content-review

name	content-review
description	Expert content writer review of in-game text. Use this when asked to review dialogue, localization, lore consistency, tone, translation quality, or when changelog entries are added or modified.

Content Review Process

Review all player-facing text in the game for quality, consistency, and authenticity. The game is set in a fictional Savoie ski resort — tone should blend professional ski operations with warm Savoyard hospitality and humor.

Scope: Only in-game content that players see (localization strings, dialogue, changelog overlay, level names). Do NOT review repository markdown files (ROADMAP.md, ARCHITECTURE.md, TESTING.md, etc.) — those are developer-facing documentation.

Phase 1: Content extraction

Use explore agents to gather all content from these sources:

1. src/config/locales/*.ts — All UI strings, dialogues, taunts, tutorials, changelog across 14 languages (fr, en, de, it, es, sv, nb, fi, cs, pl, tr, sk, ja, ko)
2. src/config/levels.ts — Level names, narrative context, difficulty descriptions

Phase 2: Review dimensions

Evaluate content across these dimensions, launching parallel explore agents:

1. Voice & tone consistency

   • Characters should have distinct voices: Jean-Pierre (gruff mentor, uses "petit"), Thierry (cautious/technical), Marie (warm café owner), Émilie (competitive/modern)
   • Characters must introduce themselves on first appearance (L2 Émilie, L5 Thierry, first restaurant Marie) — portrait shows the name, but dialogue should include a natural self-intro
   • Characters must NOT introduce themselves on subsequent appearances — no "It's Émilie" if she already spoke in a prior level
   • Characters must speak in first person when they are the introSpeaker — never refer to themselves in third person
   • Failure taunts should vary between humorous, dramatic, and sympathetic — never cruel or repetitive
   • Tutorial text should be clear and encouraging for beginners
   • Professional ski operations vocabulary where appropriate (damage, chenillette, canon à neige)

2. Savoyard authenticity

   • Food/drink references should be accurate (tartiflette, reblochon, génépi, vin chaud — not generic "cheese" or "wine")
   • Resort terminology should match real Savoie resorts (ESF, remontées mécaniques, damage)
   • Place names and character names should feel authentically Savoyard
   • Cultural references should be specific (FIS competitions, ski de fond, Beaufortain)

3. Localization quality

   • French is the source language — check that it reads naturally with proper register
   • Translations should be idiomatic, not literal — each language should feel native
   • Technical ski/grooming terms should use correct local terminology per language (German: Pistenraupe; Italian: gatto delle nevi; Spanish: máquina pisapistas)
   • The groom button label must use each language's specific grooming verb (FR=Damer, DE=Präparieren, IT=Battere, ES=Pisar, SV=Pista, FI=Tamppaa, JA=圧雪, KO=정설), not generic "prepare/maintain"
   • Humor and wordplay should be adapted, not translated literally (e.g., "yard sale" → local wipeout idiom, "soup" for slush → local ski slush term)
   • Check for missing translations, placeholder text, or English left in non-English languages — scan ALL keys, especially short UI labels (skiStyle, skiTricks, skiBestCombo, quitGame, showDebug)
   • Register consistency: the game uses informal address throughout (FR=tu, DE=du, CS=ty, SK=ty, etc.). Scan ALL taunts, tutorials, dialogues, hazard warnings, and credits for formal forms (FR=vous, DE=Sie, CS=vy/jste, SK=vy/ste). Every line must use informal register.
   • Immersion: the game is set in Savoie, France. Keep French references like Météo France, ESF — do not replace with local equivalents (SMHI, 気象庁, etc.). Do not substitute local organizations (e.g., PL "GOPR" for ski patrol — use generic terms like "patrol narciarski").
   • Factual consistency across locales: colors (jaune=yellow, not orange/amber), currency conversions, proper nouns must match the FR source exactly
   • Cross-locale term consistency: the SAME concept must use the SAME word within each language. Key terms to unify: rescue/patrol team, grooming machine, ski run, winch, fuel. Run a dedicated audit comparing taunt/dialogue keys that reference the same concept (e.g., tauntCliff3 and tauntAvalanche5 both mention the patrol team — must be the same word).
   • CJK readability: JA and KO use non-Latin scripts. Character names must use katakana (JA: ジャン=ピエール, エミリー, ティエリ, マリー) or hangul (KO: 장피에르, 에밀리, 티에리, 마리) — never Latin script mid-text. Food/drink terms should use native script (JA: ホットワイン not "vin chaud", KO: 뱅쇼 not "vin chaud"). French level names and "Météo France" may stay in Latin as recognizable proper nouns.
   • Nickname clarity: "JP" (for Jean-Pierre) is opaque outside French-speaking locales, especially in JA (JP=Japan). Use the full localized name in JA/KO; "JP" is acceptable in Latin-script locales where Jean-Pierre was introduced earlier.
   • Wordplay adaptation: Hyphenated non-words like "de-sicurato" (IT), "des-aseguraste" (ES) don't work as wordplay in the target language — rephrase using natural idioms. Only use wordplay that works in the target language (e.g., DE "entsichert" is a real word).
   • Groomer terminology by language: FR=dameuse, EN=groomer, DE=Raupe/Pistenraupe, IT=gatto delle nevi, ES=máquina, SV=maskin, NB=maskin, FI=rinnekone, CS=rolba, SK=rolba (NOT fréza), PL=ratrak, TR=araç, JA=圧雪車, KO=설상차

