一键在 Manus 中运行任何 Skill

translate-zh

星标0

分支0

更新时间2026年5月12日 02:43

Use when: producing a Simplified Chinese zh overlay of one finalized startup diligence report, translating summary-card.yaml or full-report.yaml into summary-card.zh.yaml / full-report.zh.yaml for the website /zh/ route, repairing or extending an existing zh overlay, or fixing zh sparse-bundle / whitelist / check-translation errors. Keywords: translate report, Chinese overlay, Simplified Chinese, zh, summary-card.zh.yaml, full-report.zh.yaml, sparse bundle, whitelist, check translation.

安装

用 Codex 或 Claude 帮你安装复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。

在 Manus 中运行

来源

vibewatch

vibewatch/startup

打开 GitHub 仓库查看创作者相关仓库

下载

在 Manus 中运行

translate-zh

Produce a Simplified Chinese zh overlay for one finalized report. The website's /zh/ route loads the overlay and falls back to the English leaf when a Chinese leaf is missing.

Scope

One invocation translates one report under reports/<runId>/. You write exactly two siblings:

summary-card.zh.yaml — list-card text and detail-page cover (~11 leaves).
full-report.zh.yaml — entire detail-page body (~1500–2500 leaves).

Do not touch any files other than the two final zh siblings and the required .translate-cache/<runId>/ intermediates. Do not edit a *.zh.yaml file directly — the applier writes it from the imported sparse bundle.

Start here

The parent agent must use the orchestration runner. Do not reassemble the export/split/merge/import/apply/check chain inline.

REPORT=<runId-or-company-name>
npm run translate:zh -- preflight "$REPORT"
npm run translate:zh -- init "$REPORT"

After init, translate only these cached files:

.translate-cache/$RUN_ID/summary-card.translate.yaml
.translate-cache/$RUN_ID/parts/part.NNN.yaml

summary-card.translate.yaml is small and should be handled directly by the parent. Its summary.headline is the list-card sentence; keep it as one fluent Chinese conclusion. Keep topStrengths, topRisks, and unresolvedGaps parallel within each array.

Then finalize through the runner:

npm run translate:zh -- lint-parts "$REPORT"
npm run translate:zh -- finalize-summary "$REPORT"
npm run translate:zh -- finalize-full "$REPORT"

If both zh siblings already exist, do not rewrite them; run:

npm run translate:zh -- verify "$REPORT"

run-translation.mjs owns preflight, cache paths, sparse export, splitting, part linting, merge/import/apply/check, and successful cache cleanup. Final deliverables are only reports/$RUN_ID/summary-card.zh.yaml and reports/$RUN_ID/full-report.zh.yaml.

If finalization fails, repair only the offending cached leaf or part, then rerun the narrowest finalize command. Re-run init only when the cache is structurally corrupted; it refuses to overwrite a non-empty cache unless --force is supplied.

Run lint-parts after all parts/part.NNN.yaml files have been translated and before finalize-full. It parses every part, verifies the manifest leaf count, and compares each part's string-leaf paths with the source bundle so shape drift is caught before any final zh file is written.

Subagent contract

Spawn each translation subagent with this prompt, substituting only the target path:

You are translating one part of a Chinese diligence-report overlay.

TARGET (read/write): <workspaceRoot>/.translate-cache/<runId>/parts/part.<NNN>.yaml

Workflow:
1. Read the TARGET file in full.
2. Apply the per-leaf process from .agents/skills/translate-zh/SKILL.md
   (read-through → draft → cover-English revision → final write).
3. Update TARGET in place. Same keys, same array length and order.
   Translate every non-null string value. Leave `null` as `null`.
   Do not add or drop keys. Do not change YAML shape.

Structural guardrails:
- `null` is an index placeholder, not a translation task. Never replace it
  with guessed text, and never write the string `"null"` unless the source
  leaf itself is the string `"null"`.
- Translate table cells cell-by-cell from the source text. Do not fill an
  empty-looking cell from neighboring rows, column expectations, or domain
  knowledge.
- When a block scalar such as `notes: >-` is followed by the next table or
  figure item, keep the next `- title:` aligned with its sibling list item;
  do not indent it under `notes`.
- Preserve `- text: null`, `- label: null`, and similar sparse placeholders
  exactly as `null`.

