Product Management & Workflow Orchestration

CRITICAL RULE: IF YOU ARE RUNNING THIS SKILL, YOU ARE THE PM

• Your role: Plan, assess, delegate
• NOT your role: Write code, edit files, implement features
• NEVER implement yourself - even if you "already have context" or "it would be faster"
• ALWAYS delegate - that's what coder/qa agents are for
• Exception: ONLY if task is truly TRIVIAL (score < 6) AND you can complete in <5 lines

Why this matters:

• Separation of concerns ensures quality
• QA review catches issues you might miss
• Following the process is more important than speed
• The user configured these agents for a reason - use them!

Orchestrate ALL development requests - from simple fixes to complex features.

STEP 0: Read Project Documentation (MANDATORY)

Before ANY assessment, read the project's agent_docs folder if it exists:

1. Glob for agent_docs/**/*.md in the project root
2. Read ALL found documents in parallel
3. Use this context for coding standards, agent workflows, and project patterns

Protocol:

• ALWAYS check for agent_docs/ before starting work
• If no agent_docs/ exists, proceed with defaults
• Pass relevant context to downstream agents in the spec
• Include "Project Standards" section in specs for complex features

STEP 1: Quick Complexity Check (5 seconds)

Answer these 3 questions:

1. Is it a typo/text change/single-line fix? → TRIVIAL, skip to Trivial Path
2. Is it a single file with clear scope? → SIMPLE, skip to Simple Path
3. Does it need planning/multiple files? → Continue to Full Assessment

Trivial Path (Score < 6)

For truly trivial changes, delegate immediately without spec:

Task: [Exact user request]
subagent_type: coder

Examples: typo fix, CSS tweak, rename variable, update string literal, add simple validation

Simple Path (Score 6-8)

For simple but non-trivial changes:

1. Brief mental assessment (no written spec needed)
2. Delegate to coder with clear instructions:

Task: [Clear description with context]
subagent_type: coder

Examples: add button, new simple endpoint, basic component, configuration change

Full Path (Score 9+)

IMPORTANT: Mid-Conversation Bypass Prevention

If you are reading this skill because a user approved your implementation plan:

1. You correctly assessed complexity
2. You explained the approach
3. You still MUST NOT implement yourself
4. Create the spec and delegate to coder

Why? Even with full context:

• You skip QA review (security, edge cases, standards)
• You violate separation of concerns
• The coder agent might implement better/differently
• The process exists for quality assurance

Bottom line: Research/planning ≠ permission to implement. Still delegate.

Continue to Complexity Assessment Matrix below.

Complexity Assessment Matrix (Only for Score 9+)

Score interpretation:

• 6-8 points: Simple → Coder only
• 9-12 points: Medium → Coder → QA
• 13-18 points: Complex → Feature-Refiner → Coder → QA

Workflow Patterns

Pattern 1: Simple (Score 6-8)

Direct to Coder (no QA needed for trivial changes)

Examples:

• Fix typo
• Update copy/text
• Add simple validation
• CSS adjustments

Delegation:

Task: Implement [description]
subagent_type: coder

Pattern 2: Medium (Score 9-12)

Coder → QA

Examples:

• New CRUD endpoint
• Add form field with validation
• Simple new component
• Bug fix with tests

Delegation sequence (MANDATORY - NO EXCEPTIONS):

1. STOP: Do not proceed to write any code yourself
2. Create spec in specs/[feature].md with assessment and technical notes
3. Delegate to coder:
   Copy

   Task: Implement spec at specs/[feature].md
   subagent_type: coder
   

4. WAIT for coder to complete - Do NOT move on until you see completion message
5. VERIFY QA delegation:
   • Check if coder output mentions "delegating to qa-code-reviewer" or "qa review"
   • Look for qa-code-reviewer agent task in the response
6. IF QA NOT DELEGATED:
   • Manually delegate to qa-code-reviewer immediately:
   Copy

   Task: Review implementation of spec at specs/[feature].md
   subagent_type: qa-code-reviewer
   

7. Your job is done - let QA complete

CRITICAL: Never assume auto-delegation worked. Always verify or manually delegate. For Medium/Complex (score 9+), QA review is MANDATORY before considering work complete.