4. Narrative coherence

   • Difficulty progression should be reflected in dialogue tone (encouraging → challenging → intense)
   • Character appearances should match the level context (Marie at lower pistes, Thierry at hazardous levels)
   • Dialogue must match level mechanics — don't tell players to use equipment they don't have (e.g., winch on hasWinch: false levels)
   • One-shot dialogues (steep warning, Marie intro) persist via localStorage — verify they only trigger once, not per-level
   • Lore references should be consistent across levels (no contradictions in backstory)
   • Avoid opaque acronyms or jargon that players won't know (e.g., "PIDA" → "les pisteurs", "ESF" → "l'école de ski")
   • Bonus objectives should feel achievable and well-described
   • Players traverse levels bottom-to-top — zone encounter order matters. Higher startY = lower on map = encountered FIRST. Verify that tutorial/warning text appears on the gentler zone before the deadly one.

5. Clarity & UX writing

   • Button labels should be unambiguous and action-oriented
   • Error/warning messages should tell the player what happened AND what to do
   • Tutorial hints should be concise — players should understand in <3 seconds
   • Accessibility labels should be descriptive and screen-reader friendly
   • Check for hardcoded English strings in non-locale .ts files (e.g., onStatus('Saved!') instead of t('saved')) — all player-visible text must go through t()

6. Dead key hygiene

   • After changes, search for // UNUSED keys in locale files that may now be used (grep source code for the key name as a quoted string)
   • When commenting out keys, verify they aren't dynamically constructed (e.g., showDialogue() probes key + 'Touch' / key + 'Gamepad' variants)
   • t() in dev mode throws on missing keys — E2E tests will catch any removed key that's still referenced

7. Changelog conciseness

   • Each date entry should have 3–5 items max — highlight what players will notice
   • Lead with new content (levels, wildlife, characters), not technical fixes
   • Consolidate all bug fixes into a single catch-all line (e.g. "🔧 Corrections tactiles et accessibilité")
   • Drop items players won't notice (internal refactors, signage standards, sensitivity ranges, touch target sizes)
   • Keep entries short: one emoji + 2–3 words per line. No sub-clauses, no parenthetical details, no enumerations
   • Reference style (Feb 3–8 entries):

     🏔️ Le Glacier et Coupe des Aiguilles
     🦅 Faune alpine
     🎨 Portraits des personnages
     🎮 Navigation clavier/manette dans les paramètres
     🔧 Corrections tactiles, affichage et accessibilité
     

   • Bad patterns to avoid:
     • ❌ "11 niveaux : Le Glacier (treuil) et Coupe des Aiguilles (finale FIS)" — too many details
     • ❌ "Faune alpine : bouquetins, chamois, marmottes, lièvres et renards" — enumeration belongs in gameplay, not changelog
     • ❌ "Cibles tactiles agrandies (pause, plein écran, passer)" — players don't care which targets
     • ❌ "Signalétique conforme NF S52-102" — regulatory references mean nothing to players

Phase 3: Cross-model consultation

For significant content reviews (new dialogues, major localization changes):

• When running as Opus: consult gpt-5.2 for creative writing quality and gemini-3-pro-preview for multilingual/cultural accuracy
• When running as any other model: consult claude-opus-4.6
• Ask reviewers to evaluate tone, idiom quality, and cultural accuracy — don't ask them to re-run greps or key parity checks you already did. Cross-model value is in linguistic/cultural judgment, not mechanical validation.

Skip for minor reviews (typo fixes, single-string changes).

Phase 4: Prioritized findings

Severity	Criteria	Examples
HIGH	Incorrect/misleading information, broken UX	Wrong control hints, untranslated strings, confusing error messages
MEDIUM	Tone inconsistency, weak translations, missed lore	Generic phrasing where specific Savoyard detail fits, literal translations
LOW	Polish opportunities, enhanced flavor	Additional humor variants, richer cultural references, subtle wordplay

Track findings in SQL:

CREATE TABLE IF NOT EXISTS content_review (
    id TEXT PRIMARY KEY,
    severity TEXT NOT NULL,
    category TEXT NOT NULL,
    language TEXT,
    title TEXT NOT NULL,
    description TEXT,
    status TEXT DEFAULT 'open'
);

Phase 5: Implementation

For each fix:

1. French first — Write/fix the French source text, then adapt to other languages
2. Adapt, don't translate — Each language version should feel native and natural
3. Preserve placeholders — {keys}, {groomKey}, {winchKey} must remain intact in all languages
4. Test string length — Long strings may overflow UI elements; check visually
5. Maintain changelog — Player-visible text changes should be noted in the in-game changelog

Phase 6: Verification

LLM review catches semantic issues (wrong meaning, weak voice, bad idioms) but consistently misses mechanical patterns like formal register across 300-line files. Always run both LLM review and grep validation.

