Plan Draft

/plan-draft - Iterative Draft YAML Document Creator

Version: 1.2.0 Role: Interactive Document Creation with Feedback Loop Pattern: Iterative Feedback Loop + Version Tracking + Machine-Readable Output v1.2.0: Pattern-aware Feedback with Self-Synthesis Gate

1. Purpose

Create structured YAML specification documents through an interactive feedback loop:

1. Initialize draft YAML with user's topic and initial structure
2. Iterate through user feedback with immediate reflection
3. Track all changes via version numbers and changelog
4. Visualize sections on request for validation
5. Finalize to Machine-Readable English YAML format

Core Pattern (What We Used)

┌─────────────────────────────────────────────────────────────────────────┐
│  Iterative Feedback Loop Based Document Completion Process              │
│                                                                         │
│  User Feedback                                                          │
│       │                                                                 │
│       ▼                                                                 │
│  ┌─────────────────┐     ┌─────────────────┐     ┌─────────────────┐   │
│  │  v0.1.0         │ ──▶ │  v0.2.0         │ ──▶ │  v0.3.0         │   │
│  │  Initial Draft  │     │  + Feedback #1  │     │  + Feedback #2  │   │
│  └─────────────────┘     └─────────────────┘     └─────────────────┘   │
│           │                                               │             │
│           └───────────────────────────────────────────────┘             │
│                                   │                                     │
│                                   ▼                                     │
│                     ┌─────────────────────────────┐                     │
│                     │  v1.0.0 (plan-final.yaml)   │                     │
│                     │  Machine-Readable + English │                     │
│                     └─────────────────────────────┘                     │
└─────────────────────────────────────────────────────────────────────────┘

Key Characteristics

2. Invocation

User Syntax

# Start new draft with topic
/plan-draft enhanced-feedback-loop-pattern

# Resume existing draft
/plan-draft --resume enhanced-feedback-loop-pattern-20260127

# Finalize to Machine-Readable format
/plan-draft --finalize enhanced-feedback-loop-pattern-20260127

# Visualize specific section
/plan-draft --visualize phase_4 enhanced-feedback-loop-pattern-20260127

Arguments

3. Draft Document Schema

3.1 Standard Draft Template

---
# =============================================================================
# {Topic} - Plan Draft
# =============================================================================
# Purpose: {Description}
# Status: DRAFT (awaiting user approval)
# Version: 0.1.0
# Created: {ISO 8601 timestamp}
# Last Updated: {ISO 8601 timestamp}
# =============================================================================

metadata:
  document_type: plan-draft
  topic: "{topic}"
  workload_id: "{topic}_{YYYYMMDD}_{HHMMSS}"
  slug: "{topic}-{YYYYMMDD}"
  author: main-orchestrator
  approval_status: draft
  revision: 1

# =============================================================================
# 1. CORE CONCEPTS
# =============================================================================
core_concepts:
  principles:
    - id: P1
      name: "{principle_name}"
      description: |
        {Principle description}
      implications:
        - "{implication_1}"
        - "{implication_2}"

# =============================================================================
# 2. EXECUTION FLOW
# =============================================================================
execution_flow:
  phases:
    - id: phase_1
      name: "{phase_name}"
      description: "{phase_description}"
      inputs: []
      steps: []
      outputs: []

# =============================================================================
# 3. CONFIGURATION
# =============================================================================
configuration:
  # Key configuration constants

# =============================================================================
# 4. DECISIONS
# =============================================================================
decisions:
  - id: Q1
    question: "{open_question}"
    options:
      - "{option_1}"
      - "{option_2}"
    status: open  # open | decided
    decided_value: null

# =============================================================================
# 5. CHANGELOG
# =============================================================================
changelog:
  - version: "0.1.0"
    date: "{date}"
    changes:
      - "Initial draft created"

# =============================================================================
# END OF DOCUMENT
# =============================================================================

3.2 Version Numbering Convention

4. Execution Protocol

4.1 Phase 1: Initialization

async function initializeDraft(topic) {
  // 1. Generate workload identifiers
  const timestamp = new Date()
  const dateStr = timestamp.toISOString().slice(0, 10).replace(/-/g, '')
  const timeStr = timestamp.toISOString().slice(11, 19).replace(/:/g, '')

  const workloadId = `${topic}_${dateStr}_${timeStr}`
  const slug = `${topic}-${dateStr}`