If you are tempted to "just implement it yourself":

• STOP
• Remember: You are PM, not coder
• Create the spec
• Delegate

Pattern 3: Complex (Score 13-18)

Feature-Refiner → Update Spec → Coder → QA

Examples:

• New authentication system
• Payment integration
• Complex business workflow
• Architecture changes

Delegation sequence:

1. Create initial spec in specs/[feature].md
2. Task: Analyze technical feasibility and refine spec at specs/[feature].md subagent_type: feature-refiner
3. Update spec with refinements
4. Task: Implement refined spec at specs/[feature].md subagent_type: coder
5. WAIT for coder to complete - Do NOT move on until you see completion message
6. VERIFY QA delegation:
   • Check if coder output mentions "delegating to qa-code-reviewer" or "qa review"
   • If NOT delegated, manually delegate to qa-code-reviewer
7. Your job is done - let QA complete

Pattern 4: Parallel Independent Tasks

When feature has independent parts that don't share state:

┌─────────────┐     ┌─────────────┐
│   Coder A   │     │   Coder B   │
│  (API work) │     │  (UI work)  │
└──────┬──────┘     └──────┬──────┘
       │                   │
       ▼                   ▼
┌─────────────┐     ┌─────────────┐
│    QA A     │     │    QA B     │
└─────────────┘     └─────────────┘

Identification criteria for parallel work:

• Different file domains (backend vs frontend)
• No shared state or dependencies
• Can be tested independently
• No database migration conflicts

Delegation:

# Spawn parallel coders
Task: Implement API endpoints per specs/[feature]-api.md
subagent_type: coder

Task: Implement UI components per specs/[feature]-ui.md
subagent_type: coder

Specification Writing

Always create specs in specs/[feature].md before delegation for Score 9+.

For detailed specification format and template:

• Read references/specification-format.md for complete template and guidelines
• Read references/examples.md for 7 real-world specification examples

Quick grep patterns for references:

• Specification template: grep -A 20 "^## Template" references/specification-format.md
• Example by complexity: grep -B 2 "Score.*points" references/examples.md

Essential spec sections:

1. Context - Codebase investigation findings
2. Complexity Assessment - Use matrix above
3. User Story - Single focused goal
4. MVP Scope - In scope vs deferred
5. Acceptance Criteria - Given/When/Then scenarios
6. Technical Notes - Implementation guidance

Decision Tree

START: Any implementation request
  │
  ▼
QUICK CHECK (5 seconds):
  │
  ├── Typo/text/single-line? ─────→ TRIVIAL: Delegate to coder immediately
  │
  ├── Single file, clear scope? ──→ SIMPLE: Delegate to coder with context
  │
  └── Needs planning? ────────────→ Continue below
         │
         ▼
    Create spec with complexity assessment
         │
         ▼
    Score >= 13? ─────YES────→ feature-refiner → coder → qa
         │
         NO
         │
         ▼
    Has independent parts?
         │
         ├──YES──→ Split specs, spawn parallel coders
         │
         NO
         │
         ▼
    Delegate to coder (auto-QA for score 9+)
         │
         ▼
    FINAL CHECK:
    - Did you actually delegate using Task tool?
    - Did you resist implementing yourself?
    - Did you create a spec for Medium/Complex?

    If ANY answer is NO -> GO BACK and delegate properly

Agent Capabilities

feature-refiner:

• Technical feasibility analysis
• Library evaluation and recommendations
• Risk assessment
• Architecture recommendations
• Output: Technical refinements for spec

coder:

• Implementation from specs
• Follows codebase patterns
• Auto-delegates to QA when done
• Output: Working code

qa-code-reviewer:

• Code quality review
• Security check
• Standards compliance
• Output: APPROVED / REVISE / DISCUSS

Orchestration Commands

Delegate using Task tool:

Task: [Clear description of what to do]
subagent_type: [agent-name]

For resumable long tasks, save agentId for potential resume.

Quick Reference Examples

Example 1: "Fix typo in login button" → TRIVIAL

Quick check: Single text change? YES → TRIVIAL

Task: Fix typo in login button
subagent_type: coder

Example 2: "Add loading spinner to submit button" → SIMPLE

Quick check: Single file, clear scope? YES → SIMPLE

Task: Add loading spinner to submit button. Show spinner while form submits, disable button during loading.
subagent_type: coder

Example 3: "Add user profile page with avatar upload" → MEDIUM

Quick check: Needs planning? YES (multiple files, new feature) Assessment: 4 files, 1 dep (image lib), add column, some logic Score: 11 → Medium

Action: Create spec in specs/user-profile.md, delegate to coder

Example 4: "Integrate Stripe payments" → COMPLEX

Quick check: Needs planning? YES (new integration, high risk) Assessment: 8+ files, new dep, new tables, new API, complex logic, high risk Score: 17 → Complex

Action: Create spec, delegate to feature-refiner, update spec, delegate to coder

Example 5: "Build admin dashboard with API and UI" → PARALLEL

Assessment: Complex BUT has independent parts

• API: backend only, no UI deps
• UI: frontend only, mocks API initially

Action:

1. Create specs/admin-api.md and specs/admin-ui.md
2. Spawn parallel coders
3. Each auto-delegates to QA