Files
Claude-Code-Workflow/.codex/skills/roadmap-with-file/SKILL.md

28 KiB

name, description, argument-hint
name description argument-hint
roadmap-with-file Strategic requirement roadmap with iterative decomposition and issue creation. Outputs roadmap.md (human-readable, single source) + issues.jsonl (machine-executable). Handoff to team-planex. [-y|--yes] [-c|--continue] [-m progressive|direct|auto] "requirement description"

Auto Mode

When --yes or -y: Auto-confirm strategy selection, use recommended mode, skip interactive rounds.

Roadmap-with-file Skill

Usage

$roadmap-with-file "Implement user authentication system with OAuth and 2FA"
$roadmap-with-file -m progressive "Build real-time notification system"
$roadmap-with-file -m direct "Refactor payment module"
$roadmap-with-file -m auto "Add data export feature"
$roadmap-with-file --continue "auth system"
$roadmap-with-file -y "Implement caching layer"

Flags:

  • -y, --yes: Skip all confirmations (auto mode)
  • -c, --continue: Continue existing session
  • -m, --mode: Strategy selection (progressive / direct / auto)

Context Source: cli-explore-agent (optional) + requirement analysis Output Directory: .workflow/.roadmap/{session-id}/ Core Output: roadmap.md (single source, human-readable) + issues.jsonl (global, machine-executable)


Subagent API Reference

spawn_agent

Create a new subagent with task assignment.

const agentId = spawn_agent({
  message: `
## TASK ASSIGNMENT

### MANDATORY FIRST STEPS (Agent Execute)
1. **Read role definition**: ~/.codex/agents/{agent-type}.md (MUST read first)
2. Read: .workflow/project-tech.json
3. Read: .workflow/project-guidelines.json

## TASK CONTEXT
${taskContext}

## DELIVERABLES
${deliverables}
`
})

wait

Get results from subagent (only way to retrieve results).

const result = wait({
  ids: [agentId],
  timeout_ms: 600000  // 10 minutes
})

if (result.timed_out) {
  // Handle timeout - can continue waiting or send_input to prompt completion
}

send_input

Continue interaction with active subagent (for clarification or follow-up).

send_input({
  id: agentId,
  message: `
## CLARIFICATION ANSWERS
${answers}

## NEXT STEP
Continue with plan generation.
`
})

close_agent

Clean up subagent resources (irreversible).

close_agent({ id: agentId })

Output Artifacts

Single Source of Truth

Artifact Purpose Consumer
roadmap.md Human-readable strategic roadmap with all context Human review, team-planex handoff
.workflow/issues/issues.jsonl Global issue store (appended) team-planex, issue commands

Why No Separate JSON Files?

Original File Why Removed Where Content Goes
strategy-assessment.json Duplicates roadmap.md content Embedded in roadmap.md Strategy Assessment section
exploration-codebase.json Single-use intermediate Embedded in roadmap.md Codebase Context appendix

Overview

Strategic requirement roadmap with iterative decomposition. Creates a single roadmap.md that evolves through discussion, with issues persisted to global issues.jsonl for execution.

Core workflow: Understand → Decompose → Iterate → Validate → Handoff

┌─────────────────────────────────────────────────────────────────────────┐
│                    ROADMAP ITERATIVE WORKFLOW                            │
├─────────────────────────────────────────────────────────────────────────┤
│                                                                          │
│  Phase 1: Requirement Understanding & Strategy                           │
│     ├─ Parse requirement: goal / constraints / stakeholders              │
│     ├─ Assess uncertainty level → recommend mode                         │
│     ├─ User confirms strategy (-m skips, -y auto-selects)                │
│     └─ Initialize roadmap.md with Strategy Assessment                    │
│                                                                          │
│  Phase 2: Decomposition & Issue Creation                                 │
│     ├─ cli-roadmap-plan-agent executes decomposition                     │
│     ├─ Progressive: 2-4 layers (MVP→Optimized) with convergence          │
│     ├─ Direct: Topological task sequence with convergence                │
│     ├─ Create issues via ccw issue create → issues.jsonl                 │
│     └─ Update roadmap.md with Roadmap table + Issue references           │
│                                                                          │
│  Phase 3: Iterative Refinement (Multi-Round)                             │
│     ├─ Present roadmap to user                                           │
│     ├─ Feedback: Approve | Adjust Scope | Modify Convergence | Replan    │
│     ├─ Update roadmap.md with each round                                 │
│     └─ Repeat until approved (max 5 rounds)                              │
│                                                                          │
│  Phase 4: Handoff                                                        │
│     ├─ Final roadmap.md with Issue ID references                         │
│     ├─ Options: team-planex | first wave | view issues | done            │
│     └─ Issues ready in .workflow/issues/issues.jsonl                     │
│                                                                          │
└─────────────────────────────────────────────────────────────────────────┘