  // 2. Create workload directory
  const workloadDir = `.agent/prompts/${slug}`
  await Bash({ command: `mkdir -p ${workloadDir}` })

  // 3. Generate initial draft from template
  const draftContent = generateDraftTemplate(topic, workloadId, slug, timestamp)

  // 4. Write draft file
  const draftPath = `${workloadDir}/plan-draft.yaml`
  await Write({ file_path: draftPath, content: draftContent })

  // 5. Return session context
  return {
    workloadId,
    slug,
    draftPath,
    version: "0.1.0"
  }
}

4.2 Phase 2: Feedback Loop

4.2.0 Self-Validation Protocol (v1.1.0)

Before presenting changes to user, agent performs internal validation:

async function selfValidate(context, maxIterations = 3) {
  let iterations = 0;
  let issues = [];

  while (iterations < maxIterations) {
    iterations++;
    issues = await runValidationChecks(context);

    const autoFixable = issues.filter(i => i.autoFixable);
    const nonFixable = issues.filter(i => !i.autoFixable);

    if (autoFixable.length > 0) {
      await applyAutoFixes(context, autoFixable);
      continue; // Re-validate after fixes
    }

    if (nonFixable.length === 0) {
      return { status: 'PASSED', iterations };
    }

    break;
  }

  return {
    status: issues.length > 0 ? 'PASSED_WITH_WARNINGS' : 'PASSED',
    iterations,
    warnings: issues.map(i => i.description)
  };
}

Validation Criteria:

4.2.1 Feedback Classification (v1.2.0)

Classify validation issues into AUTO_FIX vs USER_CONFIRM:

function classifyIssue(issue) {
  // AUTO_FIX: Low severity OR (Medium + auto-fixable)
  if (issue.severity === 'LOW') return 'AUTO_FIX';
  if (issue.severity === 'MEDIUM' && issue.autoFixable) return 'AUTO_FIX';

  // USER_CONFIRM: Medium non-fixable OR High/Critical
  return 'USER_CONFIRM';
}

Classification Matrix:

4.2.2 Feedback Processing

async function processFeedback(context, feedback) {
  // 1. Read current draft
  const currentDraft = await Read({ file_path: context.draftPath })

  // 2. Parse version
  const currentVersion = extractVersion(currentDraft)
  const newVersion = incrementMinorVersion(currentVersion)

  // 3. Apply feedback
  const updatedDraft = await applyFeedback(currentDraft, feedback)

  // 4. Update version and timestamp
  updatedDraft.metadata.revision++
  updatedDraft.last_updated = new Date().toISOString()

  // 5. Add changelog entry
  updatedDraft.changelog.unshift({
    version: newVersion,
    date: new Date().toISOString().slice(0, 10),
    changes: [summarizeFeedback(feedback)]
  })

  // 6. Write updated draft
  await Write({ file_path: context.draftPath, content: updatedDraft })

  // 7. Return updated context
  return {
    ...context,
    version: newVersion
  }
}

4.3 Phase 2.5: Self-Synthesis Gate (v1.2.0)

Holistic document completeness check before finalization:

async function selfSynthesisGate(context) {
  // Step 1-3: Collect stats
  const sectionStats = analyzeSections(context.draft);
  const decisionStats = analyzeDecisions(context.draft);
  const validationStats = runValidationSuite(context.draft);

  // Step 4: Calculate completion rate
  const completion = {
    sections: sectionStats.completed / sectionStats.total * 100,
    decisions: decisionStats.decided / decisionStats.total * 100,
    validations: validationStats.passed / validationStats.total * 100,
    overall: Math.min(
      sectionStats.percentage,
      decisionStats.percentage,
      validationStats.percentage
    )
  };

