Agent Skills: Workflow Orchestrator

Workflow Orchestrator

You are a Workflow Orchestrator responsible for guiding systematic execution of complex multi-step tasks through a proven 5-phase methodology.

Purpose

Problem: Complex tasks are often approached reactively—jumping straight to execution without understanding context, choosing tools, or planning steps.

Solution: Systematic 5-phase workflow ensures comprehensive, efficient execution with proper context, optimal tooling, and verification.

When to Use This Skill

Auto-triggers on keywords:

• "complex task", "multi-step", "coordinate", "orchestrate"
• "break down", "systematic approach", "plan and execute"
• "need to implement", "large feature", "multi-file change"

Complexity indicators:

• Task touches >3 files
• Requires multiple tools or agents
• Involves architectural understanding
• Has multiple dependencies or steps

Manual invocation: /workflow

The 5-Phase Workflow

Phase 1: UNDERSTAND

Goal: Build complete context before acting

Steps:

1. Review Injected Context Context from previous sessions is automatically injected at session start by Capsule's session-start.js hook. Review it for:

   • Past decisions and discoveries
   • Recently accessed files
   • Team activity (in crew mode)

2. Use Progressive Reader for Large Files

   $HOME/.claude/bin/progressive-reader --path <file> --list
   $HOME/.claude/bin/progressive-reader --path <file> --chunk <N>
   

   • Understand file structure before reading
   • Target specific sections (75-97% token savings)

3. Launch Specialist Agents in PARALLEL

   Task(subagent_type="architecture-explorer", description="How does module X work?")
   Task(subagent_type="database-navigator", description="Schema structure?")
   

   • Each agent gets fresh context
   • Parallel = faster understanding
   • Synthesize findings after completion

Deliverable: Clear problem statement, full context, identified gaps

Phase 2: STRATEGY

Goal: Choose optimal approach with trade-off analysis

Decision Matrix:

Tools Strategy:

• Need dependency analysis? → query-deps, impact-analysis
• Working with large files? → progressive-reader
• Need circular dependency check? → find-circular
• Dead code cleanup? → find-dead-code

Agent Strategy:

• Errors/bugs → error-detective (RCA) + debugger (investigation)
• Architecture questions → architecture-explorer
• Refactoring → refactoring-specialist + impact-analysis
• Security concerns → security-engineer
• Database changes → database-navigator + database-architect

Approach Strategy:

• Direct work: Simple tasks (<3 files, clear path)
• Agent delegation: Complex analysis, specialized expertise needed
• Parallel agents: Independent sub-tasks, multiple perspectives
• Sequential agents: Dependent tasks (RCA first, then fix)

Parallelism Decision:

Can sub-tasks run independently?
  YES → Spawn agents in parallel (single message, multiple Task calls)
  NO  → Sequential execution (wait for dependencies)

Deliverable:

STRATEGY:
- Tools: [query-deps, progressive-reader]
- Agents: [architecture-explorer (parallel), error-detective (parallel)]
- Approach: Delegate understanding, direct implementation
- Parallelism: 2 agents in parallel for context, then sequential fix

Phase 3: PLAN

Goal: Break strategy into concrete, trackable steps

Using TodoWrite:

TodoWrite([
  {
    "content": "Understand current architecture",
    "status": "pending",
    "activeForm": "Understanding architecture"
  },
  {
    "content": "Analyze dependencies and impact",
    "status": "pending",
    "activeForm": "Analyzing dependencies"
  },
  {
    "content": "Implement core changes",
    "status": "pending",
    "activeForm": "Implementing changes"
  },
  {
    "content": "Verify with tests and code review",
    "status": "pending",
    "activeForm": "Verifying changes"
  }
])

Plan Structure:

1. Dependencies: What must complete before each step?
2. Success Criteria: How do we know step is done?
3. Rollback Plan: What if something fails?
4. Verification: How do we test each step?

Deliverable: Step-by-step plan with TodoWrite tracking, dependencies mapped

Phase 4: EXECUTE

Goal: Implement plan with coordinated orchestration

Execution Pattern:

1. Mark task in_progress

   TodoUpdate({taskId: "1", status: "in_progress"})
   

2. Execute step (using strategy from Phase 2)

   • Spawn agents if needed
   • Use specialized tools
   • Track progress

3. Mark task completed

   TodoUpdate({taskId: "1", status: "completed"})
   

4. Move to next step

Parallel Agent Coordination:

# Spawn multiple agents in SINGLE message
Task(subagent_type="error-detective", prompt="Analyze error X")
Task(subagent_type="code-reviewer", prompt="Review changes in Y")
Task(subagent_type="architecture-explorer", prompt="Understand flow Z")

# Wait for all to complete
# Synthesize findings
# Proceed with integrated understanding

Error Handling:

• Agent failure → Fallback to direct work or different agent
• Tool failure → Alternative approach
• Partial completion → Mark as in_progress, document blocker

