agent-workflow
npx skills add https://github.com/supercent-io/skills-template --skill agent-workflow
Agent 安装分布
Skill 文档
Agent Workflow Design Principles
Ralph Wiggum ê¸°ë² ê¸°ë°ì ë©í° ìì´ì í¸ ìí¬íë¡ì° ì¤ê³ ìì¹ì ëë¤. ì¬ë¬ AI ìì´ì í¸(Opus/Sonnet/Gemini/Codex)를 ì¡°ì¨íì¬ ë³µì¡í ìì¨ ê°ë° ìì ì ìííë ë°©ë²ë¡ ì ë¤ë£¹ëë¤.
When to use this skill
- ëê·ëª¨ ì½ëë² ì´ì¤ ë¶ì: 500+ íì¼ì ìì¤ ì½ë ë¶ì íì ì
- ë³µì¡í 구í ìì : ìí¤í ì² ê²°ì , ëë²ê¹ , 리í©í ë§ì´ íìí ìì
- ìì¨ ê°ë° 루í: ì¬ëì ê°ì ìì´ ì¥ìê° ì¤íëë ê°ë° ìì
- ë©í° ëª¨ë¸ íì : Opus, Sonnet, Gemini, Codex를 í¨ê» íì©í ë
- Implementation Plan ê¸°ë° ê°ë°: ì°ì ìì ê¸°ë° ì ì§ì 구í
1. Ralph Wiggum Technique ê°ì
íµì¬ ì² í
“Ralph is a Bash loop.” – Geoffrey Huntley
ì¬ì¨ ê°ì¡±ì ìºë¦í° Ralph Wiggumì²ë¼ ë¨ìíì§ë§ ëì§ê¸´ ì ê·¼ë²ì ëë¤. AI ìì´ì í¸ìê² ëì¼í ìì ì ìë£ë ëê¹ì§ ë°ë³µì ì¼ë¡ ìíìíµëë¤.
기본 ì리
# Ralph Wiggumì íµì¬ - 5ì¤ Bash ì¤í¬ë¦½í¸
while true; do
claude --prompt "ìì
ì ê³ìíì¸ì. ìë£ëë©´ COMPLETE ì¶ë ¥"
if grep -q "COMPLETE" output.log; then break; fi
done
íµì¬ íµì°°:
- ì§í ìíë LLM 컨í ì¤í¸ ìëì°ê° ìë íì¼ê³¼ Git íì¤í 리ì ì ì¥
- 컨í ì¤í¸ê° ê°ë 차면 ìë¡ì´ ìì´ì í¸ê° ì´ì´ë°ì ìì ê³ì
- íì¼ìì¤í ì´ ìí(State), Gitì´ ë©ëª¨ë¦¬(Memory)
Contextual Pressure Cooker
모ë¸ì ì ì²´ ì¶ë ¥(ì¤í¨, ì¤í í¸ë ì´ì¤, í 루ìë¤ì´ì í¬í¨)ì ë¤ì ë°ë³µì ì ë ¥ì¼ë¡ í¼ëë°±íì¬ “문맥ì ìë ¥ì¥” í¨ê³¼ë¥¼ ë§ëëë¤.
[ë°ë³µ 1] ìì
ìë â ì¤í¨ â ì¤ë¥ ì¶ë ¥
â
[ë°ë³µ 2] ì´ì ì¤ë¥ + ìì
ìë â ë¶ë¶ ì±ê³µ
â
[ë°ë³µ 3] ì´ì ê²°ê³¼ + ìì
ìë â ìë£
2. SuperWork: Opus Subagent íì©
SuperWorkë?
SuperWorkë Opus 모ë¸ì extended thinking ë¥ë ¥ì íì©íì¬ ë³µì¡í ì¶ë¡ , ë¶ì, ìì¬ê²°ì ì ìííë ìë¸ìì´ì í¸ì ëë¤.
