Blog Translate skill

Blog Translate, SEO-Optimized Blog Translation

Bundled paths below use host Markdown substitution of ${CLAUDE_SKILL_DIR}; it is not an exported shell variable. Resolve them before execution, quote paths, and refuse nonabsolute overrides. Pass blog_reference_root resolved from ${CLAUDE_SKILL_DIR}/../blog/references, blog_template_root from ${CLAUDE_SKILL_DIR}/../blog/templates, and needed sibling roots to agents. Read the main reference orchestration-details.md before loading project context; pass only its helper-fenced output to downstream agents.

Translates an existing blog post into one or more target languages. Unlike generic translation, this skill produces SEO-optimized, publication-ready content with localized keywords, meta tags, and culturally correct formatting.

Adapted from claude-blog-multilingual by Chris Mueller (Pro Hub Challenge, March 2026). Original repository: no longer available as of 2026-10-12

Key References

Load on demand:

• ${CLAUDE_SKILL_DIR}/references/translation-rules.md, format preservation, number/date/currency formats per locale, quote handling, quality criteria.
• ${CLAUDE_SKILL_DIR}/references/cultural-adaptation.md, cultural profiles per locale (DACH, Francophone, Hispanic, Japanese, custom). This file is shared with blog-localize (do not duplicate).

Workflow

Phase 1: Input Parsing

1. Read the source file (markdown, MDX, or HTML) only after resolving it against the project root/current working directory. Reject symlinked paths, traversal outside the root, files over 10 MB, and binary files.
2. Auto-detect source language. Order of preference:
   • Frontmatter lang field.
   • HTML lang attribute.
   • Content analysis (script, common stop words).
3. Parse target languages from --to as comma-separated Google-compatible hreflang tags (de, fr, es-MX, ja, pt-BR, zh-Hant). If --to is missing, ask the user once: "Which languages should I translate to? Provide hreflang tags such as de, fr, es-MX, ja, pt-BR."
4. Normalize every code with the shared multilingual locale rules used by blog-multilingual, blog-localize, and blog-locale-audit: ISO 639-1 language in lowercase, optional ISO 15924 script in title case, optional ISO 3166-1 Alpha-2 region in uppercase. Reject invalid codes with a suggestion (jp becomes "Did you mean ja for Japanese?"). Require a region or explicit neutral mode for ambiguous language-only targets such as es, pt, and zh. Google accepts a language code by itself as hreflang (https://developers.google.com/search/docs/specialty/international/localized-versions, read 2026-10-12), but never a region by itself; this stricter rule exists for translation quality, not because Google requires a region. If a target equals the normalized source language, skip it with a notice.

Phase 2: Content Analysis

Extract the translatable surface:

• Frontmatter: title, description, tags, author (only when translatable, e.g. role labels, not personal names).
• All headings (H1, H2, H3).
• Body paragraphs.
• Image alt text and <figcaption> content.
• Chart <text> and <tspan> content; preserve SVG attributes (x, y, fill, transform), except under the overflow rule: if text overflows, widen the viewBox or reduce font-size; never truncate (see ${CLAUDE_SKILL_DIR}/references/translation-rules.md).
• FAQ questions and answers.
• Evidence-backed explanation text.
• Key Takeaways or summary box.
• CTA text.
• Internal-link zone anchor text.

Preserve unchanged:

• Markdown and HTML structure, tags, attributes.
• Image URLs, link URLs, frontmatter keys.
• Executable code fences and inline code. Translate comments only outside executable code, or when the user explicitly asks for localized tutorial comments.
• Internal-link zone markers ([INTERNAL-LINK: ...]).
• Source organization names in citations (Gartner, McKinsey, etc.).
• Person names.
• Schema JSON-LD blocks (translate user-facing content strings only; never translate Person, Organization, or Brand names, URLs, IDs, @id, or sameAs).

Identify the primary and secondary keywords for Phase 3.

Phase 3: Keyword Localization

For each target language:

1. Decide whether the source keyword is the established term in the target market. If yes (e.g., "Content Marketing" stays in German), keep it.
2. If a local equivalent has real search behavior, swap to it.
3. Apply the same logic to secondary keywords.
4. Record the mapping. The translator agent uses it to update title, meta description, and H2 headings consistently.

Phase 4: Translation

Spawn the blog-translator agent (Agent tool, one call per target language) with:

• The source content.
• The keyword localization map from Phase 3.
• The target language code.
• Pointer to ${CLAUDE_SKILL_DIR}/references/translation-rules.md.
• Instruction to limit this phase to language, register, natural topical coverage, and formatting. Reserve brand, legal, statistic, and cultural substitutions for blog-localize.

Run agents in parallel when translating into multiple languages.

When blog-multilingual invokes this skill, its invocation starts with Mode: orchestrated; see Phase 5 (output root) and Phase 7 (contract).

The agent returns the fully translated post in the same format as the input.

Phase 5: Post-Processing

For each translated version:

1. Add or update locale frontmatter:

   lang: "de"
   translatedFrom: "en"
   translatedDate: "YYYY-MM-DD"
   sourceHash: "<sha256 of the source file bytes>"
   translationHash: "<sha256 of the translated body, after the closing frontmatter>"
   slug: "wie-man-ki-slop-vermeidet"
   

   The agent cannot compute hashes (no Bash), so you add sourceHash and translationHash here; blog-locale-audit relies on them. Provenance lives in frontmatter only, with no appended HTML comment.

