Ralph - Boucle auto-référentielle avec vérification

VérifiéSûr

Système de persistence itérative pour tâches complexes avec mode PRD et exécution parallèle Ultrawork. Inclut routage intelligent de modèles et vérification d'achèvement.

Spar Skills Guide Bot
DeveloppementAvancé
2002/06/2026
Claude Code
#self-referential-loop#task-completion#prd#ultrawork

Recommandé pour

Notre avis

Boucle auto-référentielle qui itère sur une tâche jusqu'à sa complétion, avec vérification par un architecte, intégrant la génération de PRD et l'exécution parallèle.

Points forts

  • Assure une complétion systématique des tâches grâce à des boucles de vérification
  • Génère automatiquement un Product Requirements Document (PRD) pour structurer le travail
  • Exécute des sous-tâches en parallèle via Ultrawork pour maximiser l'efficacité
  • Route intelligemment les tâches vers des modèles adaptés pour économiser des tokens

Limites

  • Peut générer une surcharge pour des tâches simples qui ne nécessitent pas d'itérations
  • Dépend de la qualité de la vérification par l'architecte et de la configuration des agents
  • La configuration PRD est optionnelle et peut allonger le temps de démarrage
Quand l'utiliser

Lorsque vous devez accomplir une tâche complexe en plusieurs étapes, avec des vérifications itératives et une exécution parallèle.

Quand l'éviter

Pour des tâches triviales ou uniques où une exécution directe est plus rapide qu'une boucle de vérification.

Analyse de sécurité

Sûr
Score qualité75/100

The skill is a meta-instruction file for an AI coding agent. It does not contain any destructive or exfiltrating commands, obfuscated payloads, or instructions to bypass safety. It defines a workflow for persistent task completion with verification tiers and agent delegation, relying on existing, presumably safe, agent tools.

Aucun point d'attention détecté

Exemples

PRD-driven development
--prd Build a todo app with React and TypeScript including user authentication and a database.
Simple iterative task
Refactor the authentication module to use JWT tokens instead of session cookies. Ensure all existing tests pass and type checking succeeds.
Parallel implementation
Implement the following features in parallel: user profile page, admin dashboard, and notification system. Use Ultrawork's parallel execution.

name: ralph description: Self-referential loop until task completion with architect verification

Ralph Skill

[RALPH + ULTRAWORK - ITERATION {{ITERATION}}/{{MAX}}]

Your previous attempt did not output the completion promise. Continue working on the task.

PRD MODE (OPTIONAL)

If the user provides the --prd flag, initialize a PRD (Product Requirements Document) BEFORE starting the ralph loop.

Detecting PRD Mode

Check if {{PROMPT}} contains the flag pattern: --prd or --PRD

PRD Initialization Workflow

When --prd flag detected:

  1. Create PRD File Structure (.omc/prd.json and .omc/progress.txt)
  2. Parse the task (everything after --prd flag)
  3. Break down into user stories with this structure:
{
  "project": "[Project Name]",
  "branchName": "ralph/[feature-name]",
  "description": "[Feature description]",
  "userStories": [
    {
      "id": "US-001",
      "title": "[Short title]",
      "description": "As a [user], I want to [action] so that [benefit].",
      "acceptanceCriteria": ["Criterion 1", "Typecheck passes"],
      "priority": 1,
      "passes": false
    }
  ]
}
  1. Create progress.txt:
# Ralph Progress Log
Started: [ISO timestamp]

## Codebase Patterns
(No patterns discovered yet)

---
  1. Guidelines for PRD creation:

    • Right-sized stories: Each completable in one focused session
    • Verifiable criteria: Include "Typecheck passes", "Tests pass"
    • Independent stories: Minimize dependencies
    • Priority order: Foundational work (DB, types) before UI
  2. After PRD created: Proceed to normal ralph loop execution using the user stories as your task list

Example Usage

User input: --prd build a todo app with React and TypeScript

Your workflow:

  1. Detect --prd flag
  2. Extract task: "build a todo app with React and TypeScript"
  3. Create .omc/prd.json with user stories
  4. Create .omc/progress.txt
  5. Begin ralph loop using user stories as task breakdown

ULTRAWORK MODE (AUTO-ACTIVATED)

Ralph is a persistence wrapper that includes Ultrawork as a component for maximum parallel execution. You MUST follow these rules:

Parallel Execution Rules

  • PARALLEL: Fire independent calls simultaneously - NEVER wait sequentially
  • BACKGROUND FIRST: Use Task(run_in_background=true) for long operations (10+ concurrent)
  • DELEGATE: Route tasks to specialist agents immediately

Smart Model Routing (SAVE TOKENS)

| Task Complexity | Tier | Examples | |-----------------|------|----------| | Simple lookups | LOW (haiku) | "What does this function return?", "Find where X is defined" | | Standard work | MEDIUM (sonnet) | "Add error handling", "Implement this feature" | | Complex analysis | HIGH (opus) | "Debug this race condition", "Refactor auth module" |

Available Agents by Tier

FIRST ACTION: Before delegating any work, read the agent reference file:

Read file: docs/shared/agent-tiers.md

This provides the complete agent tier matrix, MCP tool assignments, and selection guidance.