ìí ë¶ë´
| ìì´ì í¸ | ëª¨ë¸ | ìí | ë³ë ¬ ì |
|---|---|---|---|
| SuperWork | Opus | ë¶ì, ì¢ í©, ì°ì ìì ê²°ì , ìí¤í ì² | 1 |
| Worker | Sonnet | ì½ë ê²ì, íì¼ ì½ê¸°, 구í | ìµë 500 |
| Executor | Codex | ë¹ë, í ì¤í¸, ë°°í¬ ì¤í | 1 |
| Analyst | Gemini | ëì©ë ì½ë ë¶ì (1M+ í í°) | 1 |
SuperWork ìí¬íë¡ì°
1. [Sonnet x 500] src/* ìì¤ ì½ë ë¶ì
â
2. [Sonnet x 500] specs/* ì¤í 문ìì ë¹êµ
â
3. [Opus] ë¶ì ê²°ê³¼ ì¢
í© ë° ì°ì ìì ê²°ì (SuperWork)
â
4. [Opus] @IMPLEMENTATION_PLAN.md ìì±/ì
ë°ì´í¸
â
5. [Sonnet] ì°ì ìì ììëë¡ êµ¬í
â
6. [Codex] ë¹ë ë° í
ì¤í¸ ì¤í
â
7. ë°ë³µ (ìë£ê¹ì§)
3. IMPLEMENTATION_PLAN.md ê¸°ë° ê°ë°
íì¼ êµ¬ì¡°
# Implementation Plan
## Priority 1: Critical
- [ ] Fix authentication token refresh bug
- [ ] Add rate limiting to API endpoints
## Priority 2: High
- [ ] Implement user profile caching
- [ ] Add pagination to list endpoints
## Priority 3: Medium
- [ ] Refactor database connection pooling
- [ ] Add comprehensive error logging
## Completed
- [x] Set up project structure
- [x] Configure CI/CD pipeline
ìë ì ë°ì´í¸ ê·ì¹
- ìì
ìë£ ì:
[ ]â[x]ë¡ ë³ê²½ í Completedë¡ ì´ë - ì ìì ë°ê²¬ ì: ì ì í ì°ì ììì ì¶ê°
- ë¸ë¡ì»¤ ë°ê²¬ ì: Priority 1ì¼ë¡ ì¹ê²©
- ìì¡´ì± ë°ê²¬ ì: ê´ë ¨ ìì ìì ì¡°ì
4. Multi-Agent Ralph Loop ìí¤í ì²
14 ìì´ì í¸ ìì¤í (9 Core + 5 Auxiliary)
Core Agents:
- Orchestrator: ì ì²´ ìí¬íë¡ì° ì¡°ì¨
- Planner: Implementation Plan ê´ë¦¬
- Coder: ì½ë ìì± ë° ìì
- Reviewer: ì½ë 리뷰 ë° íì§ ê²ì¦
- Tester: í ì¤í¸ ìì± ë° ì¤í
- Debugger: ë²ê·¸ ë¶ì ë° ìì
- Documenter: 문ì ìì± ë° ì ë°ì´í¸
- Refactorer: ì½ë 리í©í ë§
- Deployer: ë°°í¬ ë° ì¸íë¼ ê´ë¦¬
Auxiliary Agents:
- Security Auditor: ë³´ì ì·¨ì½ì ê²ì¬
- Performance Analyzer: ì±ë¥ ë¶ì
- Dependency Manager: ìì¡´ì± ê´ë¦¬
- Config Manager: ì¤ì ê´ë¦¬
- Adversarial Validator: ì ëì ê²ì¦
Context Preservation
Ledger ìì¤í :
.ralph/
âââ ledger.json # ì ì²´ ìí 기ë¡
âââ handoffs/ # ìì´ì í¸ ê° ì¸ìì¸ê³
â âââ coder-to-reviewer.md
â âââ reviewer-to-tester.md
âââ checkpoints/ # Git ì²´í¬í¬ì¸í¸
âââ checkpoint-001.json
âââ checkpoint-002.json
Handoff ìì:
# Handoff: Coder â Reviewer
## Completed Work
- Implemented user authentication module
- Added JWT token validation
## Files Changed
- src/auth/login.ts (new)
- src/auth/token.ts (new)
- src/middleware/auth.ts (modified)
## Notes for Reviewer
- Check token expiration logic
- Verify password hashing strength
## Next Steps
- Code review required before merge
5. í í° ê´ë¦¬ ë° Escape Hatches
í í° ì ê° ì ëµ (85-90% ê°ì)
- ì ì§ì 컨í ì¤í¸ ë¡ë©: íìí íì¼ë§ ë¡ë
- Ledger ê¸°ë° ìí ê´ë¦¬: ì ì²´ íì¤í 리 ëì ìì½
- Handoff 문ìí: ìì´ì í¸ ê° ìµì ì ë³´ë§ ì ë¬
- Git Checkpoint: ìí ë³µêµ¬ì© ì¤ë ì·
Escape Hatches (ë¹ì© ì ì´)
# ë°ëì max-iterations ì¤ì
ralph-loop --max-iterations 50 --task "Implement feature X"
# ë¹ì© íë ì¤ì
ralph-loop --max-cost $10 --task "Refactor module Y"
# ìê° ì í
ralph-loop --timeout 2h --task "Debug issue Z"
ìë ì¤ë¨ ì¡°ê±´:
- ëì¼ ì¤ë¥ 3í ì°ì ë°ì
- í í° ìì° 80% ìì§
- 무í 루í í¨í´ ê°ì§
- ì¬ì©ì ì ì ì¤ë¨ ì í¸
6. ì¤ì ìí¬íë¡ì° ìì
ìì 1: ëê·ëª¨ 리í©í ë§
# 1. ë¶ì ë¨ê³ (Sonnet x 500 ë³ë ¬)
claude task "src/* ì ì²´ íì¼ì ì½ë ì¤ë© ë¶ì"
# 2. ê³í ë¨ê³ (Opus SuperWork)
claude task --model opus "ë¶ì 결과를 ì¢
í©íê³ IMPLEMENTATION_PLAN.md ìì±. SuperWork."