Dual Modes

Mode Strategy Best For Decomposition
Progressive MVP → Usable → Refined → Optimized High uncertainty, need validation 2-4 layers, each with full convergence
Direct Topological task sequence Clear requirements, confirmed tech Tasks with explicit inputs/outputs

Auto-selection logic:

  • ≥3 high uncertainty factors → Progressive
  • ≥3 low uncertainty factors → Direct
  • Otherwise → Ask user preference

Output Structure

.workflow/.roadmap/RMAP-{slug}-{date}/
└── roadmap.md                  # ⭐ Single source of truth
                                #   - Strategy Assessment (embedded)
                                #   - Roadmap Table
                                #   - Convergence Criteria per Issue
                                #   - Codebase Context (appendix, if applicable)
                                #   - Iteration History

.workflow/issues/issues.jsonl   # Global issue store (appended)
                                #   - One JSON object per line
                                #   - Consumed by team-planex, issue commands

roadmap.md Template

# Requirement Roadmap

**Session**: RMAP-{slug}-{date}
**Requirement**: {requirement}
**Strategy**: {progressive|direct}
**Status**: {Planning|Refining|Ready}
**Created**: {timestamp}

---

## Strategy Assessment

- **Uncertainty Level**: {high|medium|low}
- **Decomposition Mode**: {progressive|direct}
- **Assessment Basis**: {factors summary}
- **Goal**: {extracted goal}
- **Constraints**: {extracted constraints}
- **Stakeholders**: {extracted stakeholders}

---

## Roadmap

### Progressive Mode
| Wave | Issue ID | Layer | Goal | Priority | Dependencies |
|------|----------|-------|------|----------|--------------|
| 1 | ISS-xxx | MVP | ... | 2 | - |
| 2 | ISS-yyy | Usable | ... | 3 | ISS-xxx |

### Direct Mode
| Wave | Issue ID | Title | Type | Dependencies |
|------|----------|-------|------|--------------|
| 1 | ISS-xxx | ... | infrastructure | - |
| 2 | ISS-yyy | ... | feature | ISS-xxx |

---

## Convergence Criteria

### ISS-xxx: {Issue Title}
- **Criteria**: [testable conditions]
- **Verification**: [executable steps/commands]
- **Definition of Done**: [business language, non-technical]

### ISS-yyy: {Issue Title}
...

---

## Risks

| Risk | Severity | Mitigation |
|------|----------|------------|
| ... | ... | ... |

---

## Iteration History

### Round 1 - {timestamp}
**User Feedback**: {feedback summary}
**Changes Made**: {adjustments}
**Status**: {approved|continue iteration}

---

## Codebase Context (Optional)

*Included when codebase exploration was performed*

- **Relevant Modules**: [...]
- **Existing Patterns**: [...]
- **Integration Points**: [...]

Issues JSONL Specification

Location & Format

Path: .workflow/issues/issues.jsonl
Format: JSONL (one complete JSON object per line)
Encoding: UTF-8
Mode: Append-only (new issues appended to end)

Record Schema

{
  "id": "ISS-YYYYMMDD-NNN",
  "title": "[LayerName] goal or [TaskType] title",
  "status": "pending",
  "priority": 2,
  "context": "Markdown with goal, scope, convergence, verification, DoD",
  "source": "text",
  "tags": ["roadmap", "progressive|direct", "wave-N", "layer-name"],
  "extended_context": {
    "notes": {
      "session": "RMAP-{slug}-{date}",
      "strategy": "progressive|direct",
      "wave": 1,
      "depends_on_issues": []
    }
  },
  "lifecycle_requirements": {
    "test_strategy": "unit",
    "regression_scope": "affected",
    "acceptance_type": "automated",
    "commit_strategy": "per-issue"
  }
}

Query Interface

# By ID (detail view)
ccw issue list ISS-20260227-001

# List all with status filter
ccw issue list --status planned,queued
ccw issue list --brief  # JSON minimal output

# Queue operations (wave-based execution)
ccw issue queue list              # List all queues
ccw issue queue dag               # Get dependency graph (JSON)
ccw issue next --queue <queue-id> # Get next task

# Execute
ccw issue queue add <issue-id>    # Add to active queue
ccw issue done <item-id>          # Mark completed

Note

: Issues are tagged with wave-N in tags[] field for filtering. Use --brief for programmatic parsing.