2. Verify structural integrity:

   • Same number of H2 and H3 sections as the original.
   • All images present with translated alt text.
   • All SVG charts present with translated text labels (length-adjusted: DE +30%, FR +15%, JA -20%, others and the overflow rule in ${CLAUDE_SKILL_DIR}/references/translation-rules.md).
   • FAQ count matches.
   • Evidence-backed explanations from the source remain intact where present.

3. Save translated files. Exactly one output root applies per run:

   • Standalone (/blog translate): translations/{lang}/{localized-slug}.{ext} under the current working directory.
   • Orchestrated (the invocation starts with Mode: orchestrated): multilingual/{lang}/{localized-slug}.{ext}, the directory the orchestrator created. Never also write under translations/.

   Before writing, slugify {localized-slug} to lowercase ASCII with only a-z, 0-9, and hyphens, rejecting empty or reserved names. Create the language directory from the normalized hreflang code only. Resolve the final path and require it to stay inside the intended output root; reject symlinked output paths.

Phase 6: Translation-Quality Guardrails

Scan the output for machine-translation artifacts before reporting done:

• Literal idioms (English idioms transliterated, not adapted).
• Unnatural word order (SOV translated as SVO into a non-SVO language, or the reverse).
• Mixed-language sentences (other than established loanwords).
• Number, date, or currency strings still in source format.
• Frontmatter strings still in the source language.

Flag every issue inline (file path, line number, fix suggestion). The translator agent should re-pass any flagged passage before delivery.

Phase 7: Final Artifact Delivery Contract

Translation can change claims, links, metadata, schema, and rendered layout.

Orchestrated mode. When invoked by blog-multilingual (orchestrated mode), skip the delivery contract; blog-multilingual runs it once per locale on the final text. Orchestrated mode is signaled by the literal line Mode: orchestrated at the top of the invocation. In that mode report each locale's status as pending (contract deferred to blog-multilingual) and never call a translation complete or publication-ready.

Standalone runs. Before calling any translated file complete or publication-ready, run the canonical sequence in ${CLAUDE_SKILL_DIR}/../blog/references/blog-delivery-contract.md against each final locale independently (this skill is the last text change for the locale). Translation-specific rules:

• Resolve BLOG_SCRIPT_DIR="${CLAUDE_BLOG_SCRIPTS_DIR:-${CLAUDE_SKILL_DIR}/../../scripts}" and reject it unless absolute. Never resolve a helper from the current project's scripts/ directory.
• Draft folder: create a per-locale delivery staging directory under the intended language output and copy the final locale file there under its lowercase localized slug. Pass --slug <localized-slug> to the gates.
• Visuals: resolve every draft marker ([INTERNAL-LINK, [IMAGE, [CHART): none may remain. Reuse the source hero (copy it as hero.<ext>); a new paid hero needs the blog-image cost and consent gate first.
• Render: blog_render.py --md <locale-file> --out-dir <locale-delivery-dir>.
• Init nonce: blog_preflight.py --draft <locale-delivery-dir> --init-review-nonce.
• Review: pass only the printed nonce to a fresh blog-reviewer against the rendered locale HTML; it returns text and you save it verbatim to <locale-delivery-dir>/review.md.
• Gates: blog_preflight.py --draft <locale-delivery-dir> --strict --slug <localized-slug>.
• Loop: after ANY change to the file, hero, HTML, or PDF, re-render, re-init the nonce, and re-review (at most 3 failed gate runs per locale draft; run blog_preflight.py --draft <locale-delivery-dir> --reset-iterations before the first gate run only when a new translation starts in a used directory or a delivered locale gets a new edit pass, never after a failed run).

Require 90+/100, zero P0, the matching nonce, final BLOCKING: false, and all other gates. A failed locale remains draft-blocked; report its diagnostic and do not describe it as complete.

Phase 8: Delivery

## Translation complete: [Original title]

### Source
- Language: [source]
- File: [source path]

### Translations
| Language | File | Keywords adapted | Status |
|----------|------|------------------|--------|
| de | translations/de/{slug}.md | [N] | passed / draft-blocked / pending (orchestrated) |
| fr | translations/fr/{slug}.md | [N] | passed / draft-blocked / pending (orchestrated) |

### Quality checks
- Structural integrity: pass / fail per language
- Meta tags localized: pass / fail per language
- Numbers, dates, currencies formatted per locale: pass / fail
- Keywords localized: [N] keywords adapted
- Machine-translation artifacts flagged: [N] (see notes above)
- Delivery contract: score, P0 status, nonce verification, and preflight report per locale (standalone only; deferred to blog-multilingual in orchestrated mode)

### Next steps
- Run `/blog localize <file> --locale <code>` for cultural deep-adaptation.
- Run `/blog locale-audit translations/` to verify completeness.
- Use `/blog multilingual` to combine write, translate, localize, hreflang
  in one command.

Error Handling

Scenario	Action
Unsupported language code	Suggest the correct Google-compatible hreflang code
Source equals a target	Skip with "Source is already in [lang]"
File not found	Report error with suggested path
Translator agent timeout	Retry once, then report partial results
Binary or non-text file	Report error, suggest correct file

Cross-References

• Next step (cultural depth): /blog localize <file> --locale <code>
• QA sweep across all language versions: /blog locale-audit <directory>
• One-command pipeline: /blog multilingual <topic> --languages <codes>