# 3. ì¤í ë¨ê³ (Ralph Loop)
ralph-loop --max-iterations 100 \
--plan IMPLEMENTATION_PLAN.md \
--task "ê³íì ë°ë¼ 리í©í ë§ ìí"
ìì 2: ë²ê·¸ ìì
# 1. ë²ê·¸ ì¬í (Codex)
codex-cli shell "npm test -- --grep 'failing test'"
# 2. ìì¸ ë¶ì (Opus SuperWork)
claude task --model opus "í
ì¤í¸ ì¤í¨ ìì¸ ë¶ì. ê´ë ¨ ì½ë ê²í . SuperWork."
# 3. ìì ë° ê²ì¦ (Ralph Loop)
ralph-loop --max-iterations 20 \
--completion-promise "All tests passing" \
--task "ë²ê·¸ ìì ë° í
ì¤í¸ íµê³¼"
ìì 3: ì ê¸°ë¥ êµ¬í
# IMPLEMENTATION_PLAN.md
## Feature: User Dashboard
### Priority 1
- [ ] Design database schema for user metrics
- [ ] Create API endpoints for dashboard data
### Priority 2
- [ ] Implement React dashboard component
- [ ] Add real-time data updates via WebSocket
### Priority 3
- [ ] Write integration tests
- [ ] Add documentation
# ì ì²´ ìëí
ralph-loop --plan IMPLEMENTATION_PLAN.md \
--max-iterations 200 \
--checkpoint-interval 10 \
--task "User Dashboard ê¸°ë¥ êµ¬í"
6.5 Skill Stack Orchestration (Validation Gates)
Full-Stack Delivery Workflow
ë³µì¡í íë¡ì í¸ìì ì¬ë¬ ì¤í¬ì ìíì±íì¬ ìëí¬ìë ë리ë²ë¦¬:
## Skill Stack Sequence
### Step 1: Define Outcome
- Target deliverable: [launch, campaign, product]
- Time constraints: [deadline]
- Quality constraints: [standards]
### Step 2: Sequence Skills
1) Marketing strategy â messaging, positioning
2) Frontend design â UI/UX, design tokens
3) Image generation â visual assets
4) Video production â promo content
5) Code simplification â refactoring, cleanup
6) Presentation deck â final pitch
Validation Gates
ê° ë¨ê³ ì¬ì´ì ê²ì¦ ê²ì´í¸ë¥¼ ë°°ì¹:
validation_gates:
gate_a:
name: "Messaging & Positioning Review"
after_skill: "marketing-strategy"
checklist:
- Brand voice consistency
- Target audience alignment
- Value proposition clarity
owner: "Analyst"
gate_b:
name: "Design System Consistency"
after_skill: "frontend-design"
checklist:
- Token usage validation
- Accessibility compliance
- Responsive behavior
owner: "Analyst"
gate_c:
name: "Asset QA"
after_skill: "image-generation, video-production"
checklist:
- Brand alignment
- Resolution/format correctness
- File naming convention
owner: "Executor"
gate_d:
name: "Code Quality"
after_skill: "code-refactoring"
checklist:
- Tests passing
- Lint clean
- Behavior preserved
owner: "Executor"
gate_e:
name: "Deck Narrative & Polish"
after_skill: "presentation-builder"
checklist:
- Story arc coherent
- Visual consistency
- Speaker notes complete
owner: "Orchestrator"
Multi-Agent Role Assignment
## Role Assignment
| Agent | Role | Responsibilities |
|-------|------|------------------|
| Claude (Opus) | Orchestrator | Plan, synthesize, approve gates |
