openclaw-i18n — Internationalization Layer for OpenClaw

What This Skill Does

OpenClaw is a powerful AI agent framework — but by default, models generate
output that often has wrong diacritics, mixed languages, or mechanically garbled
text when users communicate in non-English languages. This skill fixes that
with two layers:

Layer 1 — SKILL.md: Configures the agent to think and respond correctly
in the user's language. Language detection, diacritics enforcement, locale
formatting, cultural communication norms.

Layer 2 — Post-Processor: A Python script that mechanically cleans the
model's raw output before it reaches the user. This is the unique part — no
other ClawHub skill does this. It catches the specific mechanical errors that
models reliably produce on non-English text: missing diacritics, stray
Chinese/Russian characters, merged words, common typos.

---

Language Detection

How It Works

• - Auto-detect user language from conversation context
• Signals: explicit language statements, user names, place references, writing

style, diacritic usage

• - Confidence scoring — if below 80%, ask the user to confirm

Supported Languages

Language	Code	Status	Diacritics	Launch
Romanian	ro	Stable	ă â î ș ț	v1.0
German

de | Stable | ü ö ä ß | v1.0 |

| French | fr | Planned | à â ç é è ê ë î ï ô û ü | v1.1 | | Spanish | es | Planned | á é í ó ú ñ ü ¿ ¡ | v1.1 | | Italian | it | Planned | à è é ì ò ó ù | v1.2 | | Portuguese | pt | Planned | à á â ã ç é ê í ó ô õ ú | v1.2 |

English (en) is always the fallback when detection fails.

Configuration Commands

CODEBLOCK0

Configuration persists across sessions via OpenClaw memory tools.

---

Output Rules

Language Enforcement

• - Always respond in the detected or configured language
• Never switch languages mid-response without explicit reason
• If a user writes in Romanian, respond in Romanian — never English
• If a user writes in German, respond in German — never English

Character Rules

• - Never output Chinese, Russian, Arabic, or non-Latin characters in

non-intentional contexts. Stray characters that appear in Romanian or German output are model errors — the post-processor removes them.

• - Words are always separated by correct spaces — never merge two words

into one. Example: "sivei" should never appear as "știu" split wrongly.

• - Always use correct diacritics for the active language

Diacritics — Never Approximate

Romanian: ă â î ș ț — always correct, never write "s" instead of "ș" or "t" instead of "ț" in Romanian words.

German: ü ö ä ß — never substitute with "ue", "oe", "ae" unless user
explicitly asks for old spelling.

What NOT To Do

• - No filler phrases: "Desigur!", "Absolut!", "With pleasure!"
• No language mixing: English sentence fragments in a Romanian response
• No transliteration of diacritics: never write "sch" for "ș" in Romanian
• No "obviously", "of course", "as you know" — condescending patterns

---

Locale-Specific Formatting

Romanian (RO)

Element	Format
Date	15 martie 2026
Currency

150 lei / 30 euro / 29,99 € |

| Phone | +40 734 270 188 / 0721 234 567 | | Address | Strada Victoriei nr. 10, București |

German (DE)

Element	Format
Date	15. März 2026 / 15.03.2026
Currency

29,99 € / 150 € |

| Phone | +49 30 12345678 / 0171 2345678 | | Address | Hauptstraße 10, 10115 Berlin |

---

Cultural Communication Norms

German (DE) — Formal by Default

• - Use Sie (capitalized) unless the user explicitly switches to du
• Formal address: "Sie" not "du", "Ihnen" not "dir"
• Titles: Herr, Frau, Dr. are relevant in professional contexts
• Email tone: structured, clear subject lines, bullet points appreciated
• When in doubt: over-formal beats casual

Romanian (RO) — Warm but Professional

• - First name basis is standard once rapport is established
• Direct communication — say what you mean
• Professional but not stiff — Romanians appreciate warmth
• Avoid: excessive formality, bureaucratic language, impersonal templates

---

Post-Processor Integration

The i18n_processor.py script runs after the model generates output and
before it reaches the user. This is critical — it fixes mechanical errors
that slip through even when the model "knows" the correct form.

What the Post-Processor Fixes (deterministic, high-confidence):

1. 1. Missing diacritics in common words (RO: "vrau"→"vreau", DE: "Mueller"→"Müller")
2. Stray non-Latin characters — single Chinese/Russian chars inserted by M2.7
3. Merged words — obvious cases where two words lost their space
4. Whitespace errors — double spaces, trailing/leading spaces
5. Common typos — model-specific patterns that are consistent and fixable

What the Post-Processor Does NOT Fix:

• - Grammar errors requiring context understanding
• Wrong word choice (semantic errors)
• Stylistic improvements
• Sentence structure problems

Integration Instructions

The agent should pipe output through the processor:

CODEBLOCK1

Invoking from SKILL.md Context

When this skill is active, every model response should be treated as raw output and passed through process(response, detected_language) before delivery. The agent never sees this step — it's transparent to the user.

---

Configuration Memory

Language preference is stored via OpenClaw memory tools:

• - Key: i18n_language — the active language code (e.g., "ro", "de")
• Key: i18n_mode — "auto" or "manual"
• Key: i18n_detection_confidence — last confidence score (0.0–1.0)

When i18n_mode is "auto", the agent re-evaluates language on each session
start but respects manual overrides until reset.

---

Quick Reference

Command	Effect
"Set language to Romanian"	Lock to RO
"Setează limba la română"

Lock to RO |
| "Stelle Sprache auf Deutsch" | Lock to DE |
| "Reset language to auto-detect" | Back to auto |

Symbol	Meaning
ă â î ș ț	Romanian diacritics
ü ö ä ß

German diacritics |
| 🌐 | Language indicator |
| 🌍 | i18n skill active |

---

Error Handling

If the post-processor encounters ambiguous text:

• - Do not change if confidence < 95%
• Log the case for review
• Never risk a false positive that makes output worse

Rule: The cost of a missed diacritic is lower than the cost of a wrong one.