ralph-skill-review-loop（拉尔夫技能复习循环）

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**:

{ "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 .claude/skills/ralph-prompt-[name]/SKILL.md git commit -m "Improve ralph-prompt-[name]: [brief description]

• [Change 1]
• [Change 2]

Part of Ralph skills self-improvement loop iteration [N]"

### 3.3 Track Improvements

Update RALPH_REVIEW_STATE.json:

{ "improvements_made": [ { "iteration": N, "skill": "ralph-prompt-X", "issue": "description", "fix": "what was done" } ] }

---

## STEP 4: POST-IMPROVEMENT VERIFICATION

After implementing fixes:

### 4.1 Verify Each Skill Still Works

For each modified skill:

- YAML frontmatter is valid
- All sections render correctly
- Examples are syntactically correct
- No broken references

### 4.2 Check for Regressions

- No content accidentally deleted
- Cross-references still valid
- Templates still complete

### 4.3 Run Syntax Check

Verify YAML frontmatter

for f in .claude/skills/ralph-prompt-*/SKILL.md; do head -20 "$f" | grep -E "^(name:|description:)" done

---

## STEP 5: UPDATE STATE

Update RALPH_REVIEW_STATE.json:

If review was CLEAN (zero recommendations):

{ "consecutive_clean_reviews": [previous + 1], "last_review_result": "CLEAN", "last_review_timestamp": "[timestamp]" }

If review was NEEDS_WORK:

{ "consecutive_clean_reviews": 0, "last_review_result": "NEEDS_WORK", "improvements_this_iteration": [count], "last_review_timestamp": "[timestamp]" }

Update RALPH_REVIEW_LOG.md:

Iteration [N] - [timestamp]

Review Result

[CLEAN/NEEDS_WORK]

Issues Found

• [Issue 1]
• [Issue 2]

Fixes Applied

• [Fix 1]
• [Fix 2]

State After

• Consecutive clean reviews: [N]
• Total improvements to date: [N]

---

## STEP 6: LOOP DECISION

### Check Termination Condition

Read RALPH_REVIEW_STATE.json:

cat RALPH_REVIEW_STATE.json | jq '.consecutive_clean_reviews'

### If consecutive_clean_reviews >= 2:

Skills have passed two consecutive reviews with zero recommendations.

Create RALPH_SKILLS_VALIDATION_COMPLETE.md:

Ralph Skills Validation Complete

Summary

• Total iterations: [N]
• Total improvements made: [count]
• Final state: All skills validated

Skills Validated

1. ralph-prompt-builder - PASSED
2. ralph-prompt-single-task - PASSED
3. ralph-prompt-multi-task - PASSED
4. ralph-prompt-project - PASSED
5. ralph-prompt-research - PASSED

Validation Criteria Met

All 16 review criteria passing for all 5 skills.

Timestamp

[ISO timestamp]

Output: RALPH_SKILLS_PERFECTED

### If consecutive_clean_reviews < 2:

Continue to next iteration (loop back to STEP 1)

---

## REVIEW CRITERIA REFERENCE (Quick Check)

### Ralph Technique Alignment

1. Clear completion criteria
2. Self-verification commands
3. TDD/iteration approach
4. "If Stuck" guidance
5. tags used correctly
6. Appropriate max-iterations recommendations
7. Deterministically bad philosophy

### Skill Structure Quality

1. Complete YAML frontmatter
2. Progressive disclosure
3. Actionable Quick Start
4. Realistic examples
5. Best practices section
6. Ralph loop integration docs

### Content Completeness

1. No placeholders
2. Matching examples
3. Real verification commands

---

## ESCAPE HATCH

If stuck after 50 iterations without reaching 2 consecutive clean reviews:

1. Document the recurring issues in RALPH_REVIEW_BLOCKERS.md
2. List which criteria keep failing
3. Identify if criteria are too strict
4. Output: RALPH_REVIEW_BLOCKED

---

## PROGRESS TRACKING

Every 5 iterations, summarize:

PROGRESS SUMMARY - Iteration [N] ================================ Started: [timestamp] Current: [timestamp] Consecutive clean reviews: [N]/2

Skills Status:

• ralph-prompt-builder: [X/16 criteria passing]
• ralph-prompt-single-task: [X/16 criteria passing]
• ralph-prompt-multi-task: [X/16 criteria passing]
• ralph-prompt-project: [X/16 criteria passing]
• ralph-prompt-research: [X/16 criteria passing]

Improvements made: [total count] Remaining issues: [count]

---

## COMPLETION CONDITIONS

Output RALPH_SKILLS_PERFECTED ONLY when:

- All 5 skills reviewed
- All 16 criteria checked per skill
- Zero recommendations in current review
- Zero recommendations in previous review
- consecutive_clean_reviews >= 2 in state file
- RALPH_SKILLS_VALIDATION_COMPLETE.md created
- All changes committed

---

## SAFETY LIMITS

- Maximum iterations: 100
- Expected completion: 20-40 iterations
- Budget alert: If > 50 iterations, evaluate if criteria are too strict

Usage Instructions

1. Initialize State Files

Before running, create the initial state:

# Create state file
cat > RALPH_REVIEW_STATE.json << 'EOF'
{
  "current_iteration": 0,
  "consecutive_clean_reviews": 0,
  "skills_reviewed": [],
  "improvements_made": [],
  "last_review_timestamp": "",
  "status": "NOT_STARTED"
}
EOF

# Create log file
cat > RALPH_REVIEW_LOG.md << 'EOF'
# Ralph Skills Review Log

## Overview
Self-improving review loop for Ralph Wiggum prompt generator skills.

## Target: Two consecutive clean reviews

---

## Iteration History

EOF

2. Run the Loop

/ralph-wiggum:ralph-loop "[THE PROMPT ABOVE]" \
  --completion-promise "RALPH_SKILLS_PERFECTED" \
  --max-iterations 100

3. Monitor Progress

# Check current state
cat RALPH_REVIEW_STATE.json | jq '.'

# See recent activity
tail -30 RALPH_REVIEW_LOG.md

# Check how many clean reviews
cat RALPH_REVIEW_STATE.json | jq '.consecutive_clean_reviews'

4. After Completion

Review the outputs:

• RALPH_REVIEW_STATE.json - Final state
• RALPH_REVIEW_LOG.md - Complete history
• REVIEW_FINDINGS.md - Last review details
• RALPH_SKILLS_VALIDATION_COMPLETE.md - Success certificate
• Git log - All improvements committed

Why This Works

1. State Tracking: JSON state file persists across iterations
2. Clear Criteria: 16 specific, measurable review criteria
3. Self-Correction: Each iteration reads previous results and fixes issues
4. Termination Condition: Two consecutive clean reviews ensures stability
5. Evidence-Based: All findings documented, all fixes tracked
6. Git Integration: Every improvement committed for auditability

Expected Behavior

Iteration 1-5: Discovery phase

• Identify initial issues across all skills
• Begin fixing critical issues

Iteration 6-15: Improvement phase

• Systematic fixes
• Quality improvements
• Cross-consistency

Iteration 16-25: Stabilization phase

• Fewer issues found
• Polish and edge cases
• Approaching clean reviews

Iteration 26-40: Validation phase

• First clean review achieved
• Verify no regressions
• Second clean review achieved
• Completion

Customization

Stricter Review

Add more criteria to the review framework.

Faster Completion

Reduce to "one clean review" by changing:

consecutive_clean_reviews >= 1

Focus on Specific Skills

Modify the target skills list in the prompt.