Hard constraints — do NOT:
- run `node`, `npm`, or any other command;
- invoke any script under `.agents/skills/translate-zh/scripts/`
  (those belong to the parent);
- write anywhere except the single TARGET path (no `/tmp`, no extra
  `.json` files);
- touch the English source under `reports/<runId>/`;
- read or modify any other `parts/part.*.yaml`;
- fetch URLs or invent facts.

Return when TARGET has been updated in place. The parent merges all parts.

Authoritative whitelist

scripts/whitelist.mjs is the single source of truth for which leaves are translatable. The exporter only emits whitelisted paths; the applier silently drops anything else; the validator byte-compares everything else. If a new visible-text path appears (new figure shape, new section), stop and report the whitelist gap. Whitelist changes are repo-development work, not part of a report translation run.

Use references/glossary.zh.yaml for recurring terms. Do not edit the glossary during a report translation; if a recurring term is missing, pick one consistent rendering within the cache and mention the glossary gap after the overlay is complete.

Invariants — must not change

Schema shape: keys, key order, array length, array order. Address list items by index — never reorder, merge, or split.
IDs, refs, slugs, URLs, dates, enums, numerics, schemaVersion, artifact stay byte-identical to English.
Proper nouns (companies, products, models, SKUs) keep Latin spelling. Never add a parenthetical Chinese alias (OpenAI（开放人工智能） — never).
Source titles, publishers, and keyQuote text are excluded by the whitelist; they stay in the original publication's language.
No new facts. No hedging shift. may is not 将. fails to is not 尚未. roughly, at least, reportedly, only when, unless, provided that survive at the same strength.
Punctuation: 全角 inside Chinese sentences (，。；：？！「」). Half-width with one space on each side around embedded English tokens and numbers: ARR 增长 80%，由 Microsoft Azure 渠道贡献.

Translation philosophy — re-author, not translate

The English file is the source of facts, not of sentence shape. Read the whole leaf, then write Chinese the way an investment analyst at 晚点 LatePost / 海外独角兽 / 远川研究所 would write it from scratch. If your draft mirrors English clause order, rewrite the whole sentence — patching words never removes 翻译腔.

Self-check: could an experienced Chinese analyst have written this from scratch, without ever seeing the English?

Sentence shape — apply before drafting

Time, place, condition lead. Y is true when X → X 之后，Y. Never Y，当 X 时.
Topic to the front; drop 对……来说. For thin AppSec teams, manual reproduction is the bottleneck → AppSec 团队人手紧，手工复现就是瓶颈.
Long noun phrases become short verb clauses. Three or more 的 in a row is a smell — flatten to a verb or split.
Split long sentences. A 30-word English sentence with two relative clauses becomes 2–3 short Chinese clauses joined by ， ； —— or a period. Subordination is English; Chinese is parataxis.
Use concrete verbs (落地 / 拼出 / 卡住 / 砸钱 / 吃掉 / 跑通 / 挤压 / 撬动 / 顶住 / 守住 / 打穿) over 实现 / 进行 / 做出 / 完成 / 形成. Collapse 进行 / 做出 + 名词 (做出决策 → 决定; 进行验证 → 验证).
Active over passive.

Tone

Investor-memo register: confident, concise, analytical, dense with verbs and numbers. Avoid 学术腔 / 公文味 / 营销文案腔 (市场进入策略、 相关性分析、形成竞争优势).

Length parity, not character parity. A natural Chinese sentence is shorter than its English source. If your draft is dramatically longer, you over-translated qualifiers (approximately, essentially, in the context of, in order to, with respect to) — cut them. Specificity (numbers, names, dates, mechanisms, direct quotes) stays intact.

Anti-patterns — rewrite the sentence, do not patch

Anti-pattern	Rewrite toward
长定语堆砌：`一个针对……的、能够……的、并且……的产品`	短句串联：`一个产品。它针对……、能够……，也……`
段首 `对……来说 / 对于……而言`	主语前置：`团队需要……`
`通过…… 来 ……`	`靠 / 借助 / 凭 / 用` + 动词，或直接动宾
`在 …… 的过程中 / 在 …… 上`	删除，或用 `时 / 中 / 里` 一字带过
被动逐字：`被设计为 / 被要求 / 被使用`	主动：`团队设计成……`、`监管要求……`
`A 和 B 和 C` 串列	顿号：`A、B、C`；必要时加 `等` 收束
`正在 …… 中`	直接动词：`公司在调整`
跨句用 `这 / 这一 / 这些` 指代	重复关键名词，或 `上述 / 该`
`做出……决策 / 进行……尝试`	直接动词：`决定 / 尝试`
`随着…… 的 ……`	`当…… / 一旦…… / ……之后`
数字前后 `约……的……的……`	`约 X` 后接动词，避免 `的` 连用
`是一个……的……`	判断句 `X 就是 Y`，或拆为两句
段段 `这意味着…… / 这表明……`	`因此 / 也就是说 / 反过来 / 换句话说`，或直接给结论
`相关 / 相应 / 对应 / 该等 / 此项 / 之`	删掉或换具体名词
同义堆叠：`努力和尝试 / 机会与可能`	取一个
`此外，/ 然而，/ 而且，` 机械直译	删掉，让句子自己接续
`公司称其预期…… / 公司表示其将……`	合一：`公司预期……`