| Claude (Sonnet) | Worker | Execute skills, generate content |
| Gemini | Analyst | Deep analysis, QA reviews |
| Codex | Executor | Commands, builds, validation |
Example: B2B SaaS Launch
# Full skill stack for 2-week launch
# Week 1
## Day 1-2: Strategy
claude task "Run marketing-automation skill for positioning"
# â Gate A: Messaging review
## Day 3-4: Design
claude task "Run frontend-design skill for landing page"
# â Gate B: Design consistency check
## Day 5: Assets
claude task "Run image-generation skill for hero/social"
# â Gate C: Asset QA
# Week 2
## Day 1-2: Development
claude task "Run code-refactoring skill for codebase cleanup"
# â Gate D: Code quality check
## Day 3: Video
claude task "Run video-production skill for promo video"
# â Gate C: Asset QA
## Day 4-5: Presentation
claude task "Run presentation-builder skill for investor deck"
# â Gate E: Deck review
# Final: Launch readiness
claude task --model opus "Synthesize all outputs, final launch checklist"
Metrics Tracking
## Delivery Metrics
| Metric | Target | Actual |
|--------|--------|--------|
| Gate pass rate | 100% | - |
| Skill completion | 6/6 | - |
| Rework cycles | < 2 | - |
| Total duration | 10 days | - |
7. Best Practices
DO (ê¶ì¥)
- íì max-iterations ì¤ì : 무í 루í ë°©ì§
- IMPLEMENTATION_PLAN.md ì¬ì©: ì§í ìí© ì¶ì
- Git ì²´í¬í¬ì¸í¸ íì©: 롤백 ê°ë¥íëë¡
- Handoff 문ìí: ìì´ì í¸ ê° ì»¨í ì¤í¸ ì ë¬
- ì ì í ëª¨ë¸ ì í:
- Opus: ë³µì¡í ì¶ë¡ , ìí¤í ì² ê²°ì
- Sonnet: ë³ë ¬ ìì , ì½ë ìì±
- Codex: ëª ë ¹ ì¤í
- Gemini: ëì©ë ë¶ì
DON’T (ê¸ì§)
- 무í 루í ë°©ì¹: ë¹ì© íë° ìí
- ë¨ì¼ ìì´ì í¸ ê³¼ë¶í: ìí ë¶ë´ íì
- ìí 미ì ì¥: íì¼ìì¤í /Git íì© íì
- Handoff ìëµ: 컨í ì¤í¸ ìì¤ ë°ì
- Escape Hatch 미ì¤ì : íì ì¤ë¨ ì¡°ê±´ íì
8. Constraints
íì ê·ì¹ (MUST)
- ë¹ì© ì ì´: íì max-iterations ëë max-cost ì¤ì
- ìí ì ì¥: 모ë ì§í ìí©ì íì¼ìì¤í ì 기ë¡
- ìí ë¶ë¦¬: ìì´ì í¸ë³ ëª íí ì± ì ì ì
- ê²ì¦ ë¨ê³: ê° ë¨ê³ ìë£ í ê²ì¦ ìí
ê¸ì§ ì¬í (MUST NOT)
- 무ì í 루í: íì ì¢ ë£ ì¡°ê±´ íì
- ë¨ì¼ ìì´ì í¸ ë ì : ë³µì¡í ìì ì ë¶ì°
- ìë ê°ì ìì¡´: ìì¨ ì¤í ê°ë¥íëë¡ ì¤ê³
- 컨í ì¤í¸ ëë¹: íìí ì ë³´ë§ ì ë¬
References
- How Ralph Wiggum went from ‘The Simpsons’ to the biggest name in AI right now | VentureBeat
- The Ralph Wiggum Approach: Running AI Coding Agents for Hours | DEV Community
- A Brief History of Ralph | HumanLayer Blog
- GitHub – ghuntley/how-to-ralph-wiggum
- GitHub – mikeyobrien/ralph-orchestrator
- GitHub – alfredolopez80/multi-agent-ralph-loop
- Ralph Orchestrator Guide
Metadata
ë²ì
- íì¬ ë²ì : 1.0.0
- ìµì¢ ì ë°ì´í¸: 2026-01-10
- í¸í íë«í¼: Claude, ChatGPT, Gemini, Opencode
ê´ë ¨ ì¤í¬
íê·¸
#SuperWork #multi-agent #ralph-wiggum #orchestration #autonomous #extended-thinking #opus #sonnet