Implementation

Session Initialization

const getUtc8ISOString = () => new Date(Date.now() + 8 * 60 * 60 * 1000).toISOString()

// Parse flags
const AUTO_YES = $ARGUMENTS.includes('--yes') || $ARGUMENTS.includes('-y')
const continueMode = $ARGUMENTS.includes('--continue') || $ARGUMENTS.includes('-c')
const modeMatch = $ARGUMENTS.match(/(?:--mode|-m)\s+(progressive|direct|auto)/)
const requestedMode = modeMatch ? modeMatch[1] : 'auto'

// Clean requirement text (remove flags)
const requirement = $ARGUMENTS
  .replace(/--yes|-y|--continue|-c|--mode\s+\w+|-m\s+\w+/g, '')
  .trim()

const slug = requirement.toLowerCase()
  .replace(/[^a-z0-9\u4e00-\u9fa5]+/g, '-')
  .substring(0, 40)
const dateStr = getUtc8ISOString().substring(0, 10)
const sessionId = `RMAP-${slug}-${dateStr}`
const sessionFolder = `.workflow/.roadmap/${sessionId}`

// Auto-detect continue mode
if (continueMode || file_exists(`${sessionFolder}/roadmap.md`)) {
  // Resume existing session
  const existingRoadmap = Read(`${sessionFolder}/roadmap.md`)
  // Extract current phase and continue from there
}

Bash(`mkdir -p ${sessionFolder}`)

Phase 1: Requirement Understanding & Strategy

Objective: Parse requirement, assess uncertainty, select decomposition strategy, initialize roadmap.md.

Steps:

  1. Parse Requirement

    • Extract: goal, constraints, stakeholders, keywords
  2. Assess Uncertainty

    const uncertaintyFactors = {
      scope_clarity: 'low|medium|high',
      technical_risk: 'low|medium|high',
      dependency_unknown: 'low|medium|high',
      domain_familiarity: 'low|medium|high',
      requirement_stability: 'low|medium|high'
    }
    
    // Calculate recommendation
    const highCount = Object.values(uncertaintyFactors).filter(v => v === 'high').length
    const lowCount = Object.values(uncertaintyFactors).filter(v => v === 'low').length
    
    let recommendedMode
    if (highCount >= 3) recommendedMode = 'progressive'
    else if (lowCount >= 3) recommendedMode = 'direct'
    else recommendedMode = 'progressive'  // default safer choice
    
  3. Strategy Selection (skip if -m specified or AUTO_YES)

    let selectedMode
    
    if (requestedMode !== 'auto') {
      selectedMode = requestedMode
    } else if (AUTO_YES) {
      selectedMode = recommendedMode
    } else {
      const answer = ASK_USER([
        {
          id: "strategy",
          type: "choice",
          prompt: `Decomposition strategy:\nUncertainty: ${uncertaintyLevel}\nRecommended: ${recommendedMode}`,
          options: [
            { value: "progressive", label: recommendedMode === 'progressive' ? "Progressive (Recommended)" : "Progressive" },
            { value: "direct", label: recommendedMode === 'direct' ? "Direct (Recommended)" : "Direct" }
          ],
          default: recommendedMode
        }
      ])  // BLOCKS (wait for user response)
    
      selectedMode = answer.strategy
    }
    
  4. Initialize roadmap.md

    const roadmapContent = `# Requirement Roadmap
    
    

Session: ${sessionId} Requirement: ${requirement} Strategy: ${selectedMode} Status: Planning Created: ${getUtc8ISOString()}


Strategy Assessment

  • Uncertainty Level: ${uncertaintyLevel}
  • Decomposition Mode: ${selectedMode}
  • Assessment Basis: ${factorsSummary}
  • Goal: ${extractedGoal}
  • Constraints: ${extractedConstraints}
  • Stakeholders: ${extractedStakeholders}

Roadmap

To be populated after Phase 2 decomposition


Convergence Criteria Details

To be populated after Phase 2 decomposition


Risks

To be populated after Phase 2 decomposition


Iteration History

To be populated during Phase 3 refinement


Codebase Context (Optional)