Worked example

Source: "OpenAI's product experience depends on model availability, API/tool orchestration, SDK integration, and trust controls; partner channels add deployment reach but also opacity."
Stiff: OpenAI 的产品体验依赖于模型可用性、API/工具编排、SDK 集成和信任控制；合作伙伴渠道增加了部署触达，但也带来了不透明度。
Native: OpenAI 的产品体验靠几样东西撑着：模型够稳、API 和工具编排能跑、SDK 接得上、信任控制守得住；合作伙伴渠道把部署做远了，但也把这条链做模糊了。

The fix: pivot 依赖于 + 名词串 to 靠几样东西撑着 plus short verb clauses; turn the abstract noun 信任控制 into the action 守得住.

Soundcheck — mental grep your draft

If any of these survive, the leaf is not done:

三个 的 连用。
段首 对……来说、对于……而言、关于……方面、随着……、 通过……来……、在……的过程中、在……方面。
空动词 + 名词：做出决定、进行尝试、产生影响、实现增长。
被动逐字：被设计为、被要求、被认为是、被使用。
段段 这 / 这一 / 这些 指代上文。
公文词：相关 / 相应 / 对应 / 该等 / 此项 / 之。
正在……中。
同义堆叠：努力和尝试、机会与可能。
段段 此外，/ 然而，/ 而且，。

Per-leaf process

For every prose leaf and every multi-word table cell:

Read the part end-to-end first. Lock recurring terms, proper nouns, and acronyms so the same English term gets the same Chinese form throughout the file. Mark mechanical leaves (numbers, currency, dates, IDs, mermaid blocks) — copy them verbatim. $60M is $60M, never 6 万美元.
Draft using the sentence-shape rules above. Note hedges (may, likely, roughly, at least, reportedly, estimated) and constraints (only when, unless, provided that) — they must survive at the same strength.
Cover the English. Read the Chinese alone. Walk the soundcheck and the anti-pattern table. Rewrite any sentence that makes you pause as a whole sentence, not word-by-word.
Re-open the English. Confirm no fact, number, hedge, or qualifier was lost, strengthened, softened, or invented.
Save in place to the bundle (*.translate.yaml) or your assigned parts/part.NNN.yaml. The applier writes *.zh.yaml from the imported bundle — never edit it directly. The runner's finalize commands perform the final strict translation checks.

Table cells

Most leaves in full-report.yaml come from tables/[]/rows/[]/[] — short cells, not sentences. They follow stricter rules than prose:

Pure numbers / units / currency / percent / dates / IDs: do not translate. "$2.6T–$4.4T", "13.8", "2024", "n/a", "—", "null", "T+1" all stay verbatim. Currency symbols stay too — write "$852B", not "8520 亿美元", inside a table cell, even though the prose convention is the opposite. Cells must align across rows; a column that is numeric in 9 rows and prose in 1 row should keep the same form.
Single-word status / level / enum-like values: translate to a consistent short term and reuse it across rows. "High / Medium / Low" → "高 / 中 / 低". "Yes / No" → "是 / 否". "Confirming / Adverse / Neutral" → "证实 / 反向 / 中性". Pick one rendering per column and stick to it; do not vary by row.
Phrase cells (2–10 words): translate as natural Chinese noun phrases. Drop articles ("the", "a") and reorder modifiers as needed. Keep proper nouns in Latin spelling.
Mixed proper nouns + descriptive English: keep the proper noun, translate the ordinary descriptor. Do not leave a whole cell in English just because one token is a company, product, API, or market acronym. "U.S. enterprise sample" → "美国企业样本"; "Global retail/CPG" → "全球零售 / CPG"; "API platform / Responses API" → "API 平台 / Responses API".
Sentence cells: apply the prose translation philosophy above.
Cells that are mostly proper nouns + a small connector: keep the proper nouns and translate only the connector. "Microsoft & Azure" → "Microsoft 与 Azure".
Pure model / version / SKU lists: keep Latin tokens and normalize separators if useful. "GPT-5.x, GPT-4.1, o1, GPT-4o" → "GPT-5.x、GPT-4.1、o1、GPT-4o" is acceptable because every token is a model/version name.

