简明技术中文 (STC). A controlled language for Chinese technical writing: procedures, API docs, troubleshooting guides, and explanatory answers from LLMs. The goal is that a reader gets it right on the first read.

It follows the method of ASD-STE100 (Simplified Technical English): a numbered rule set, a controlled vocabulary where one meaning has one word, and sentence constraints a script can check. It is not a translation of STE and is not affiliated with ASD.

The Chinese README is README.md. The spec and vocabulary are in Chinese, because they are about Chinese.

Why

In October 2026 Andrej Karpathy posted that asking an LLM to explain things in ASD-STE100 makes the output far more readable. STE is a controlled English built for aircraft maintenance manuals: 53 writing rules plus a dictionary of about 900 words, each with one meaning.

Chinese had no equivalent. The existing Chinese style guides cover typography, punctuation and document structure. None of them does the core of STE: numbered rules, a controlled vocabulary, and sentence constraints that a tool can enforce. This project fills that gap.

Example

Before, from examples/before.md:

用户首先需要登陆到目标服务器，然后切换到 /opt 目录并且创建一个用于存放采集器配置文件以及运行时产生的临时数据的工作目录。

"The user first needs to log in [misspelt] to the target server, then switch to the /opt directory and also create a working directory that is used for storing the collector's configuration files as well as the temporary data produced at runtime."

After, from examples/after.md:

登录目标服务器。 Log in to the target server.

切换到 /opt 目录。 Change to the /opt directory.

创建工作目录 /opt/collector。工作目录存放配置文件和临时数据。 Create the working directory /opt/collector. It holds the configuration files and temporary data.

What the checker says about the "before" sentence:

examples/before.md:11: [1.3][必须] 禁用词「登陆」，改用：登录        # wrong character for "log in"
examples/before.md:11: [2.1][必须] 操作句 56 字，不超过 30 字      # instruction is 56 units, limit 30
examples/before.md:11: [4.1][建议] 禁用词「然后」，改用：拆成编号步骤  # "then": split into numbered steps (advisory)
examples/before.md:11: [4.1][必须] 禁用词「并且」，改用：拆句       # "and also": split the sentence

The whole "before" document has 34 must-fix findings. The rewrite has 0.

What is in the box

File	Content
规范.md	The rules. 6 sections, 40 rules, each with a rejected and an approved example. Two levels: must and should
词表.md	The vocabulary. Banned words with replacements, a one-meaning-one-word table, and meaning restrictions on common verbs. About 150 entries
tools/check.py	The checker. Zero dependencies, Python 3.9+. It reads the vocabulary tables straight from the Markdown
SKILL.md	Agent skill for Claude Code, Codex, Cursor and others
examples/	A full before/after rewrite

The rules that do most of the work:

One concept, one word, for the whole document.
No 进行 / 加以 + verb (the Chinese equivalent of "perform an installation"). Use the verb.
Instructions: at most 30 units. Descriptions: at most 40. A unit is one Chinese character, or one Latin word or number.
Instructions are imperatives that start with the verb. No 请 ("please"), no 您 ("you"), no "the user needs to".
Conditions before actions. Warnings in their own paragraph, before the step they apply to.
No 以上 / 以下 after a number (Chinese "and above / and below" is ambiguous about the endpoint). Write 大于 / 不小于 ("greater than / not less than").
打开 ("open") is only for files, windows and pages. Features are 启用 ("enabled"), programs are 启动 ("started").
可以 only means "may". Capability is 支持 ("supports").

Rules that are specific to Chinese and have no STE counterpart: nominalised verbs, long pre-nominal modifiers (the 的 chains), the endpoint ambiguity of 以上/以下, the permission/ability ambiguity of 可以, and relative time words. The mapping to STE is in appendix B of the spec.

Use it with an agent

npx skills add mzopedia/simplified-technical-chinese

Or copy the repository into your agent's skill folder. For Claude Code:

git clone --depth 1 https://github.com/mzopedia/simplified-technical-chinese.git ~/.claude/skills/simplified-technical-chinese

Without the skill, paste this into the system prompt. It is the "normal mode", roughly Karpathy's "80% of the way to STE":

用简明技术中文回答。规则：一句只说一件事，操作句不超过 30 字，描述句不超过 40 字；操作步骤用编号列表，动词开头，不写「请」「您」；用主动句；条件写在动作前面；一个概念全文只用一个词；不用「进行」「相关」「等」「以上」「尽快」「大概」；不用反问和感叹；数字后不用「以上」「以下」，写「大于」「不小于」。

For the full rule set, give the model 规范.md.

The checker

python3 tools/check.py file.md [more files]
    --strict            exit 1 when there is any must-level finding
    --json              JSON output
    --kind 操作|描述     force one sentence type; use 描述 for chat answers
    --max-op 30         unit limit for instructions
    --max-desc 40       unit limit for descriptions

It reliably catches: banned words, rejected synonyms, sentence and paragraph length, passive voice and modal / degree / estimate words inside instructions, trailing conditions, double negation, open-ended lists (…等), relative dates, Chinese numerals, "reduced by N times", questions and exclamations.

It cannot judge facts, term choice, logical order, meaning restrictions, modifier length, pronoun clarity, acronym consistency, or warnings and reasons inside steps. Those checks scored too low on real documents during calibration and now live only in the spec text, for humans. See docs/校准报告.md. It treats numbered list items as instructions. Every finding needs a human.

It skips code, inline code, HTML comments, text inside 「」 (UI labels), YAML front matter, lines that start with 不批准 ("rejected:"), and anything between  and .

Status

Version 0.2. The rules, the thresholds and the auto-checked vocabulary were calibrated on 3.5 million characters of real Chinese technical documentation (TiDB, Kubernetes, MDN, Vue, Vite, Ant Design, the Rust book and the ES6 tutorial, 2856 files): every rule the checker enforces was sampled and judged on real text, and the ones that were wrong more often than right were removed from the checker. Warnings on good documents dropped from 194 to 88 per 10,000 characters; more than half of what remains is sentence length. Excluding length, from 151 to 40. The length limits sit at the 90th percentile of good documents (90% of instructions are under 33 units, 90% of descriptions under 41). Details: docs/校准报告.md.

Known limits:

The vocabulary has about 170 entries and covers general technical writing only.
No controlled reading-comprehension study yet. The calibration measured false positives, not reading speed or error rates.
The checker is syntactic only.

The spec, the Chinese README and SKILL.md pass the checker with 0 must-level findings, and CI re-checks them on every push. This English README is not checked: the checker is for Chinese text.

Licence

Text of the spec and vocabulary: CC BY 4.0. Checker and tests: MIT.