Aso Review

Update status as you progress: in_progress when starting, completed when done.

Workflow

Phase 0: Route

Verify the review target is ASO metadata.

Valid targets:

• **/metadata/** — fastlane metadata (subtitle, keywords, description per locale)

Redirect: If .tsx LP files or docusaurus.config.* detected, tell user: "LP pages should use /lp-review instead."

No metadata found: If no fastlane metadata directory exists, ask the user for the path before proceeding.

Phase 1: Discover Target Files

Glob("**/metadata/**")     → fastlane metadata

• Read ALL discovered files
• Build locale inventory (expect 14+ languages)
• Log discovered locales: ja, en, ko, zh-Hans, de, es, fr, ...

If fewer than expected locales found: Inform the user how many locales were detected and confirm whether to proceed or wait for missing locales.

Phase 2: Build Shared Context

Gather before dispatching to reviewers:

1. Full metadata text per locale (subtitle, keywords, description)
2. Locale inventory with file paths
3. Brand voice reference (see references/_localization-principles.md)
4. App category and target audience

Phase 3: Form Review Team & Run Reviews

MUST call TeamCreate to create the review team:

TeamCreate("aso-review-team")

Then spawn ALL 3 agents in ONE message using Task tool. Do NOT launch sequentially.

Each agent receives:

• Complete metadata text for all locales
• The checklist from references/_checklist-aso.md
• The localization principles from references/_localization-principles.md
• Instruction to produce per-locale findings grouped by locale

CRITICAL: Each agent MUST review EVERY detected locale independently. Not just JP and EN.

Per-agent instructions:

aso-reviewer (LEAD):

• Per-locale keyword research (keywords are NOT translated — research local search terms independently)
• Subtitle optimization within character limits (CJK ~15, Latin ~30)
• Competitive differentiation per market (competitors differ by region)
• Search volume analysis for primary keywords
• Makes final call on conflicts between ASO and naturalness

naturalness-reviewer:

• Detect "translated feel" per locale (source-language structure leaking through)
• Check formality alignment (consumer app = casual register in most locales)
• Flag literally translated idioms and keyword stuffing disguised as natural text
• Suggest specific rewrites for every flagged item

messaging-reviewer:

• Cross-locale value proposition consistency (core benefit must align)
• Tone alignment with brand voice across all locales
• Feature naming consistency (translated vs kept original)
• Flag contradicting claims between locales

Phase 4: Cross-Review Discussion

MUST use SendMessage between agents. Do NOT skip this phase.

1. SendMessage aso-reviewer's keyword suggestions to naturalness-reviewer — "Are these keyword changes still natural?"
2. SendMessage naturalness-reviewer's findings to aso-reviewer — "Can we preserve keywords while fixing naturalness?"
3. SendMessage messaging consistency issues to all agents — "Locale X contradicts locale Y on value prop"

Each agent responds with agreements, disagreements, and proposed resolutions.

The aso-reviewer (LEAD) makes final call on conflicts — especially ASO keyword vs naturalness trade-offs.

Phase 5: Synthesize Results

1. Collect final feedback from all agents after cross-review
2. Produce a consolidated report per locale

Report format:

# ASO Review Report

## Summary
[Overall assessment across all locales]

## Per-Locale Findings

### ja (Japanese)
- [Finding with specific before/after suggestion]

### ko (Korean)
- [Finding with specific before/after suggestion]

[Repeat for every detected locale]

## Cross-Locale Issues
- [Issues spanning multiple locales]

## Conflicts Resolved
- [Conflict]: [Resolution by lead reviewer]

3. MUST call TeamDelete("aso-review-team") after synthesizing results.
4. Present report to user and wait for approval before applying fixes.

Phase 6: Apply Fixes

Only after user approval.

1. Apply approved changes to metadata files
2. Run pnpm run build if applicable
3. Present diff summary

Examples

ASO Subtitle

Bad (keyword stuffing):

タスク管理・時間管理・プロジェクト管理・チーム管理アプリ

Good (natural with keyword):

チームのタスクと時間をまとめて管理

Why: The bad version crams every keyword into the subtitle, sacrificing readability. The good version includes the primary keyword naturally while communicating a clear benefit.

Korean: Translated vs Localized

Bad (translated from JP, sounds like a manual):

작업을 관리할 수 있습니다

Good (localized, casual and scannable):

할 일, 한눈에 정리

Why: The bad version uses formal -습니다 endings and technical 작업 (work/task), reading like a translated manual. The good version uses casual 할 일 (to-do) with a noun-ending phrase natural to Korean App Store copy.

German: Translated vs Localized

Bad (subject omission from JP source):

Verwaltet Aufgaben und Zeit

Good (addresses user directly):

Deine Aufgaben und Zeit im Griff

Why: The bad version omits the subject (natural in Japanese, awkward in German) and uses a bare verb. The good version addresses the user directly with "Deine" (du-form), standard for German consumer apps.

Chinese: Translated vs Localized

Bad (reads like a spec sheet):

任务管理和时间跟踪

Good (conversational, benefit-oriented):

轻松搞定每日待办

Why: The bad version lists features in formal written register. The good version uses colloquial phrasing native to the Chinese App Store, leading with the benefit.

English: AI Vocabulary in Metadata

Bad:

Delve into comprehensive task management

Good:

See your tasks. Check them off.

Why: The bad version uses AI vocabulary ("delve", "comprehensive") that signals machine-generated text. The good version is concrete, action-oriented, and sounds human-written.

Anti-Patterns

1. Do NOT launch agents sequentially — all 3 in ONE Task call
2. Do NOT skip cross-review discussion — Phase 4 is mandatory
3. Do NOT translate keywords — research local search terms independently per locale
4. Do NOT review only JP and EN — MUST review ALL detected locales
5. Do NOT merge results without discussion — agents must cross-check findings
6. Do NOT review LP pages — redirect user to /lp-review
7. Do NOT forget TeamDelete — always clean up the review team after Phase 5

Reference Files