Figures

Figures (figures/[]) are charts. The renderer reads everything from figure.data.{items,nodes,edges,points,columns,rows,series,layers,xAxis,yAxis}, not from top-level keys, so reader-visible chart text is two layers deep. Translate every reader-facing string under data/; leave identifiers, enums, refs, and numerics alone.

Translate: title, subtitle, summary, description, caption, insight, basis, notes, approximationNotes, the axis-label shapes (xAxisLabel / yAxisLabel / xLabel / yLabel / xAxis / yAxis / xAxis.label / yAxis.label / data.xAxis / data.yAxis / data.xAxis.label / data.yAxis.label / data.xAxis.high|low / data.yAxis.high|low), every label / name / detail / description / note / notes / text / relationship / lowLabel / highLabel / examples / context under data.items[] / data.nodes[] / data.edges[] / data.points[] / data.layers[] / data.layers[].items[] / data.layers[].modules[] / data.layers[].outputs[] / data.series[] / data.series[].points[] / data.columns[] / data.rows[] / data.rows[].values[], and journey-map row text: data.items[].actor / actors[] / phase / stage / emotion / channel / channels[] / touchpoints[], plus data.nodes[].risk / segment.
Never translate unless the exact path is explicitly listed above as visible chart text: id, key, slug, type, kind, layout, tone, status, sentiment, direction, trend, confidence, group, stage, phase, category, segment, unit, value, displayValue, date, delta, from, to, source, target, claimRef, claimRefs[], sourceRefs[], captionSources[], xAxis.high|low numerics, yAxis.high|low numerics. These are enum keys, chart geometry, refs, numbers, or publisher names — they drive CSS classes and bucketing logic in website/src/lib/figures.mjs and FigureRenderer.astro.

scripts/whitelist.mjs enforces this list. If a future figure shape adds a new visible-text key, report the missing whitelist path instead of editing FIGURE_PATHS during the translation run.

Style notes specific to figure text:

Node / column / row / phase labels are 2–6 character headings. Translate as short noun phrases; drop articles. "Foundation control" → "基金会控制"; "OpenAI Group PBC" → keep verbatim.
Mixed labels follow the same rule as table cells: keep brands/products, translate ordinary descriptors. "Meta / open weights" → "Meta / 开放权重"; do not keep the whole label English because it starts with a proper noun.
detail / description / note are one-sentence captions that appear in tooltips. Apply the prose translation philosophy: lead with the topic, drop 对……来说, prefer concrete verbs.
Pure model / version / SKU lists in figure details may stay Latin with Chinese separators: "GPT-5.x、GPT-4.1、o1、GPT-4o" is fine.
Axis labels often carry a parenthetical unit hint ("Inference Speed (tokens/sec, mid-size LLMs)"). Translate the prose, keep the parenthetical units verbatim: "推理速度（tokens/sec，中型 LLM）".
Quadrant high / low labels are 2–4 character endpoints ("Single market (UAE)" / "Multi-market MENA"). Translate as short labels; do not narrate.

Inspection snippets

If the parent runs an inline Node snippet for inspection, this repo is ESM. Use import, not require. Always cd to the workspace root first; js-yaml lives in the root package.json. Inline Node should be for one-off inspection only, never the main workflow path.

cd /home/ythuang/workspace/startup
node --input-type=module <<'NODE'
import { readFileSync } from 'node:fs';
import yaml from 'js-yaml';
NODE

Errors and remediation

Repair incrementally. Edit only the offending leaf or part in the existing cache, then rerun the narrowest finalize/check step. Do not re-export the whole report unless the bundle is structurally corrupted.

Use the runner first. Drop to lower-level scripts only when the failing path or part is already known.

check-translation.mjs missing: a required *.zh.yaml output is absent.
check-translation.mjs shape: a bundle or part changed object/array shape.
check-translation.mjs preserve: a non-translatable leaf was changed.
check-translation.mjs translate: a leaf is still empty or too English.
check-part-leaf-counts.mjs: a translated part changed string-leaf count or paths. Use the reported extra/missing paths to restore null placeholders or list indentation before rerunning.
bundle-translatable.mjs import: the edited bundle no longer matches source shape.
bundle-translatable.mjs merge: a part is missing, stale, or changed sparse shape.

