Applies to every agent-facing file in the package —
skills/ppt-master/SKILL.md,references/*.md,workflows/**/*.md— and to repository documentation underdocs/.
Each Markdown file is single-language. New files mirror the language of their siblings in the same directory. Never mix English scaffolding with Chinese paragraphs (or the reverse) inside one file.
docs/zh/, README_CN.md, SPONSORS_CN.md, and other wholly Chinese files are complete translations, not exceptions to this rule — they are single-language too.
Inside an English file, a non-English string is allowed only when it is the subject matter. It is never the wording of an instruction to the model.
| Allowed — the string is content | Forbidden — the string is the rule |
|---|---|
A trigger phrase the user literally types (继续生成 projects/<name>) |
A decision rule written in Chinese (语速：edge 默认 +0%…) |
| A sample field value in an example JSON / notes block | A selection criterion phrased in Chinese (财报 → 稳重男声) |
A label the deck renders on the page (情景数据) |
An English sentence with a Chinese word standing in for its own term |
A pronunciation or typography example (百分之六十八 for TTS) |
A section heading or table header in Chinese |
A proper noun with English context (印章, 新中式) |
A prose paragraph of instructions in Chinese |
| A directory or brand name that is genuinely Chinese |
Hard rule: an instruction the model must follow is written in the file’s own language. A rule written in one language biases the model’s output toward that language and silently overrides SKILL.md’s “match the user’s language” discipline.
Do not write a rule that fixes which language the model replies in — write a one-line Chinese description is a defect. Say “in the user’s chat language” instead.
A user-facing message template written in one concrete language is content under §2, so it stays; but it must carry an explicit “translate to the user’s chat language if different” note.
SKILL.md’s description is English prose. Do not append per-language keyword lists to it, and do not name a specific agent host or harness in it. When a request intent triggers unreliably, widen the English intent vocabulary (nouns and verbs users actually type, including language-neutral tokens such as PPT / PPTX) rather than adding another language’s keywords — the language list has no natural end, and every entry costs budget on a file loaded on every run.