dualface/ste-zh

ste‑zh – 中文技术报告规范 Skill

What it is

• An Agent skill that tells large‑language‑model (LLM) agents (Claude, Cursor, Codex, etc.) to write their status‑reports in a controlled, unambiguous Chinese style. It adapts the aerospace‑industry standard ASD‑STE100 (Simplified Technical English) to Mandarin, enforcing a “key‑info‑first, evidence‑based, fixed‑state” format.

Why it matters

• AI agents often produce verbose, speculative, or incomplete updates when fixing bugs or delivering features. ste‑zh forces the agent to:
  1. State the result up front (e.g., 已完成…),
  2. Cite concrete evidence (file/line numbers, test results),
  3. Separate completion from verification, and
  4. Present any pending decisions as numbered options.
• This reduces misunderstanding, speeds up human review, and makes multi‑agent workflows (e.g., with the author’s Kander board) easier to scan.

Core features

• Key‑info first – the first sentence always declares the report type (result, progress, summary, etc.).
• Fixed status vocabulary – a limited set of status words (已完成, 已验证, 未验证, …) defined in references/terminology.md.
• Evidence‑driven – every claim must be backed by a verifiable artifact (code line, test case, command output).
• Active voice & one‑word‑one‑meaning – eliminates vague verbs and ambiguous phrasing.
• Decision options – when human input is required, the skill auto‑formats a numbered list with a recommended choice.

Installation

• Preferred: use Vercel‑Labs skills CLI, which registers the skill under the canonical name ste.

  npx skills add dualface/ste-zh -g          # global install
  npx skills add dualface/ste-zh            # project‑local install
  npx skills add dualface/ste-zh -g -a claude-code  # install for a specific agent
  

• Manual: clone the repo into the agent’s skills directory and name the folder ste (e.g., ~/.claude/skills/ste).

How to use

• In any conversation with a supported agent, type one of the trigger commands:
  • /ste
  • “用 STE 规范输出”
  • “按 ASD‑STE100 汇报”
• The agent will then format all subsequent replies according to the STE‑ZH rules until you send stop ste (or a similar stop command).
• If the agent’s output drifts (e.g., after a long context), resend the trigger to re‑activate the style.

Typical workflow

1. Agent works on a task (bug fix, refactor, etc.).
2. You ask for a report – the agent replies with a STE‑ZH‑styled block:

   已完成：登录刷新逻辑的代码修改。iOS 端未验证。
   
   原因：两个请求同时刷新 token，后完成的刷新使先拿到的 token 失效。
   
   证据：
   - auth/refresh.ts:42 — 刷新前不检查是否已有刷新在进行。
   …
   

3. You make a decision by replying with the option number (e.g., 1).
4. Agent proceeds based on your choice.

Who benefits

• Teams that rely on LLM agents for code review, debugging, or documentation and need concise, verifiable updates.
• Users of multi‑agent orchestration tools like Kander, where many tasks appear on a board and quick scanning is essential.
• Any developer who wants a consistent Chinese reporting style without writing custom prompts each time.

License

• MIT (see LICENSE).

Related projects by the author

• Kander – a rule‑driven multi‑agent kanban scheduler.
• Ullage‑cli – a local daemon/CLI for tracking usage of Claude, ChatGPT, Grok, Cursor, etc.
• QuickTUI – a mobile‑first terminal UI for coding agents.

---

All information above is taken directly from the repository’s README.