Common repair order:

Fix the offending leaf or part in .translate-cache/<runId>.
For full-report part edits, rerun lint-parts until it passes.
Rerun finalize-summary or finalize-full.
Only if shape is badly corrupted, rerun init and re-translate the affected bundle or part.

Common pitfalls

Translating an enum value in YAML (the website maps recommendation: research-more to 继续研究 via displayLabel; the YAML stays English).
Reordering, merging, or splitting list items.
Translating source titles, publishers, or keyQuote.
Adding parenthetical Chinese aliases for companies or products.
Translating a mechanical table cell (number, currency, date, n/a).
Leaving byte-identical short labels that include an English descriptor. Proper nouns stay Latin, but the descriptor still needs Chinese: U.S. federal courts → 美国联邦法院; Series E at $61.5B → Series E 轮，估值 $61.5B; Responsible Scaling Policy v3.2 → Responsible Scaling Policy v3.2（负责任扩展政策）.

name	translate-zh
description	Use when: producing a Simplified Chinese zh overlay of one finalized startup diligence report, translating summary-card.yaml or full-report.yaml into summary-card.zh.yaml / full-report.zh.yaml for the website /zh/ route, repairing or extending an existing zh overlay, or fixing zh sparse-bundle / whitelist / check-translation errors. Keywords: translate report, Chinese overlay, Simplified Chinese, zh, summary-card.zh.yaml, full-report.zh.yaml, sparse bundle, whitelist, check translation.
argument-hint	<runId-or-company-name>
user-invocable	true

translate-zh

Produce a Simplified Chinese zh overlay for one finalized report. The website's /zh/ route loads the overlay and falls back to the English leaf when a Chinese leaf is missing.

Scope

One invocation translates one report under reports/<runId>/. You write exactly two siblings:

summary-card.zh.yaml — list-card text and detail-page cover (~11 leaves).
full-report.zh.yaml — entire detail-page body (~1500–2500 leaves).

Start here

The parent agent must use the orchestration runner. Do not reassemble the export/split/merge/import/apply/check chain inline.

REPORT=<runId-or-company-name>
npm run translate:zh -- preflight "$REPORT"
npm run translate:zh -- init "$REPORT"

After init, translate only these cached files:

.translate-cache/$RUN_ID/summary-card.translate.yaml
.translate-cache/$RUN_ID/parts/part.NNN.yaml

Then finalize through the runner:

npm run translate:zh -- lint-parts "$REPORT"
npm run translate:zh -- finalize-summary "$REPORT"
npm run translate:zh -- finalize-full "$REPORT"

If both zh siblings already exist, do not rewrite them; run:

npm run translate:zh -- verify "$REPORT"

Subagent contract

Spawn each translation subagent with this prompt, substituting only the target path:

You are translating one part of a Chinese diligence-report overlay.

TARGET (read/write): <workspaceRoot>/.translate-cache/<runId>/parts/part.<NNN>.yaml

Workflow:
1. Read the TARGET file in full.
2. Apply the per-leaf process from .agents/skills/translate-zh/SKILL.md
   (read-through → draft → cover-English revision → final write).
3. Update TARGET in place. Same keys, same array length and order.
   Translate every non-null string value. Leave `null` as `null`.
   Do not add or drop keys. Do not change YAML shape.

Structural guardrails:
- `null` is an index placeholder, not a translation task. Never replace it
  with guessed text, and never write the string `"null"` unless the source
  leaf itself is the string `"null"`.
- Translate table cells cell-by-cell from the source text. Do not fill an
  empty-looking cell from neighboring rows, column expectations, or domain
  knowledge.
- When a block scalar such as `notes: >-` is followed by the next table or
  figure item, keep the next `- title:` aligned with its sibling list item;
  do not indent it under `notes`.
- Preserve `- text: null`, `- label: null`, and similar sparse placeholders
  exactly as `null`.

Hard constraints — do NOT:
- run `node`, `npm`, or any other command;
- invoke any script under `.agents/skills/translate-zh/scripts/`
  (those belong to the parent);
- write anywhere except the single TARGET path (no `/tmp`, no extra
  `.json` files);
- touch the English source under `reports/<runId>/`;
- read or modify any other `parts/part.*.yaml`;
- fetch URLs or invent facts.

