| name | translate-blog |
| description | Translate English technical blog posts to Chinese. Triggers: 翻译, translate, 翻译一下, 翻一下, translate this. Use when the user wants to translate a technical article (EN → ZH) with a natural, developer-friendly voice. |
| argument-hint | [file path or URL to translate] |
Translate Blog
Translate English technical blog posts into idiomatic, developer-friendly Chinese. Single-pass workflow: analyze → translate → review → highlight → summarize.
Input
| Type | Action |
|---|
Local .md file | Read directly |
| URL | Fetch with scrapling, save as {slug}.md to target directory, then translate |
URL Fetching
Preferred method — use Jina AI Reader API for clean, AI-optimized markdown:
curl "https://r.jina.ai/<url>" \
-H "Authorization: Bearer $JINA_API_KEY" \
-H "Accept: text/markdown" \
-o <output-path>
Jina Reader returns clean article markdown — no nav, ads, sidebars. Usually no cleanup needed.
Fallback — scrapling for pages Jina cannot reach (paywalled, blocked). Apply progressive fallback:
Level 1 — Lightweight (static pages, no bot protection):
scrapling extract get '<url>' <output-path> --ai-targeted --timeout 30
Level 2 — Heavy (JS-rendered, Cloudflare-protected):
scrapling extract stealthy-fetch '<url>' <output-path> \
--real-chrome --no-headless \
--solve-cloudflare \
--network-idle --wait 5000 \
--timeout 90000 \
--ai-targeted
Level 3 — Manual fallback: If both Jina and scrapling fail, ask user to copy-paste article content.
{slug}: 2-4 word kebab-case from article title (e.g. building-a-react-hook)--ai-targeted: optimizes extracted content for AI consumption (removes nav, ads, sidebars)--real-chrome --no-headless: uses visible Chrome instance to bypass bot detection--network-idle --wait 5000: waits for JS to finish rendering before extraction
URL Content Cleanup
Jina Reader output is typically clean — skip cleanup unless visible noise remains. For scrapling output, strip non-article noise from saved markdown before translating:
- Navigation: site headers, breadcrumbs, sidebars, menus
- Footer clutter: copyright notices, "related posts", "read next", newsletter signup forms, comment sections
- Sharing links: "Share on Twitter/LinkedIn", "Discuss on Hacker News", etc.
- Author bio boxes: author blurbs, "written by" cards at the bottom
Keep only article body: title, publication date, article text, code blocks, embedded images. Result reads as clean standalone article.
Format repair: scraped content often has formatting issues — broken paragraphs, missing line breaks between sections, orphaned list items, inconsistent heading levels, stray punctuation, garbled whitespace. Fix these during cleanup so source is well-structured before translation.
Output
Directory Selection
Determine the output directory by inspecting the current project structure:
- If directories like
posts/, articles/, content/, blog/ exist, save source and translation there.
- Otherwise, save to the current working directory.
Do not ask the user — infer from the project layout.
File Naming
Translation saved as {basename}-zh.md in the target directory.
Example: posts/react-hooks.md → posts/react-hooks-zh.md
If the output file already exists, rename existing to {basename}-zh.backup-{timestamp}.md.
Workflow
Step 1: Analyze
Read the source and identify:
Content type — classify the article and set translation strategy:
| Type | Strategy |
|---|
| Tutorial / How-to | Precision > Fluency. Instructions must be unambiguous |
| Opinion piece | Preserve author's voice and stance, even if controversial |
| Research paper explainer | Highest terminology precision, add translator notes for obscure concepts |
| Documentation | Strict heading/code correspondence, structural fidelity |
| Blog / Narrative | Fluency > Precision. Allow liberal adaptation for natural flow |
Terminology: technical terms, proper nouns, acronyms, framework/library names. Cross-reference with ~/.claude/skills/translate-blog/glossary.md. For terms not in the glossary, determine standard Chinese translations.
Multi-meaning terms: for glossary entries with multiple translation options (e.g. Grounding → 基础化/落地), examine the article context and lock in one translation for the entire article. Record the choice to maintain consistency.
Tone: formal or conversational? Humor, metaphors, cultural references?
Challenges: passages that need creative adaptation (figurative language, long sentences, culturally-specific references).
Step 2: Translate
Apply the following principles:
Style: adapt precision/fluency balance based on content type identified in Step 1. Technical terms must be accurate and consistent across the full text. For narrative passages (intros, analogies, conclusions), rewrite into natural Chinese as if a native writer composed it — break long English sentences into shorter idiomatic ones, interpret metaphors by intended meaning not word-for-word.
Audience: Chinese-speaking developers/engineers. Assume familiarity with common technical terms — no annotation needed for words like API, container, framework, runtime.
Translator notes: only add concise explanations （**译注**：...） for genuinely niche or ambiguous terms the audience may not know (obscure startup names, inside jokes, rarely-used acronyms). Keep annotations minimal.
Code blocks: translate comments to Chinese. Leave variable names, function names, class names, and string literals in English.
Frontmatter: translate title and description fields to Chinese. Keep all other fields (tags, date, author, slug, etc.) as-is. Tags stay in English for search compatibility.
Images: preserve all image references as-is. Do not check or warn about image text language.
Markdown formatting: preserve all headings, links, lists, tables, and formatting exactly.
Step 3: Review Pass
Re-read the full translation against the source. Check for:
- Terminology consistency — every glossary term translated identically across all paragraphs. Multi-meaning terms use the translation locked in Step 1.
- Natural Chinese — no translationese (翻译腔). Flag and rewrite any sentence that reads as a word-for-word translation.
- Completeness — no skipped paragraphs, sentences, or list items.
- Markdown integrity — all headings, links, lists, tables, code blocks, and formatting preserved correctly.
Fix issues in-place before proceeding. Track the number and nature of fixes for the summary.
Step 4: Highlight Key Sentences
Identify sentences in the translated text that match these criteria:
- The article's central thesis or core argument (usually in intro or conclusion)
- Counter-intuitive or surprising claims that challenge common assumptions
- Original analogies or metaphors coined by the author
- Highly quotable summary statements that capture a key insight
Count: ~1 per 500 source words. Minimum 2, maximum 8.
Wrap selected sentences in bold (**...**) inline. A golden sentence is usually a standalone statement. Exclude: transitional sentences, purely descriptive sentences, example-illustration sentences.
Step 5: Summary
After translation completes, display:
**Translation complete**
Source: {source-path} ({word-count} words)
Output: {output-path}
Content type: {type}
Translation stats:
Glossary terms applied: {count} ({new-count} new terms added)
Golden sentences highlighted: {count}
Translator notes added: {count}
Review notes:
- {issues found and fixed during review, e.g. "3 sentences rewritten for natural flow"}
If no issues were found during review, display "No issues found."
Glossary
Terminology mappings are maintained locally at ~/.claude/skills/translate-blog/glossary.md.
Initialization: On first use, if the file does not exist, copy the seed glossary from references/glossary.md (relative to this skill) to the local path.
Updates: When new terms are encountered during translation, append them to the local glossary immediately. Do not wait for user confirmation.
Format: | English | Chinese | Notes | markdown table.
