- Copy
shared/language-template.md - Fill in all sections marked
[LANGUAGE] - Add to these files per the checklist in the template:
shared/burned-words.md- Burned words + Empty intensifiersshared/ai-markers.md- Detection patterns (openers, conclusions, transitions, hedging, punctuation tells, human-language description)shared/tone-profiles.md- Per-tone language markers (all 7 profiles)shared/specificity-ladder.md- Rung 0 to 4 examplesshared/rhythm-tables.md- Opener categories + Conjunction listSKILL.md- Stage 0 table, Stage 5 final checks, verify flag, output formatREADME.mdandREADME.ru.md- Language table row
- Add 2+ examples in
examples/[lang]-*.md - Add 1+ test file in
tests/benchmark/ai-texts/[lang]/ - Run
powershell -File scripts/validate.ps1- must pass with 0 errors - Submit PR
- Create
scenarios/[name].md - Include these sections:
- Use when: description
- Default tone: one of the 7 profiles
- Key priorities: 3-5 ordered items
- Rhythm targets: fragment spacing, max consecutive category, etc.
- What to cut: specific list
- Pipeline Recommendations: likely needed / often skippable / never skip
- Add to
SKILL.mdfile tree (line ~440) - Add to
README.mdandREADME.ru.mdarchitecture trees - Run validation script
- Submit PR
- Add to
shared/tone-profiles.md:- Universal signature (who's speaking, key traits)
- Per-language markers for all 9 languages
- Add row to length mix table in
shared/rhythm-tables.md - Add to
SKILL.mdtone table and output format - Update all scenario files that reference tone lists
- Run validation script
- Submit PR
- Start with a verb: Add, Fix, Update, Remove, Refactor
- Be specific: "Add Japanese ai-markers and burned words" not "Update shared files"
- One concern per commit where practical
Always run before submitting:
# Windows
powershell -File scripts/validate.ps1
# Linux/macOS
bash scripts/validate.shThe validator checks 11 categories of cross-reference integrity. It must pass with 0 errors. Warnings are acceptable if they are expected (e.g., em-dash references in documentation files).
When adding or editing language data:
- Markers must be authentic. Native speaker review required for new languages. Machine-translated templates are not acceptable for tone markers.
- Examples must be culturally relevant. Don't translate the WordPress security example. Use a domain that resonates with that language's audience.
- Variants noted. If a language has major regional variants (PT-BR vs PT-EU, ES vs LATAM), note the differences where they affect the humanization rules.
- Filler words in context. "Du coup" is not just a filler - it has specific usage patterns. Know them before listing them.