
bkit-rules
by popup-studio-ai
bkit Vibecoding Kit - PDCA methodology + Claude Code mastery for AI-native development
SKILL.md
name: bkit-rules description: | Core rules for bkit plugin. PDCA methodology, level detection, agent auto-triggering, and code quality standards. These rules are automatically applied to ensure consistent AI-native development.
Use proactively when user requests feature development, code changes, or implementation tasks.
Triggers: bkit, PDCA, develop, implement, feature, bug, code, design, document, 개발, 기능, 버그, 코드, 설계, 문서, 開発, 機能, バグ, 开发, 功能, 代码
Do NOT use for: documentation-only tasks, research, or exploration without code changes. hooks: PreToolUse: - matcher: "Write|Edit" hooks: - type: command command: "node ${CLAUDE_PLUGIN_ROOT}/scripts/pre-write.js" timeout: 5000 PostToolUse: - matcher: "Write" hooks: - type: command command: "node ${CLAUDE_PLUGIN_ROOT}/scripts/pdca-post-write.js" timeout: 5000
bkit Core Rules
Automatically applied rules that don't require user commands.
1. PDCA Auto-Apply Rules
No Guessing: If unsure, check docs → If not in docs, ask user SoR Priority: Code > CLAUDE.md > docs/ design documents
| Request Type | Claude Behavior |
|---|---|
| New feature | Check docs/02-design/ → Design first if missing |
| Bug fix | Compare code + design → Fix |
| Refactoring | Current analysis → Plan → Update design → Execute |
| Implementation complete | Suggest Gap analysis |
Template References
| Document Type | Template Path |
|---|---|
| Plan | ${CLAUDE_PLUGIN_ROOT}/templates/plan.template.md |
| Design | ${CLAUDE_PLUGIN_ROOT}/templates/design.template.md |
| Analysis | ${CLAUDE_PLUGIN_ROOT}/templates/analysis.template.md |
| Report | ${CLAUDE_PLUGIN_ROOT}/templates/report.template.md |
2. Level Auto-Detection
Detection Order
- Check CLAUDE.md for explicit Level declaration
- File structure based detection
Enterprise (2+ conditions met)
- infra/terraform/ folder
- infra/k8s/ or kubernetes/ folder
- services/ folder (2+ services)
- turbo.json or pnpm-workspace.yaml
- docker-compose.yml
- .github/workflows/ (CI/CD)
Dynamic (1+ conditions met)
- bkend settings in .mcp.json
- lib/bkend/ or src/lib/bkend/
- supabase/ folder
- firebase.json
Starter
None of the above conditions met.
Level-specific Behavior
| Aspect | Starter | Dynamic | Enterprise |
|---|---|---|---|
| Explanation | Friendly, avoid jargon | Technical but clear | Concise, use terms |
| Code comments | Detailed | Core logic only | Architecture only |
| Error handling | Step-by-step guide | Technical solutions | Brief cause + fix |
| PDCA docs | Simple | Feature-specific | Detailed architecture |
| Primary Agent | starter-guide | bkend-expert | enterprise-expert |
| Reference Skill | starter | dynamic | enterprise |
Level Upgrade Signals
- Starter → Dynamic: "Add login", "Save data", "Admin page"
- Dynamic → Enterprise: "High traffic", "Microservices", "Own server"
Hierarchical CLAUDE.md Rules
project/
├── CLAUDE.md # Project-wide (always reference)
├── services/CLAUDE.md # Backend work context
├── frontend/CLAUDE.md # Frontend work context
└── infra/CLAUDE.md # Infrastructure context
Rule: Area-specific rules > Project-wide rules
3. Agent Auto-Trigger Rules
Level-Based Selection
When user requests feature development:
- Detect project level
- Invoke appropriate agent automatically
Task-Based Selection
| User Intent | Auto-Invoke Agent |
|---|---|
| "code review", "security scan" | code-analyzer |
| "design review", "spec check" | design-validator |
| "gap analysis" | gap-detector |
| "report", "summary" | report-generator |
| "QA", "log analysis" | qa-monitor |
| "pipeline", "which phase" | pipeline-guide |
Proactive Suggestions
After completing major tasks, suggest relevant agents.
Do NOT Auto-Invoke When
- User explicitly declines
- Task is trivial
- User wants to understand process
- Agent already invoked for same task
4. Code Quality Standards
Pre-coding Checks
- Does similar functionality exist? Search first
- Check utils/, hooks/, components/ui/
- Reuse if exists; create if not
Core Principles
DRY: Extract to common function on 2nd use SRP: One function, one responsibility No Hardcoding: Use meaningful constants Extensibility: Write in generalized patterns
Self-Check After Coding
- Same logic exists elsewhere?
- Can function be reused?
- Hardcoded values present?
- Function does only one thing?
When to Refactor
- Same code appears 2nd time
- Function exceeds 20 lines
- if-else nests 3+ levels
- Same parameters passed to multiple functions
5. Task Classification
Classify tasks to apply appropriate PDCA level:
| Classification | Content Size | PDCA Level | Action |
|---|---|---|---|
| Quick Fix | < 50 chars | None | Execute immediately |
| Minor Change | 50-200 chars | Lite | Show summary, proceed |
| Feature | 200-1000 chars | Standard | Check/create design doc |
| Major Feature | > 1000 chars | Strict | Require design, user confirmation |
Classification Keywords
Quick Fix: fix, typo, correct, adjust, tweak Minor Change: improve, refactor, enhance, optimize, update Feature: add, create, implement, build, new feature Major Feature: redesign, migrate, architecture, overhaul, rewrite
Score
Total Score
Based on repository quality metrics
SKILL.mdファイルが含まれている
ライセンスが設定されている
100文字以上の説明がある
GitHub Stars 100以上
1ヶ月以内に更新
10回以上フォークされている
オープンIssueが50未満
プログラミング言語が設定されている
1つ以上のタグが設定されている
Reviews
Reviews coming soon