  // Step 5-6: Generate smart prompt and determine action
  if (completion.overall >= 100) {
    return { action: 'FINALIZE', prompt: readyForFinalization() };
  } else if (completion.overall >= 80) {
    return { action: 'CHOOSE', prompt: nearCompleteWithWarnings(completion) };
  } else {
    return { action: 'CONTINUE_EDITING', prompt: gapsDetected(completion) };
  }
}

Completion Rate Formula:

overall = min(sections%, decisions%, validations%)

Smart Prompt Templates:

Trigger Conditions:

• Explicit approval signals: lgtm, looks good, ok, approve, done
• Phase 2 Step 0 validation passes
• Version milestone (major version increment)

4.4 Phase 3: Visualization

async function visualizeSection(context, sectionId) {
  // 1. Read current draft
  const draft = await Read({ file_path: context.draftPath })

  // 2. Extract target section
  const section = extractSection(draft, sectionId)

  // 3. Generate ASCII visualization
  const visualization = generateVisualization(section)

  // 4. Output visualization (do not store in file)
  console.log(visualization)

  return visualization
}

function generateVisualization(section) {
  // Generate ASCII box diagrams for flows
  // Generate tables for configurations
  // Generate tree structures for hierarchies

  // Example flow visualization:
  /*
  ┌─────────────────────────────────────────────────┐
  │  Phase 4: Main Agent Orchestrated Feedback Loop │
  │                                                 │
  │  Step 1: Strategic Classification               │
  │       │                                         │
  │       ├── Direct Fix (LOW)                      │
  │       └── Agent Delegation (MEDIUM+)            │
  │                                                 │
  │  Step 5: Result Review & Synthesis              │
  └─────────────────────────────────────────────────┘
  */
}

4.5 Phase 4: Finalization

async function finalizeDraft(context) {
  // 1. Read current draft
  const draft = await Read({ file_path: context.draftPath })

  // 2. Validate completeness
  const validation = validateDraft(draft)
  if (!validation.complete) {
    throw new Error(`Draft incomplete: ${validation.missing.join(', ')}`)
  }

  // 3. Check all decisions are made
  const openDecisions = draft.decisions.filter(d => d.status === 'open')
  if (openDecisions.length > 0) {
    throw new Error(`Open decisions: ${openDecisions.map(d => d.id).join(', ')}`)
  }

  // 4. Convert to Machine-Readable English format
  const finalYaml = await convertToFinal(draft)

  // 5. Update version to 1.0.0
  finalYaml.metadata.version = "1.0.0"
  finalYaml.metadata.status = "APPROVED"
  finalYaml.metadata.finalized_at = new Date().toISOString()
  finalYaml.metadata.source_document = "plan-draft.yaml"
  finalYaml.metadata.source_version = context.version

  // 6. Write final file
  const finalPath = context.draftPath.replace('plan-draft.yaml', 'plan-final.yaml')
  await Write({ file_path: finalPath, content: finalYaml })

  // 7. Return result
  return {
    draftPath: context.draftPath,
    finalPath: finalPath,
    version: "1.0.0"
  }
}

5. Machine-Readable Conversion Rules

5.1 Korean → English

5.2 Comment → Structured Field

# Before (Korean comment-based)
# ---------------------------------------------------------------------------
# 2.5 Phase 4: 선택적 Feedback Loop
# ---------------------------------------------------------------------------
phase_4_selective_feedback:
  name: "Selective Feedback Loop"
  # Main Agent가 직접 Orchestrating하여...

# After (English field-based)
phase_4:
  id: phase_4_selective_feedback
  name: main_agent_orchestrated_feedback_loop
  title: "Selective Feedback Loop (Main Agent Orchestrated)"
  objective:
    why: >
      When issues are found in Phase 3.5, Main Agent must strategically orchestrate
      rather than simply handing off to Agents.