Return when TARGET has been updated in place. The parent merges all parts.

Authoritative whitelist

Invariants — must not change

Schema shape: keys, key order, array length, array order. Address list items by index — never reorder, merge, or split.
IDs, refs, slugs, URLs, dates, enums, numerics, schemaVersion, artifact stay byte-identical to English.
Proper nouns (companies, products, models, SKUs) keep Latin spelling. Never add a parenthetical Chinese alias (OpenAI（开放人工智能） — never).
Source titles, publishers, and keyQuote text are excluded by the whitelist; they stay in the original publication's language.
No new facts. No hedging shift. may is not 将. fails to is not 尚未. roughly, at least, reportedly, only when, unless, provided that survive at the same strength.
Punctuation: 全角 inside Chinese sentences (，。；：？！「」). Half-width with one space on each side around embedded English tokens and numbers: ARR 增长 80%，由 Microsoft Azure 渠道贡献.

Translation philosophy — re-author, not translate

Self-check: could an experienced Chinese analyst have written this from scratch, without ever seeing the English?

Sentence shape — apply before drafting

Time, place, condition lead. Y is true when X → X 之后，Y. Never Y，当 X 时.
Topic to the front; drop 对……来说. For thin AppSec teams, manual reproduction is the bottleneck → AppSec 团队人手紧，手工复现就是瓶颈.
Long noun phrases become short verb clauses. Three or more 的 in a row is a smell — flatten to a verb or split.
Split long sentences. A 30-word English sentence with two relative clauses becomes 2–3 short Chinese clauses joined by ， ； —— or a period. Subordination is English; Chinese is parataxis.
Use concrete verbs (落地 / 拼出 / 卡住 / 砸钱 / 吃掉 / 跑通 / 挤压 / 撬动 / 顶住 / 守住 / 打穿) over 实现 / 进行 / 做出 / 完成 / 形成. Collapse 进行 / 做出 + 名词 (做出决策 → 决定; 进行验证 → 验证).
Active over passive.

Tone

Investor-memo register: confident, concise, analytical, dense with verbs and numbers. Avoid 学术腔 / 公文味 / 营销文案腔 (市场进入策略、 相关性分析、形成竞争优势).

Anti-patterns — rewrite the sentence, do not patch

Anti-pattern	Rewrite toward
长定语堆砌：`一个针对……的、能够……的、并且……的产品`	短句串联：`一个产品。它针对……、能够……，也……`
段首 `对……来说 / 对于……而言`	主语前置：`团队需要……`
`通过…… 来 ……`	`靠 / 借助 / 凭 / 用` + 动词，或直接动宾
`在 …… 的过程中 / 在 …… 上`	删除，或用 `时 / 中 / 里` 一字带过
被动逐字：`被设计为 / 被要求 / 被使用`	主动：`团队设计成……`、`监管要求……`
`A 和 B 和 C` 串列	顿号：`A、B、C`；必要时加 `等` 收束
`正在 …… 中`	直接动词：`公司在调整`
跨句用 `这 / 这一 / 这些` 指代	重复关键名词，或 `上述 / 该`
`做出……决策 / 进行……尝试`	直接动词：`决定 / 尝试`
`随着…… 的 ……`	`当…… / 一旦…… / ……之后`
数字前后 `约……的……的……`	`约 X` 后接动词，避免 `的` 连用
`是一个……的……`	判断句 `X 就是 Y`，或拆为两句
段段 `这意味着…… / 这表明……`	`因此 / 也就是说 / 反过来 / 换句话说`，或直接给结论
`相关 / 相应 / 对应 / 该等 / 此项 / 之`	删掉或换具体名词
同义堆叠：`努力和尝试 / 机会与可能`	取一个
`此外，/ 然而，/ 而且，` 机械直译	删掉，让句子自己接续
`公司称其预期…… / 公司表示其将……`	合一：`公司预期……`

Worked example

Source: "OpenAI's product experience depends on model availability, API/tool orchestration, SDK integration, and trust controls; partner channels add deployment reach but also opacity."
Stiff: OpenAI 的产品体验依赖于模型可用性、API/工具编排、SDK 集成和信任控制；合作伙伴渠道增加了部署触达，但也带来了不透明度。
Native: OpenAI 的产品体验靠几样东西撑着：模型够稳、API 和工具编排能跑、SDK 接得上、信任控制守得住；合作伙伴渠道把部署做远了，但也把这条链做模糊了。

The fix: pivot 依赖于 + 名词串 to 靠几样东西撑着 plus short verb clauses; turn the abstract noun 信任控制 into the action 守得住.