6a. Mechanical grep validation (run these exact commands)

cd src/config/locales

# Formal register — must return zero non-UNUSED lines
grep -n 'vous\|Vous\|votre' fr.ts | grep -v UNUSED
grep -n 'jste\|ejte\|ujte\|ějte' cs.ts | grep -v UNUSED
grep -n 'ste \|ajte\|ujte' sk.ts | grep -v UNUSED
grep -n 'ınız\|iniz\|unuz\|ünüz\|sınız\|siniz' tr.ts | grep -v UNUSED

# Untranslated English UI keys — must return zero
for f in fr de it es sv nb fi cs pl sk ja ko tr; do
  grep -P '"(Quit Game|Best combo|Debug Overlay)"' ${f}.ts | grep -v UNUSED
done

# Wrong marker colors — must return zero
grep -rn 'orange\|オレンジ\|주황\|turuncu\|pomarańcz\|oranž' *.ts | grep -v UNUSED | grep -i intro

# Local weather agency — must return zero
grep -rn 'SMHI\|気象庁\|기상청\|Ilmatieteenlaitos\|Meteorologisk' *.ts | grep -v UNUSED

# Placeholder parity — counts must match en.ts
for key in winchKey groomKey keys; do
  en=$(grep -c "{$key}" en.ts)
  for f in fr de it es sv nb fi cs pl sk ja ko tr; do
    n=$(grep -c "{$key}" ${f}.ts)
    [ "$n" != "$en" ] && echo "MISMATCH: $f {$key} $n vs EN $en"
  done
done

Cross-locale term consistency check — for each language, verify patrol/team term is the same in tauntCliff3, tauntSkiWipeout3, tauntAvalanche3, tauntAvalanche5:

cd src/config/locales
for f in fr en de it es sv nb fi cs pl sk ja ko tr; do
  terms=$(grep -oP 'tauntCliff3:.*"|tauntSkiWipeout3:.*"|tauntAvalanche3:.*"|tauntAvalanche5:.*"' ${f}.ts | head -4)
  echo "=== $f ===" && echo "$terms"
done

CJK character name check — JA must use katakana, KO must use hangul for all character names:

# Must return zero (no Latin character names in JA/KO text values)
grep -n 'Jean-Pierre\|Émilie\|Thierry\b\|Marie\b' ja.ts ko.ts | grep -v UNUSED | grep -v '//' | grep -v 'Briefing'

# SK must use rolba, not fréza
grep -n 'fréz' sk.ts | grep -v UNUSED

Any grep hit (excluding known false positives like DE "Sie" meaning "she") must be fixed before proceeding.

6b. Automated tests

1. Run npx vitest run tests/unit-js/localization.test.js — key parity, changelog structure
2. Run TypeScript compilation (npx tsc --noEmit) — type checking

6c. Cross-model review (for large changes)

For reviews touching 5+ locales, get semantic confirmation from at least one other model (gemini-3-pro-preview or gpt-5.1-codex) — idiom quality, voice consistency, terminology accuracy. Grep validation does not need cross-model confirmation.

Content style guide

Tone spectrum by context

Context	Tone	Example
Tutorial	Warm, clear, encouraging	"Utilise les flèches pour te déplacer sur la piste"
Success	Celebratory, proud	"Magnifique ! La piste est parfaite pour les skieurs"
Failure (mild)	Sympathetic, humorous	"Même les meilleurs se prennent une gamelle parfois"
Failure (dramatic)	Theatrical, never mean	"400 000€ de chenillette dans le ravin... l'assurance va adorer"
Hazard warning	Urgent, professional	"Attention — risque d'avalanche élevé dans ce secteur"
Lore/flavor	Authentic, local color	"Marie a préparé sa fameuse tartiflette au reblochon"

Character voice reference

• Jean-Pierre: 30+ years experience, dry humor, uses professional jargon, occasionally nostalgic about "the old days"
• Thierry: Safety-conscious, precise, slightly anxious about hazards, references technical specs
• Marie: Warm hospitality, food-focused, knows everyone in the resort, motherly
• Émilie: Younger generation, enthusiastic, competitive, uses modern expressions

Forbidden patterns

• Generic English gaming clichés ("Game Over", "You Died", "Try Again")
• Condescending failure messages
• Cultural stereotypes or caricatures
• Mixing registers within a single character's dialogue
• Placeholder text left in production (TODO, FIXME, lorem ipsum)