Ralph Skill Review Loop

Overview

A meta-skill that uses the Ralph Wiggum technique to review and improve the Ralph Wiggum prompt generator skills themselves. Runs a continuous improvement loop until skills pass review twice consecutively with no recommendations.

Quick Start

Copy and run this prompt in a Ralph loop:

/ralph-wiggum:ralph-loop "[paste prompt below]" --completion-promise "RALPH_SKILLS_PERFECTED" --max-iterations 100

---

THE SELF-IMPROVING REVIEW LOOP PROMPT

# Task: Self-Improving Review of Ralph Wiggum Skills

## Objective
Review and improve the Ralph Wiggum prompt generator skills until they pass two consecutive reviews with zero improvement recommendations.

## Target Skills
1. ralph-prompt-builder (Master orchestrator)
2. ralph-prompt-single-task (Single task generator)
3. ralph-prompt-multi-task (Multi-task generator)
4. ralph-prompt-project (Project generator)
5. ralph-prompt-research (Research generator)

Location: .claude/skills/ralph-prompt-*/SKILL.md

## Reference Materials
- RALPH-WIGGUM-TECHNIQUE-COMPREHENSIVE-RESEARCH.md (12,000+ words of best practices)
- skill-builder-package/research/ (skill building best practices)
- skill-builder-package/examples/ (production skill patterns)

---

## STATE MANAGEMENT

### Required State Files
Create these files to track progress:

**RALPH_REVIEW_STATE.json**:
```json
{
  "current_iteration": 1,
  "consecutive_clean_reviews": 0,
  "skills_reviewed": [],
  "improvements_made": [],
  "last_review_timestamp": "",
  "status": "IN_PROGRESS"
}

RALPH_REVIEW_LOG.md:

# Ralph Skills Review Log

## Iteration History
[Append each iteration's findings here]

---

STEP 1: ORIENTATION (Every Iteration)

Read current state:

cat RALPH_REVIEW_STATE.json
cat RALPH_REVIEW_LOG.md | tail -50
git log --oneline -5
ls -la .claude/skills/ralph-prompt-*/

Check: How many consecutive clean reviews do we have?

• If 2 or more: Output RALPH_SKILLS_PERFECTED
• If less than 2: Continue to Step 2

---

STEP 2: COMPREHENSIVE SKILL REVIEW

Review Framework

For EACH skill in ralph-prompt-*, evaluate against:

2.1 Ralph Technique Alignment (from research)

• Clear completion criteria defined
• Includes self-verification commands
• Has TDD/iteration approach
• Includes "If Stuck" guidance
• Uses completion tags correctly
• Recommends appropriate max-iterations
• Follows "deterministically bad" philosophy (failures are fixable)

2.2 Skill Structure Quality

• YAML frontmatter complete (name, description with triggers)
• Progressive disclosure (overview → details → examples)
• Quick Start section exists and is actionable
• Examples are realistic and complete
• Best practices section included
• Integration with Ralph loop documented

2.3 Content Completeness

• All sections properly filled (no placeholders)
• Examples match the skill type
• Verification commands are real and runnable
• Edge cases addressed
• Cross-references to related skills

2.4 Prompt Template Quality

• Templates follow research best practices
• Success criteria are measurable
• Phase structure is clear (for multi-phase)
• State tracking included
• Progress tracking pattern included

Review Process

For each skill:

1. Read the SKILL.md file completely
2. Compare against RALPH-WIGGUM-TECHNIQUE-COMPREHENSIVE-RESEARCH.md
3. Check against all 16 criteria above
4. Document findings in REVIEW_FINDINGS.md

Review Output Format

Create/update REVIEW_FINDINGS.md:

# Review Findings - Iteration [N]

## Summary
- Skills reviewed: [count]
- Total issues found: [count]
- Critical issues: [count]
- Improvements needed: [count]

## ralph-prompt-builder
### Passing
- [x] Criterion that passes

### Issues Found
- [ ] [CRITICAL/HIGH/MEDIUM/LOW] Issue description
  - Location: [section/line]
  - Current: [what exists]
  - Should be: [what it should be]
  - Fix: [specific fix]

## ralph-prompt-single-task
[... same format]

## ralph-prompt-multi-task
[... same format]

## ralph-prompt-project
[... same format]

## ralph-prompt-research
[... same format]

## Recommendations Summary
### Must Fix (Critical/High)
1. [Recommendation 1]
2. [Recommendation 2]

### Should Fix (Medium)
1. [Recommendation 3]

### Nice to Have (Low)
1. [Recommendation 4]

## Review Result
- [ ] CLEAN (zero recommendations)
- [ ] NEEDS_WORK (has recommendations)

---

STEP 3: IMPLEMENT IMPROVEMENTS

If REVIEW_FINDINGS.md shows NEEDS_WORK:

3.1 Prioritize Fixes

Work in this order:

1. Critical issues (breaks functionality)
2. High issues (significantly impacts quality)
3. Medium issues (improves quality)
4. Low issues (polish)

3.2 Implement Each Fix

For each recommendation:

1. Read the target skill file
2. Implement the specific fix
3. Verify the fix addresses the issue
4. Commit the change:

git add

...