Localization Pass

Localization Pass · 7-Dimension Adaptation + Multilingual Audit

Single-article locale adaptation OR portfolio-wide multilingual quality audit.

When to invoke

• During /article pipeline when state.brief.target_market_locale != "en-US"
• /locale-audit <directory> — portfolio audit of existing multilingual content
• After /article translation runs to verify per-language SEO parity
• /locale-check workspace/{task}/draft.md — single-file dialect check

Mode 1 · Per-article dialect gate + 7-dimension adaptation

Wired v3.35 (2026-07-04): the dialect check is now the mandatory locale-spelling-check orchestrator stage — it runs on EVERY article (including en-US, where its production value is catching en-GB spellings leaking from model training data: "colour/organise/whilst" in en-US prose is an AI-tell). The executor existed since v5.0 with zero invocations (Rule 6). Evidence + thresholds: references/seo/serp-feature-value-2026.md §5.

Applied during the Optimize phase; the FULL 7-dimension adaptation applies when the target locale differs from US-English defaults.

The 7 dimensions

Workflow

bashCopy code
# Step 1 (WIRED, mandatory stage — runs automatically in the pipeline):
python -m scripts.lint.spelling_dialect_check --workspace {task_id} --json
# Resolves target dialect from state.brief.target_market_locale (en-US default).
# Exempt zones: References/Further Reading tail, blockquotes, code, link URLs,
# Capitalized mid-sentence proper nouns ("Labour Party", "Centre for ...").
# Gate: FAIL at drift_count >= 3 (systemic drift); 1-2 hits = warnings.
# Output: workspace/{task}/locale-spelling-lint.json (passed flag enforced by
# the orchestrator, run_pipeline._GATE_STAGES, and the pre-publish gate).

# Step 2 (repair, when the gate FAILs): apply drift_words[].use_instead
# replacements directly (deterministic word swaps — surgical Edit, cheapest
# route), then the freshness rule re-runs Step 1 automatically.
# NOTE (correction, v3.35): the pre-v3.35 claim that "the humanizer agent
# handles the locale layer" was FALSE — the humanizer has no locale code. For
# non-English locales the dialect gate no-ops and the 7-dimension adaptation
# below remains an LLM editing task guided by references/localization/{locale}.md.

# Legacy single-file form (manual diagnostics):
python -m scripts.lint.spelling_dialect_check --input {path}.md --target en-UK --json

Per-locale references (loaded on demand)

• references/localization/{locale}.md — currency / units / regulation / cultural per market
• references/localization/spelling/en-{US,UK,AU,CA,NZ}.md — 5 English dialect mappings

Currently supported locales (with detail-level references):

• en-US (default), en-UK, en-AU, en-CA, en-NZ
• de, fr, es, it, pt, nl
• jp, kr, cn, tw, hk, sg

For new locales: ship with skeleton reference + flag for user to fill domain-specific knowledge.

Mode 2 · /locale-audit <directory> — multilingual quality control

Audits a directory of multilingual content for completeness, consistency, hreflang correctness, meta-tag parity, freshness.

Phase 1: Discovery

1. Scan target directory. Group blog posts by language using:
   • Subdirectory names (en/, de/, fr/)
   • Frontmatter lang and translatedFrom fields
   • hreflang-map.json if present
2. Build content matrix mapping which post exists in which languages
3. Detect source language (most common translatedFrom target, or sourceLanguage field in hreflang-map.json)

Phase 2: Completeness audit

Translation coverage matrix:

Copy code
| Post (EN) | DE | FR | ES | JA |
|-----------|----|----|----|----|
| how-to-avoid-ai-slop | ok | ok | missing | missing |
| content-marketing-2026 | ok | missing | ok | missing |

Coverage: 60% (6 of 10 expected translations present)
Missing: 4 translations needed

Phase 3: Content parity audit

For every post that exists in multiple languages:

Flag every significant deviation as an issue.

Phase 4: SEO parity audit

For every language version verify:

Phase 5: Hreflang audit

If hreflang-tags.html, hreflang-sitemap.xml, or hreflang-map.json exists:

If no hreflang files exist: critical gap. Offer regeneration via translate skill.

Phase 6: Freshness audit

For posts with translatedDate frontmatter:

Emit actionable commands per stale file:

Copy code
3 translations are stale:
- de/ki-trends-2026.md (source updated 2 days ago)
  → Run: /translate en/ai-trends-2026.md --to de
- fr/ki-trends-2026.md (source updated 2 days ago)
  → Run: /translate en/ai-trends-2026.md --to fr
- es/tendencias-ia-2026.md (translation >90 days old)
  → Run: /translate en/ai-trends-2026.md --to es

Phase 7: Report

Output to projects/{slug}/locale-audit-{date}.md:

markdownCopy code
## Multilingual content audit report

### Summary
- Posts audited: [N] across [N] languages
- Overall health: [score] / 100
- Critical issues: [N]
- Warnings: [N]

### Translation coverage
[Matrix from Phase 2]

### Issues found

#### Critical
- [Issue with file references]

#### Warnings
- [Issue with file references]

#### Passed
- [Checks that passed]

### Prioritized fixes
1. [Highest-impact action]
2. [...]

### Stale-translation alerts
[Runnable commands from Phase 6]

### Quick fixes
- Run `/translate <file> --to <missing-langs>` for [N] missing translations
- Run `/multilingual` to regenerate hreflang assets
- Run `/locale <file> --locale <code>` for weak cultural adaptations

Error handling

Locale-specific quirks (high-leverage examples)

German (de): 20-30% longer than English; preserve compound nouns (Marketingautomatisierung); use Sie (formal) by default unless brand voice says otherwise.

Japanese (ja): 20% shorter than English; 敬語 (formal) for B2B; consider character-level character counting not byte counting; date format YYYY-MM-DD or 令和N年.

French (fr): Espace insécable before ; ? ! :, after «»; numbers use comma decimal (3,14 not 3.14); EUR uses €.

Simplified Chinese (cn): ~50% shorter than English; full-width punctuation in body, half-width in code; date YYYY-MM-DD or YYYY年MM月DD日; avoid Taiwan-specific vocabulary.

Traditional Chinese (tw / hk): Same as cn but full-width punctuation throughout + Traditional characters + region-specific vocabulary.

Cost

For typical article localization (US → UK / non-EN):

• 1× Claude Opus pass through 7 dimensions: ~$0.15
• Spell/dialect lint runs: free
• WebFetch for cultural references if needed: free
• Total: ~$0.15 per locale

For audit mode (no rewriting):

• Pure deterministic checks: free
• Optional LLM summary at end: ~$0.05

See also

• references/localization/{locale}.md — per-market detail
• scripts/lint/spelling_dialect_check.py — verification linter
• subskills/optimize/humanizer/SKILL.md — voice + purpose system used during adaptation