Soundcheck — mental grep your draft

If any of these survive, the leaf is not done:

三个 的 连用。
段首 对……来说、对于……而言、关于……方面、随着……、 通过……来……、在……的过程中、在……方面。
空动词 + 名词：做出决定、进行尝试、产生影响、实现增长。
被动逐字：被设计为、被要求、被认为是、被使用。
段段 这 / 这一 / 这些 指代上文。
公文词：相关 / 相应 / 对应 / 该等 / 此项 / 之。
正在……中。
同义堆叠：努力和尝试、机会与可能。
段段 此外，/ 然而，/ 而且，。

Per-leaf process

For every prose leaf and every multi-word table cell:

Read the part end-to-end first. Lock recurring terms, proper nouns, and acronyms so the same English term gets the same Chinese form throughout the file. Mark mechanical leaves (numbers, currency, dates, IDs, mermaid blocks) — copy them verbatim. $60M is $60M, never 6 万美元.
Draft using the sentence-shape rules above. Note hedges (may, likely, roughly, at least, reportedly, estimated) and constraints (only when, unless, provided that) — they must survive at the same strength.
Cover the English. Read the Chinese alone. Walk the soundcheck and the anti-pattern table. Rewrite any sentence that makes you pause as a whole sentence, not word-by-word.
Re-open the English. Confirm no fact, number, hedge, or qualifier was lost, strengthened, softened, or invented.
Save in place to the bundle (*.translate.yaml) or your assigned parts/part.NNN.yaml. The applier writes *.zh.yaml from the imported bundle — never edit it directly. The runner's finalize commands perform the final strict translation checks.

Table cells

Most leaves in full-report.yaml come from tables/[]/rows/[]/[] — short cells, not sentences. They follow stricter rules than prose:

Pure numbers / units / currency / percent / dates / IDs: do not translate. "$2.6T–$4.4T", "13.8", "2024", "n/a", "—", "null", "T+1" all stay verbatim. Currency symbols stay too — write "$852B", not "8520 亿美元", inside a table cell, even though the prose convention is the opposite. Cells must align across rows; a column that is numeric in 9 rows and prose in 1 row should keep the same form.
Single-word status / level / enum-like values: translate to a consistent short term and reuse it across rows. "High / Medium / Low" → "高 / 中 / 低". "Yes / No" → "是 / 否". "Confirming / Adverse / Neutral" → "证实 / 反向 / 中性". Pick one rendering per column and stick to it; do not vary by row.
Phrase cells (2–10 words): translate as natural Chinese noun phrases. Drop articles ("the", "a") and reorder modifiers as needed. Keep proper nouns in Latin spelling.
Mixed proper nouns + descriptive English: keep the proper noun, translate the ordinary descriptor. Do not leave a whole cell in English just because one token is a company, product, API, or market acronym. "U.S. enterprise sample" → "美国企业样本"; "Global retail/CPG" → "全球零售 / CPG"; "API platform / Responses API" → "API 平台 / Responses API".
Sentence cells: apply the prose translation philosophy above.
Cells that are mostly proper nouns + a small connector: keep the proper nouns and translate only the connector. "Microsoft & Azure" → "Microsoft 与 Azure".
Pure model / version / SKU lists: keep Latin tokens and normalize separators if useful. "GPT-5.x, GPT-4.1, o1, GPT-4o" → "GPT-5.x、GPT-4.1、o1、GPT-4o" is acceptable because every token is a model/version name.

Figures

Translate: title, subtitle, summary, description, caption, insight, basis, notes, approximationNotes, the axis-label shapes (xAxisLabel / yAxisLabel / xLabel / yLabel / xAxis / yAxis / xAxis.label / yAxis.label / data.xAxis / data.yAxis / data.xAxis.label / data.yAxis.label / data.xAxis.high|low / data.yAxis.high|low), every label / name / detail / description / note / notes / text / relationship / lowLabel / highLabel / examples / context under data.items[] / data.nodes[] / data.edges[] / data.points[] / data.layers[] / data.layers[].items[] / data.layers[].modules[] / data.layers[].outputs[] / data.series[] / data.series[].points[] / data.columns[] / data.rows[] / data.rows[].values[], and journey-map row text: data.items[].actor / actors[] / phase / stage / emotion / channel / channels[] / touchpoints[], plus data.nodes[].risk / segment.
Never translate unless the exact path is explicitly listed above as visible chart text: id, key, slug, type, kind, layout, tone, status, sentiment, direction, trend, confidence, group, stage, phase, category, segment, unit, value, displayValue, date, delta, from, to, source, target, claimRef, claimRefs[], sourceRefs[], captionSources[], xAxis.high|low numerics, yAxis.high|low numerics. These are enum keys, chart geometry, refs, numbers, or publisher names — they drive CSS classes and bucketing logic in website/src/lib/figures.mjs and FigureRenderer.astro.