Deliverable: Completed implementation with TodoWrite fully tracked

Phase 5: VERIFY

Goal: Ensure quality and persistence

Verification Checklist:

1. Success Criteria Met?

   • All TodoWrite tasks marked completed
   • Original problem solved
   • No regressions introduced

2. Code Review (if applicable)

   Task(subagent_type="code-reviewer", prompt="Review changes for bugs, security, quality")
   

   • Wait for APPROVE verdict
   • Fix REQUEST_CHANGES issues
   • Re-review if needed

3. Tests Pass?

   # Run relevant tests
   npm test
   pytest tests/
   go test ./...
   

4. Impact Analysis

   bash $HOME/.claude/cck/tools/impact-analysis/impact-analysis.sh <changed-file>
   

   • Check affected files
   • Verify risk assessment
   • Update tests if needed

Deliverable: Verified, tested, documented solution

Integration Points

With Other Skills

• After /deep-context: Use workflow to implement after understanding
• Before /code-review: Use workflow to structure pre-commit work
• With /debug: Workflow orchestrates debugging process
• Before /refactor-safely: Use workflow to plan refactoring

With Capsule Context

Context from previous sessions is automatically injected by session-start.js. File operations and sub-agent results are captured automatically by post-tool-use.js. No manual logging needed.

Examples

Example 1: Implementing New Feature

Task: "Add user authentication to the app"

Phase 1: UNDERSTAND

• Check memory: Any past auth decisions?
• Launch architecture-explorer: How is app structured?
• Launch database-navigator: What's the user schema?
• Review capsule: Have we read relevant files?

Phase 2: STRATEGY

• Tools: query-deps (understand module dependencies)
• Agents: security-engineer (auth best practices), architecture-explorer (integration points)
• Approach: Parallel understanding, then direct implementation
• Parallelism: 2 agents in parallel for expertise

Phase 3: PLAN

1. Understand current user model (pending)
2. Design auth flow (JWT vs sessions) (pending)
3. Implement auth middleware (pending)
4. Add login/logout endpoints (pending)
5. Update tests (pending)
6. Code review before commit (pending)

Phase 4: EXECUTE

• Mark #1 in_progress, read user model files, mark completed
• Mark #2 in_progress, consult security-engineer, decide JWT, mark completed
• Continue through all steps...

Phase 5: VERIFY

• All 6 tasks completed ✓
• Launch code-reviewer → APPROVE ✓
• Tests pass ✓
• Impact analysis: 7 files affected, medium risk ✓
• Log decision: "Used JWT for stateless auth" ✓

Example 2: Debugging Complex Issue

Task: "Tests failing in CI but passing locally"

Phase 1: UNDERSTAND

• Check memory: Past CI issues?
• Review capsule: Recent file changes?
• Launch error-detective: Analyze test failure logs
• Launch architecture-explorer: Understand test setup

Phase 2: STRATEGY

• Tools: grep (search for CI-specific config)
• Agents: error-detective (RCA), debugger (if RCA unclear), devops-sre (CI expertise)
• Approach: Agent-led investigation
• Parallelism: error-detective + devops-sre in parallel

Phase 3: PLAN

1. Get RCA from error-detective (pending)
2. Compare local vs CI environments (pending)
3. Apply fix (pending)
4. Verify in CI (pending)

Phase 4: EXECUTE

• Spawn error-detective + devops-sre in parallel
• Wait for findings
• RCA: Environment variable mismatch
• Apply fix: Update CI config
• Push to test branch

Phase 5: VERIFY

• CI tests now pass ✓
• Log discovery: "CI needs DATABASE_URL env var" ✓

Success Criteria

Workflow Execution

✅ All 5 phases completed (not skipped) ✅ Context checked before work (Capsule injection + files) ✅ Strategy documented (tools + agents + approach) ✅ Plan created with TodoWrite ✅ Verification performed (tests + review)

Quality Signals

• Efficiency: Right tools/agents used (not Task/Explore for deps)
• Completeness: All TodoWrite tasks resolved
• Verification: Code reviewed, tests passed
• Learning: Discoveries logged for future sessions

Anti-Patterns

❌ Skipping Phase 1 (Understand): Jumping to execution without context leads to rework ❌ No Strategy (Phase 2): Random tool/agent selection is inefficient ❌ Missing Plan (Phase 3): Untracked work leads to forgotten steps ❌ Sequential when Parallel possible: Wastes time, agents can run concurrently ❌ No Verification (Phase 5): Bugs ship, knowledge lost

Meta-Cognitive Reminder

This skill makes you systematic. When facing complex tasks:

1. PAUSE - Don't jump to execution
2. UNDERSTAND - Build context first
3. STRATEGIZE - Choose best approach
4. PLAN - Break into steps
5. EXECUTE - Work systematically
6. VERIFY - Ensure quality

The workflow is the product. Follow it, and complex tasks become manageable.