CRITICAL: Always pass model parameter explicitly!

Task(subagent_type="oh-my-gemini:architect-low", model="gemini-flash-latest", prompt="...")
Task(subagent_type="oh-my-gemini:executor", model="gemini-pro-latest", prompt="...")
Task(subagent_type="oh-my-gemini:architect", model="gemini-pro-latest", prompt="...")

Background Execution Rules

Run in Background (set run_in_background: true):

  • Package installation (npm install, pip install, cargo build, etc.)
  • Build processes (project build command, make, etc.)
  • Test suites (project test command, etc.)
  • Docker operations: docker build, docker pull

Run Blocking (foreground):

  • Quick status checks: git status, ls, pwd
  • File reads, edits
  • Simple commands

COMPLETION REQUIREMENTS

Before claiming completion, you MUST:

  1. Verify ALL requirements from the original task are met
  2. Ensure no partial implementations
  3. Check that code compiles/runs without errors
  4. Verify tests pass (if applicable)
  5. TODO LIST: Zero pending/in_progress tasks

VERIFICATION BEFORE COMPLETION (IRON LAW)

NO COMPLETION CLAIMS WITHOUT FRESH VERIFICATION EVIDENCE

Before outputting the completion promise:

Steps (MANDATORY)

  1. IDENTIFY: What command proves the task is complete?
  2. RUN: Execute verification (test, build, lint)
  3. READ: Check output - did it actually pass?
  4. ONLY THEN: Proceed to Architect verification

Red Flags (STOP and verify)

  • Using "should", "probably", "seems to"
  • About to output completion without fresh evidence
  • Expressing satisfaction before verification

Evidence Chain

  1. Fresh test run output showing pass
  2. Fresh build output showing success
  3. lsp_diagnostics showing 0 errors
  4. THEN Architect verification
  5. THEN completion promise

Skipping verification = Task NOT complete

VERIFICATION PROTOCOL (TIERED)

Ralph uses tiered verification to save tokens while maintaining quality.

Verification Tier Selection

Before spawning architect for verification, determine the appropriate tier:

| Change Profile | Tier | Agent | |----------------|------|-------| | <5 files, <100 lines, full tests | LIGHT | architect-low (haiku) | | Standard changes | STANDARD | architect-medium (sonnet) | | >20 files, security/architectural | THOROUGH | architect (opus) |

Ralph Minimum Verification Tier

Floor: STANDARD (architect-medium / sonnet)

Even for small changes (<5 files), ralph requires at least STANDARD tier verification. The LIGHT tier (haiku) is insufficient for ralph's completion guarantee. When tier selection returns LIGHT, upgrade to STANDARD.

Verification Flow

  1. Collect change metadata: Count files, lines, detect security/architectural patterns
  2. Select tier: Apply rules from /docs/shared/verification-tiers.md
  3. Spawn appropriate architect:
    // LIGHT - small, well-tested changes
    Task(subagent_type="oh-my-gemini:architect-low", model="gemini-flash-latest", prompt="Quick verification: [describe changes]")
    
    // STANDARD - most changes
    Task(subagent_type="oh-my-gemini:architect-medium", model="gemini-pro-latest", prompt="Verify implementation: [describe changes]")
    
    // THOROUGH - large/security/architectural changes
    Task(subagent_type="oh-my-gemini:architect", model="gemini-pro-latest", prompt="Full verification: [describe changes]")
    
  4. Wait for verdict
  5. If approved: Run /oh-my-gemini:cancel to cleanly exit
  6. If rejected: Fix issues and re-verify (same tier)

For complete tier selection rules, read: docs/shared/verification-tiers.md

ZERO TOLERANCE

  • NO Scope Reduction - deliver FULL implementation
  • NO Partial Completion - finish 100%
  • NO Premature Stopping - ALL TODOs must be complete
  • NO TEST DELETION - fix code, not tests

STATE CLEANUP ON COMPLETION

IMPORTANT: Use the cancel skill for proper state cleanup

When work is complete and Architect verification passes, run /oh-my-gemini:cancel to cleanly exit ralph mode. This handles:

  • Deletion of ralph state files (both local and global)
  • Cleanup of linked ultrawork or ecomode state
  • Proper termination of the ralph loop

This ensures clean state for future sessions without leaving stale state files behind.

INSTRUCTIONS

  • Review your progress so far
  • Continue from where you left off
  • Use parallel execution and background tasks
  • When FULLY complete AND Architect verified: Run /oh-my-gemini:cancel to cleanly exit and clean up all state files
  • Do not stop until the task is truly done

Original task: {{PROMPT}}

Skills similaires