scripts/whitelist.mjs enforces this list. If a future figure shape adds a new visible-text key, report the missing whitelist path instead of editing FIGURE_PATHS during the translation run.

Style notes specific to figure text:

Node / column / row / phase labels are 2–6 character headings. Translate as short noun phrases; drop articles. "Foundation control" → "基金会控制"; "OpenAI Group PBC" → keep verbatim.
Mixed labels follow the same rule as table cells: keep brands/products, translate ordinary descriptors. "Meta / open weights" → "Meta / 开放权重"; do not keep the whole label English because it starts with a proper noun.
detail / description / note are one-sentence captions that appear in tooltips. Apply the prose translation philosophy: lead with the topic, drop 对……来说, prefer concrete verbs.
Pure model / version / SKU lists in figure details may stay Latin with Chinese separators: "GPT-5.x、GPT-4.1、o1、GPT-4o" is fine.
Axis labels often carry a parenthetical unit hint ("Inference Speed (tokens/sec, mid-size LLMs)"). Translate the prose, keep the parenthetical units verbatim: "推理速度（tokens/sec，中型 LLM）".
Quadrant high / low labels are 2–4 character endpoints ("Single market (UAE)" / "Multi-market MENA"). Translate as short labels; do not narrate.

Inspection snippets

cd /home/ythuang/workspace/startup
node --input-type=module <<'NODE'
import { readFileSync } from 'node:fs';
import yaml from 'js-yaml';
NODE

Errors and remediation

Use the runner first. Drop to lower-level scripts only when the failing path or part is already known.

check-translation.mjs missing: a required *.zh.yaml output is absent.
check-translation.mjs shape: a bundle or part changed object/array shape.
check-translation.mjs preserve: a non-translatable leaf was changed.
check-translation.mjs translate: a leaf is still empty or too English.
check-part-leaf-counts.mjs: a translated part changed string-leaf count or paths. Use the reported extra/missing paths to restore null placeholders or list indentation before rerunning.
bundle-translatable.mjs import: the edited bundle no longer matches source shape.
bundle-translatable.mjs merge: a part is missing, stale, or changed sparse shape.

Common repair order:

Fix the offending leaf or part in .translate-cache/<runId>.
For full-report part edits, rerun lint-parts until it passes.
Rerun finalize-summary or finalize-full.
Only if shape is badly corrupted, rerun init and re-translate the affected bundle or part.

Common pitfalls

Translating an enum value in YAML (the website maps recommendation: research-more to 继续研究 via displayLabel; the YAML stays English).
Reordering, merging, or splitting list items.
Translating source titles, publishers, or keyQuote.
Adding parenthetical Chinese aliases for companies or products.
Translating a mechanical table cell (number, currency, date, n/a).
Leaving byte-identical short labels that include an English descriptor. Proper nouns stay Latin, but the descriptor still needs Chinese: U.S. federal courts → 美国联邦法院; Series E at $61.5B → Series E 轮，估值 $61.5B; Responsible Scaling Policy v3.2 → Responsible Scaling Policy v3.2（负责任扩展政策）.

translate-zh

同仓库更多 Skills

同仓库更多 Skills

translate-zh

Scope

Start here

Subagent contract

Authoritative whitelist

Invariants — must not change

Translation philosophy — re-author, not translate

Sentence shape — apply before drafting

Tone

Anti-patterns — rewrite the sentence, do not patch

Worked example

Soundcheck — mental grep your draft

Per-leaf process

Table cells

Figures

Inspection snippets

Errors and remediation

Common pitfalls

translate-zh

Scope

Start here

Subagent contract

Authoritative whitelist

Invariants — must not change

Translation philosophy — re-author, not translate

Sentence shape — apply before drafting

Tone

Anti-patterns — rewrite the sentence, do not patch

Worked example

Soundcheck — mental grep your draft

Per-leaf process

Table cells

Figures

Inspection snippets

Errors and remediation

Common pitfalls