To be populated if codebase exploration was performed `

Write(${sessionFolder}/roadmap.md, roadmapContent)


**Success Criteria**:
- roadmap.md created with Strategy Assessment
- Strategy selected (progressive or direct)
- Uncertainty factors documented

---

### Phase 2: Decomposition & Issue Creation

**Objective**: Execute decomposition via `cli-roadmap-plan-agent`, create issues, update roadmap.md.

**Steps**:

1. **Optional Codebase Exploration** (if codebase detected)
```javascript
const hasCodebase = Bash(`
  test -f package.json && echo "nodejs" ||
  test -f go.mod && echo "golang" ||
  test -f Cargo.toml && echo "rust" ||
  test -f pyproject.toml && echo "python" ||
  test -d src && echo "generic" ||
  echo "none"
`).trim()

let codebaseContext = null

if (hasCodebase !== 'none') {
  const exploreAgentId = spawn_agent({
    message: `
## TASK ASSIGNMENT

### MANDATORY FIRST STEPS (Agent Execute)
1. **Read role definition**: ~/.codex/agents/cli-explore-agent.md (MUST read first)
2. Read: .workflow/project-tech.json
3. Read: .workflow/project-guidelines.json

---

## Exploration Context
- **Requirement**: ${requirement}
- **Strategy**: ${selectedMode}
- **Project Type**: ${hasCodebase}
- **Session**: ${sessionFolder}

## Exploration Focus
- Identify modules/components related to the requirement
- Find existing patterns that should be followed
- Locate integration points for new functionality
- Assess current architecture constraints

## Output
Return findings as JSON with schema:
{
"project_type": "${hasCodebase}",
"relevant_modules": [{name, path, relevance}],
"existing_patterns": [{pattern, files, description}],
"integration_points": [{location, description, risk}],
"architecture_constraints": [string],
"tech_stack": {languages, frameworks, tools}
}
`
  })

  const exploreResult = wait({
    ids: [exploreAgentId],
    timeout_ms: 120000
  })

  close_agent({ id: exploreAgentId })

  if (exploreResult.status[exploreAgentId].completed) {
    codebaseContext = exploreResult.status[exploreAgentId].completed
  }
}
  1. Execute Decomposition Agent
    const decompositionAgentId = spawn_agent({
      message: `
    

TASK ASSIGNMENT

MANDATORY FIRST STEPS (Agent Execute)

  1. Read role definition: ~/.codex/agents/cli-roadmap-plan-agent.md (MUST read first)
  2. Read: .workflow/project-tech.json
  3. Read: .workflow/project-guidelines.json

Roadmap Decomposition Task

Input Context

  • Requirement: ${requirement}
  • Selected Mode: ${selectedMode}
  • Session ID: ${sessionId}
  • Session Folder: ${sessionFolder}

Strategy Assessment

${JSON.stringify(strategyAssessment, null, 2)}

Codebase Context

${codebaseContext ? JSON.stringify(codebaseContext, null, 2) : 'No codebase detected - pure requirement decomposition'}


Mode-Specific Requirements

${selectedMode === 'progressive' ? `Progressive Mode:

  • 2-4 layers from MVP to full implementation
  • Each layer: id (L0-L3), name, goal, scope, excludes, convergence, risks, effort, depends_on
  • L0 (MVP) must be a self-contained closed loop with no dependencies
  • Scope: each feature belongs to exactly ONE layer (no overlap)
  • Layer names: MVP / Usable / Refined / Optimized` :

`Direct Mode:

  • Topologically-sorted task sequence
  • Each task: id (T1-Tn), title, type, scope, inputs, outputs, convergence, depends_on, parallel_group
  • Inputs must come from preceding task outputs or existing resources
  • Tasks in same parallel_group must be truly independent`}

Convergence Quality Requirements

  • criteria[]: MUST be testable (can write assertions or manual verification steps)
  • verification: MUST be executable (command, script, or explicit steps)
  • definition_of_done: MUST use business language (non-technical person can judge)

Expected Output

  1. Update ${sessionFolder}/roadmap.md with Roadmap table + Convergence sections
  2. Create issues via ccw issue create - append to .workflow/issues/issues.jsonl

Issue Format (for ccw issue create)

  • id: ISS-YYYYMMDD-NNN (auto-generated)
  • title: [LayerName] goal or [TaskType] title
  • context: Markdown with goal, scope, convergence criteria, verification, DoD
  • priority: small→4, medium→3, large→2
  • tags: ["roadmap", mode, wave-N, layer-name]
  • extended_context.notes: {session, strategy, wave, depends_on_issues}

Execution Steps

  1. Analyze requirement and build decomposition context

  2. Execute decomposition (internal reasoning)

  3. Validate records, check convergence quality

  4. For each decomposed item:

