Curing the corporate throat-clear in AI Chinese

What it does

human-writing is a craftsmanship manual packaged as an agent Skill: markdown rulebooks and a Python linter that coach LLMs through writing Chinese prose with a specific human voice rather than the usual generic fluency. It enforces three stages—material validation, progression control, and linguistic tuning—so that reality writing is fact-checked and fiction is grounded in action, every paragraph advances new information, and the prose sheds report tone, model tics, and formulaic reversal structures. A check_prose.py script audits drafts against hard bans, and a distilled two-thousand-word version exists for direct use in chat windows.

The interesting bit

Version 1.1 treats robotic prose as a signal-processing problem. Instead of maintaining a blacklist of phrases like “not… but…,” it bans the underlying rhetorical move—fabricating a misunderstanding just to overturn it—regardless of wording. The linter also tracks sentence-length variation and conjunction density as statistical proxies for AI monotony, which is a more surgical approach than playing whack-a-mole with buzzwords.

Key highlights

• Genre-aware rulebooks split reality writing (sourced facts, verified quotes) from fiction (scene goals, causality) and cover Zhihu answers, essays, reviews, and broadcast scripts.
• The check_prose.py linter flags structural tics—repetition, AI parallelism, lyrical metaphor abuse, and rhythmic monotony—rather than just hunting banned words.
• A 2,000-word distilled version compresses the full ruleset for direct use in standard chat windows.
• The entire package is original rules and tooling; no scraped articles, training data, or model weights are included.

Caveats

• The authors acknowledge rule conflicts and false positives happen, and that performance varies across different models.
• It is fundamentally a set of editorial heuristics encoded for agents, not a trained model, so efficacy depends on how obediently your agent reads long markdown instructions.

Verdict

Worth a look if you publish Chinese content through LLMs and are tired of the faceless, throat-clearing cadence that makes every article sound like a corporate report dressed in casual clothes. Skip it if you are not working in Chinese prose or if you expect a fine-tuned model rather than a rule-based linter and prompt framework.