KhamLao (ຄຳລາວ)

Modular Lao language skill for Claude.

KhamLao makes Claude write authentic, concise Lao instead of slow Thai-influenced verbose text. It cuts ~40-60% of output tokens while enforcing correct Lao register, script, and idiom.

Inspired by pordee (Thai compression skill), KhamLao adds Lao-specific quality rules on top of compression.

Architecture — one skill, one command

KhamLao is a single skill (/khamlao). The data is kept split by domain in data/*.json for ease of editing, but build_skills.py merges everything into one SKILL.md — so users see just one command, never a domain picker.

skills/khamlao/SKILL.md   (~26 KB)
├─ Quality rules + critical false-friends   (always)
└─ Vocabulary reference
   ├─ Everyday   time, family, numbers, common verbs/adj
   ├─ Cooking    food, herbs, methods, utensils, dishes
   ├─ School     education vocabulary
   ├─ Nature     animals, body parts, nature
   └─ Web / UI   auth, actions, nav, booking, forms, status

Total vocabulary: 390 entries, sourced from:

• Gemini CLI bootstrap (verified Lao output)
• Native-speaker reviewer (@nouvath07)
• Lao MOE primary school textbooks (ປ.1, ປ.2, ປ.4)
• Web/UI terms for building Lao-language interfaces (native-reviewed)

Note on scale: the whole vocabulary loads into context with the skill. That's fine at this size; once it grows past ~1–2K entries the plan is to move the long tail to a lookup tool (see "What's next").

Install

Via Claude Code plugin marketplace

claude plugin marketplace add nouvath07/khamlao
claude plugin install khamlao@khamlao

Via npx skills add

npx skills add https://github.com/nouvath07/KhamLao --skill khamlao

This installs the single khamlao skill to your universal agents directory.

Manual

Copy skills/khamlao/ to your ~/.claude/skills/.

Usage

Slash commands

Lao keyword triggers (no slash needed)

How it loads

The single khamlao skill loads when Lao output is requested. It contains:

• 7 quality rules (script, register, tense, tone, no Thai-isms, transliteration, technical IDs)
• 9 critical false-friends (most common Claude leaks)
• Compression rules
• The full vocabulary reference (everyday, cooking, school, nature)

One skill, one command — no domain picker. The whole skill (~21 KB) loads together. Most response latency for Lao comes from generating Lao output tokens (Lao tokenizes to ~5–10× English), not from this context, so a single merged skill keeps the UX simple without a meaningful speed cost at the current size.

Quality

Quality rules (always when writing Lao)

• Lao script only — never substitutes Thai characters
• Modern register — ຂ້ອຍ/ເຈົ້າ for conversation; never archaic ຂ້ານ້ອຍ
• Correct tense — ສິ for future (not Thai-influenced ຈະ)
• Lao tone marks only — ່ and ້
• No Thai-isms — ໃຜ not ใคร, ໃສ not ที่ไหน, ບໍ່ not ไม่
• Foreign words → ທັບສັບ — ໄຟລ໌, ເຊັດຊັນ, ຄອມພິວເຕີ້

Compression (token savings)

• Drops politeness particles, hedging, fillers, pleasantries
• Verbose → terse swaps (e.g., ເນື່ອງຈາກ → ເພາະ)
• Two levels: lite (gentle) and full (aggressive)

Safety boundaries (NEVER compressed)

• Code blocks, error messages, stack traces — verbatim
• File paths, URLs, identifiers — verbatim
• Security warnings, irreversible commands — written in clear standard Lao

Repo structure

khamlao/
├── data/                    ← Source-of-truth (edit here)
│   ├── critical_fixes.json
│   ├── everyday.json
│   ├── cooking.json
│   ├── school.json
│   └── nature.json
├── skills/                  ← Generated (do not edit directly)
│   └── khamlao/SKILL.md     ← Single merged skill (all domains)
├── tools/
│   ├── build_skills.py      ← Merge data/*.json into the single SKILL.md
│   ├── khamlao_checker.py   ← Quality detector (Thai-leak, register, tone)
│   ├── scrape_lao_corpus.py ← Corpus pipeline (PDF → OCR → JSONL)
│   └── archive/             ← Historical one-off scripts
├── .claude-plugin/          ← Plugin metadata
├── .github/                 ← Issue & PR templates
├── README.md, CONTRIBUTING.md, LICENSE

To add or fix vocabulary

1. Edit the relevant data/*.json file
2. Run python tools/build_skills.py
3. Commit both the data/ change and the regenerated skills/khamlao/SKILL.md

Contributing

KhamLao is community-driven. Native and fluent Lao speakers — your help is essential.

• 🐛 Report Thai-influenced output — when Claude leaks Thai vocabulary/script
• 📚 Add vocabulary — Thai → Lao mappings the current rules miss
• 🌾 Add domain glossaries — propose new domains beyond the current 4
• 📝 Fix register/tone errors

See CONTRIBUTING.md. You don't need to be a programmer — language data is most valuable.

For developers/contributors

• [ARCHITECTURE.md](ARCHITECTURE.md) — how the system works, data flow, schemas, build process
• [ROADMAP.md](ROADMAP.md) — open tasks organized by impact and difficulty
• [CONTRIBUTING.md](CONTRIBUTING.md) — how to submit changes

What's next

KhamLao today is a prompt-injected skill — it corrects Claude/Gemini's Lao output at generation time. The next step is to make those same rules usable outside an LLM, as a standalone quality layer for any Lao NLP system:

🔬 Quality / correctness layer (in design)

A Python library + eval suite that any Lao OCR, TTS, or translation pipeline can call to check its own output. Three components:

┌─────────────────────────────────────────────┐
│  Layer 3: Eval suite (benchmark runner)     │  → measure model output quality
├─────────────────────────────────────────────┤
│  Layer 2: Detector library (Python pkg)     │  → Thai-leak + register checks
├─────────────────────────────────────────────┤
│  Layer 1: Rules + test data (JSON)          │  → already 80% in data/*.json
└─────────────────────────────────────────────┘

Detector example:

from khamlao_checker import check
check("ຂ້ອຍຈະໄປເຮັດວຽກ")
# → {"thai_word_leaks": ["ຈະ"], "suggested": "ສິ",
#    "register": "modern", "score": 0.72}

Why it matters: ML teams optimize average accuracy, but no one currently checks whether Lao output is actually good Lao — free of Thai-script leaks, archaic pronouns, or wrong tense markers. That gap is what KhamLao's rules already encode for LLMs; the next step is exposing them as a reusable checker.

📚 Corpus pipeline (shipped, v1.1)

tools/scrape_lao_corpus.py — download → render → Gemini OCR → quality-filter → JSONL. Used to harvest Lao text from MOE textbooks (ປ.1–ມ.7) for training and evaluation. See data/corpora/README.md.

🛣 Roadmap details

See ROADMAP.md — 6 themes, prioritized, with effort estimates. Contributors welcome at any tier of complexity.

Known limitations

• Claude's Lao training data is significantly weaker than English/Thai/Chinese. Domain glossaries help but the skill works best paired with Gemini CLI (which has stronger native Lao).
• Skill auto-loading by context is heuristic — sometimes a domain glossary won't load when you expect it. Use the /khamlao trigger to force the core skill.
• Vocabulary is targeted at Vientiane standard Lao. Regional variants (Luang Prabang, Champasak) can be added via contribution.

License

MIT