    • Run: ccw issue create --title "..." --context "..." --tags "..." --priority N
    • Record returned Issue ID
  5. Update roadmap.md with Issue ID references

  6. Return brief completion summary with Issue IDs ` })

    const decompositionResult = wait({ ids: [decompositionAgentId], timeout_ms: 300000 // 5 minutes for complex decomposition })

    close_agent({ id: decompositionAgentId })

    if (!decompositionResult.status[decompositionAgentId].completed) { throw new Error('Decomposition agent failed to complete') }

    const issueIds = decompositionResult.status[decompositionAgentId].completed.issueIds || []

    
    

Success Criteria:

  • Issues created in .workflow/issues/issues.jsonl
  • roadmap.md updated with Issue references
  • No circular dependencies
  • Convergence criteria testable

Phase 3: Iterative Refinement

Objective: Multi-round user feedback to refine roadmap.

Steps:

  1. Display Current Roadmap

    • Read and display Roadmap table + key Convergence criteria
    • Show issue count and wave breakdown
  2. Feedback Loop (skip if AUTO_YES)

    let round = 0
    let approved = false
    
    while (!approved && round < 5) {
      round++
    
      const feedback = ASK_USER([
        {
          id: "feedback",
          type: "choice",
          prompt: `Roadmap validation (round ${round}):\n${issueIds.length} issues across ${waveCount} waves. Feedback?`,
          options: [
            { value: "approve", label: "Approve", description: "Proceed to handoff" },
            { value: "scope", label: "Adjust Scope", description: "Modify issue scopes" },
            { value: "convergence", label: "Modify Convergence", description: "Refine criteria/verification" },
            { value: "replan", label: "Re-decompose", description: "Change strategy/layering" }
          ]
        }
      ])  // BLOCKS (wait for user response)
    
      if (feedback.feedback === 'approve') {
        approved = true
      } else {
        // CONSTRAINT: All modifications below ONLY touch roadmap.md and issues.jsonl
        // NEVER modify source code or project files during interactive rounds
        switch (feedback.feedback) {
          case 'scope':
            // Collect scope adjustments
            const scopeAdjustments = ASK_USER([
              { id: "adjustments", type: "text", prompt: "Describe scope adjustments needed:" }
            ])  // BLOCKS
    
            // Update ONLY roadmap.md Roadmap table + Convergence sections
            Edit({ path: `${sessionFolder}/roadmap.md`, /* scope changes */ })
            // Update ONLY issues.jsonl entries (scope/context fields)
    
            break
    
          case 'convergence':
            // Collect convergence refinements
            const convergenceRefinements = ASK_USER([
              { id: "refinements", type: "text", prompt: "Describe convergence refinements needed:" }
            ])  // BLOCKS
    
            // Update ONLY roadmap.md Convergence Criteria section
            Edit({ path: `${sessionFolder}/roadmap.md`, /* convergence changes */ })
    
            break
    
          case 'replan':
            // Return to Phase 2 with new strategy
            const newStrategy = ASK_USER([
              {
                id: "strategy",
                type: "choice",
                prompt: "Select new decomposition strategy:",
                options: [
                  { value: "progressive", label: "Progressive" },
                  { value: "direct", label: "Direct" }
                ]
              }
            ])  // BLOCKS
    
            selectedMode = newStrategy.strategy
            // Re-execute Phase 2 (updates roadmap.md + issues.jsonl only)
            break
        }
    
        // Update Iteration History in roadmap.md
        const iterationEntry = `
    

Round ${round} - ${getUtc8ISOString()}

User Feedback: ${feedback.feedback} Changes Made: ${changesMade} Status: continue iteration Edit({ path:${sessionFolder}/roadmap.md, old_string: "## Iteration History\n\n> To be populated during Phase 3 refinement", new_string: ## Iteration History\n${iterationEntry}` }) } }


3. **Finalize Iteration History**
```javascript
// Update final status in roadmap.md
Edit({
  path: `${sessionFolder}/roadmap.md`,
  old_string: "**Status**: Planning",
  new_string: "**Status**: Ready"
})

Success Criteria:

  • User approved OR max rounds reached
  • All changes recorded in Iteration History
  • roadmap.md reflects final state

Phase 4: Handoff

Objective: Present final roadmap, offer execution options.

