three real things to say here," a tight 600-character piece is the method working correctly, not failing to follow instructions.
- Feed it your actual material up front (your experience, the specific numbers, the real quotes) — the method is explicit that a realistic, lived piece comes from material it can trace, and withholding it just triggers the three-question round.
- Use it as the generative front half and a dedicated AI-tell audit as the back half if you want belt-and-suspenders — this method bans the rhetorical moves while drafting, and a separate line-by-line audit catches anything that slips through.
Limitations
- Chinese-language-specific by design — the banned constructions (翻案腔, signpost colons, the four-part parallel) and the韵律 it's tuning for are features of Chinese internet long-form prose and don't map onto English AI-writing cleanup.
- It's a writing-craft and editorial discipline, not a fact-checker — for non-fiction it insists every real claim be traceable to user-supplied or verifiable material, but it can't itself verify that a source is accurate; that still needs a real check.
- The method deliberately favors substance over hitting a word count, so it will sometimes deliver less than you asked for when the material doesn't support more — that's a design choice, and a request that genuinely needs length still needs you to supply or allow the research that backs it.