Manager → TaskList(taskId)
         → Task(create-skill, taskId)
         → Task(audit-skill, taskId)
           → Writes auditResult to TaskList(taskId, metadata)
         → TaskGet(taskId) checks metadata.auditResult
           → If FAIL: TaskUpdate(taskId, {metadata: {attempt: N+1}}) → Retry create
           → If PASS: Complete

Skill(skill-manager)

Provide:
- skill_name: your-skill-name
- description: "What it does. Use when conditions. Not for exclusions."
- (optional) skills_to_load: helper-skill-1, helper-skill-2

You → Skill(skill-manager)
      → TaskList: Create task with {skillName, attempt: 1, maxAttempts: 3}
      → Task(create-skill-files, taskId)
        → Creates skill files
        → TaskUpdate(taskId, {metadata: {draftPath: "/path/to/SKILL.md"}})
      → Task(skill-auditor, taskId)
        → TaskGet(taskId) reads metadata.draftPath
        → Reads draft, validates
        → TaskUpdate(taskId, {metadata: {auditResult: {status: "PASS"|"FAIL", issues: [...]}}})
      → TaskGet(taskId) checks metadata.auditResult
        → If FAIL & attempt < maxAttempts:
          → TaskUpdate(taskId, {metadata: {attempt: N+1, previousIssues: [...]}})
          → Retry Task(create-skill-files, taskId)
        → If PASS: Return success

You see only: "Skill 'your-skill-name' created in N attempts. All quality gates passed."

Create a skill named 'data-exporter' that exports data to CSV.
Use when users need to export data, generate reports, or create CSV files.
Not for importing data or JSON exports.

❌ Without fork (auditor sees history):
"OK I've been trying for 5 messages to get this right,
the user is getting frustrated, I should just pass it"

✅ With fork (auditor sees ONLY draft):
"I don't know how many attempts this took.
I only see the draft. It has issues.
FAIL."

skills:
  - quality-standards    # For validation
  - skill-development    # For template knowledge

.claude/skills/{skill_name}/
└── SKILL.md    # Complete skill with frontmatter, navigation, content