Steps:

  1. Display Summary

    ## Roadmap Complete
    
    - **Session**: RMAP-{slug}-{date}
    - **Strategy**: {progressive|direct}
    - **Issues Created**: {count} across {waves} waves
    - **Roadmap**: .workflow/.roadmap/RMAP-{slug}-{date}/roadmap.md
    
    | Wave | Issue Count | Layer/Type |
    |------|-------------|------------|
    | 1 | 2 | MVP / infrastructure |
    | 2 | 3 | Usable / feature |
    
  2. Offer Options (skip if AUTO_YES)

    let nextStep
    
    if (AUTO_YES) {
      nextStep = "done"  // Default to done in auto mode
    } else {
      const answer = ASK_USER([
        {
          id: "next",
          type: "choice",
          prompt: `${issueIds.length} issues ready. Next step:`,
          options: [
            { value: "planex", label: "Execute with team-planex (Recommended)", description: `Run all ${issueIds.length} issues via team-planex` },
            { value: "wave1", label: "Execute first wave", description: "Run wave-1 issues only" },
            { value: "view", label: "View issues", description: "Display issue details" },
            { value: "done", label: "Done", description: "Save and exit" }
          ]
        }
      ])  // BLOCKS (wait for user response)
    
      nextStep = answer.next
    }
    
  3. Execute Selection

    switch (nextStep) {
      case 'planex':
        // Launch team-planex with all issue IDs
        Bash(`ccw skill team-planex ${issueIds.join(' ')}`)
        break
    
      case 'wave1':
        // Filter issues by wave-1 tag
        Bash(`ccw skill team-planex --tag wave-1 --session ${sessionId}`)
        break
    
      case 'view':
        // Display issues from issues.jsonl
        Bash(`ccw issue list --session ${sessionId}`)
        break
    
      case 'done':
        // Output paths and end
        console.log([
          `Roadmap saved: ${sessionFolder}/roadmap.md`,
          `Issues created: ${issueIds.length}`,
          '',
          'To execute later:',
          `  $team-planex ${issueIds.slice(0, 3).join(' ')}...`,
          `  ccw issue list --session ${sessionId}`
        ].join('\n'))
        break
    }
    

Success Criteria:

  • User selection executed
  • Session complete
  • All artifacts accessible

Error Handling

Error Resolution
cli-explore-agent fails Skip code exploration, proceed with pure requirement decomposition
cli-roadmap-plan-agent fails Retry once, fallback to manual decomposition prompt
No codebase Normal flow, skip exploration step
Circular dependency detected Prompt user, re-decompose
User timeout in feedback loop Save roadmap.md, show --continue command
Max rounds reached Force proceed with current roadmap
Session folder conflict Append timestamp suffix

Core Rules

  1. Start Immediately: First action is session initialization, then Phase 1 execution
  2. Single Source: All context embedded in roadmap.md, no separate JSON files
  3. Iterate on Roadmap: Use feedback rounds to refine, not recreate
  4. Testable Convergence: criteria = assertions, DoD = business language
  5. Explicit Lifecycle: Always close_agent after wait completes to free resources
  6. DO NOT STOP: Continuous workflow until handoff complete
  7. Plan-Only Modifications: Interactive feedback (Phase 3) MUST only update roadmap.md and issues.jsonl. NEVER modify source code, configuration files, or any project files during interactive rounds. Code changes happen only after handoff (Phase 4) via team-planex or other execution skills

Best Practices

  1. Clear Requirements: Detailed description → better decomposition
  2. Iterate on Roadmap: Use feedback rounds to refine convergence criteria
  3. Testable Convergence: criteria = assertions, DoD = business language
  4. Use Continue Mode: Resume to iterate on existing roadmap
  5. Wave Execution: Start with wave-1 (MVP) to validate before full execution

Usage Recommendations

When to Use Roadmap vs Other Skills:

Scenario Recommended Skill
Strategic planning, need issue tracking $roadmap-with-file
Quick task breakdown, immediate execution $lite-plan
Collaborative multi-agent planning $collaborative-plan-with-file
Full specification documents $spec-generator
Code implementation from existing plan $workflow-lite-planex

Now execute roadmap-with-file for: $ARGUMENTS