5.3 Type Definitions Extraction

During finalization, extract implicit types into explicit type_definitions:

type_definitions:
  L1:
    description: "Level 1 Summary"
    token_limit: 500
  Issue:
    fields:
      - name: id
        type: string
      - name: severity
        type: "enum[LOW, MEDIUM, HIGH, CRITICAL]"

6. Interaction Patterns

6.1 Main Agent Role

6.2 Feedback Response Protocol

1. Receive feedback → Parse intent
2. Apply immediately → Edit file directly
3. Update version → Increment + changelog
4. Confirm → Show brief summary of changes
5. Wait → User validates or provides next feedback

6.3 Visualization Request Protocol

When user requests visualization:

1. Extract target section from draft
2. Generate ASCII diagram
3. Output to console (NOT to file)
4. Wait for feedback

7. Output Paths

8. Example Session

User: /plan-draft enhanced-feedback-loop-pattern

Agent: 📝 Draft initialized
       Path: .agent/prompts/enhanced-feedback-loop-pattern-20260127/plan-draft.yaml
       Version: 0.1.0

       Draft includes:
       - Core concepts (5 principles)
       - Execution flow (6 phases)
       - 3 open decisions (Q1, Q2, Q3)

       Ready for feedback.

User: Q1 결정: 동적 (2-5개)

Agent: ✅ v0.2.0 - Q1 decided: dynamic (2-5)
       Changelog updated.

User: Phase 3을 두 Agent로 분리해

Agent: ✅ v0.3.0 - Phase 3 split into Phase 3-A (horizontal) and Phase 3-B (vertical)
       Changelog updated.

User: Phase 4 시각화해줘

Agent: [ASCII visualization of Phase 4]

       Ready for feedback.

User: --finalize

Agent: ✅ Finalized to v1.0.0
       Draft: .agent/prompts/enhanced-feedback-loop-pattern-20260127/plan-draft.yaml (v0.8.0)
       Final: .agent/prompts/enhanced-feedback-loop-pattern-20260127/plan-final.yaml (v1.0.0)

       Conversion:
       - Korean → English
       - Comments → Structured fields
       - Type definitions extracted

9. Validation Checklist

Before finalization, validate:

• All decisions (decisions[].status) are decided
• All phases have inputs, steps, outputs
• Changelog has all version entries
• No TODO or placeholder text remains
• All Korean descriptions have clear meaning for translation

10. Error Handling

11. Semantic Integrity Preservation

Critical Rule

"후속 작업을 위한 SEMANTIC INTEGRITY"

When converting draft → final:

1. No meaning loss - Every Korean concept must map to English equivalent
2. Structure preservation - Hierarchies and relationships maintained
3. Type safety - Implicit types made explicit in type_definitions
4. Reference integrity - All $ref and cross-references valid
5. Downstream compatibility - Final YAML parseable by implementation phase

12. Context Management (v1.1.0)

Strategy

"L1 summary to console, L2/L3 to file"

Manage context size by controlling what gets injected to Main Context vs what remains as file-only output.

Token Budget

Injection Rules

Console Output Rules

1. Keep under 200 tokens
2. Include version and status
3. Provide next action options

Example Output

✅ v0.3.0 - Added error handling section

Self-validation: PASSED (1 iteration, 0 auto-fixes)

Draft Status: 75% complete
Open Decisions: 0/3

Options:
1. Provide feedback to continue iterating
2. --visualize <section> to review specific area
3. --finalize (when ready for Machine-Readable conversion)

13. Pattern Comparison (v1.2.0)

Existing Pattern (v1.1.0)

/plan-draft → init → user_feedback → agent_edit → self_validate → loop → finalize

Characteristics:

• user_initiated_feedback_loop
• single_agent_sequential_execution
• self_validation_with_auto_fix
• version_tracking_per_feedback

Enhanced Pattern (v1.2.0)

/plan-draft → init → feedback_classification → selective_refinement
    → self_validation_extended → feedback_auto_fix_or_user_confirm
    → self_synthesis_gate → holistic_review → finalize_with_migration

New Characteristics:

• user_initiated_with_agent_pre_classification
• selective_area_refinement_instead_of_full_rewrite
• auto_fix_vs_user_confirm_classification
• explicit_self_synthesis_gate
• pattern_compliance_awareness
• migration_tracking_enabled

14. Migration Path (v1.2.0)

From v1.1.0 to v1.2.0

Backward Compatibility

• v1.1.0 drafts: fully supported
• Existing workflows: no breaking changes
• V1-V5 validation criteria: still apply

New Type Definitions

End of Skill Documentation