메뉴
Iterative quality improvement — generate, evaluate, feedback, re-generate until threshold met
기본적으로 소스를 먼저 확인하는 Prompt가 선택됩니다. 직접 명령으로 전환하거나 로컬 사본을 다운로드할 수도 있습니다.
설치 여부를 결정하기 전에 SKILL.md와 SkillsMP에 표시된 보조 파일을 읽어 보세요.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
직접 명령은 검토 Prompt를 거치지 않습니다. 실행하기 전에 소스를 확인하세요.
npx skills add https://github.com/fabioc-aloha/Alex_Plug_In --skill ralph-loop명령은 한 줄로 유지됩니다. 복사하기 전에 가로로 스크롤해 전체 내용을 확인하세요.
로컬 사본을 원하시나요? SkillsMP에서 현재 제공할 수 있는 파일을 다운로드하세요.
SOC 직업 분류 기준
SKILL.md 표시 중
| name | ralph-loop |
| description | Iterative quality improvement — generate, evaluate, feedback, re-generate until threshold met |
| applyTo | **/*quality*,**/*eval*,**/*iterate*,**/*ralph* |
| domain | Quality Engineering |
| category | Iterative Improvement |
| tier | advanced |
| dependencies | ["testing-strategies","code-review"] |
| created | "2026-04-14T00:00:00.000Z" |
| author | Alex |
| source | microsoft/skills (Sensei technique by Shayne Boyer) |
Generate → Evaluate → Feedback → Re-generate until quality threshold met.
Inspired by the Sensei iterative quality improvement patterns from GitHub Copilot for Azure:
"LLMs improve dramatically when given structured feedback about what went wrong and specific guidance on how to fix it."
— Shayne Boyer (@spboyer), Microsoft
┌─────────────────────────────────────────────────────────────┐
│ RALPH LOOP │
├─────────────────────────────────────────────────────────────┤
│ │
│ ┌─────────┐ ┌──────────┐ ┌──────────┐ │
│ │ Generate │────▶│ Evaluate │────▶│ Analyze │ │
│ └─────────┘ └──────────┘ └──────────┘ │
│ ▲ │ │
│ │ ┌──────────────────────┘ │
│ │ ▼ │
│ │ ┌──────────┐ ┌───────────┐ │
│ └────│ Feedback │◀────│ Threshold │ │
│ └──────────┘ │ Met? │ │
│ └───────────┘ │
│ │ Yes │
│ ▼ │
│ ┌────────┐ │
│ │ Done │ │
│ └────────┘ │
└─────────────────────────────────────────────────────────────┘
| Parameter | Default | Description |
|---|---|---|
maxIterations | 5 | Maximum iterations before stopping |
qualityThreshold | 80 | Score (0-100) to consider quality met |
improvementThreshold | 5 | Minimum improvement required per iteration |
earlyStopOnPerfect | true | Stop immediately when score reaches 100 |
includeFeedback | true | Include feedback in re-generation prompts |
qualityThresholdimprovementThreshold for 2+ iterationsmaxIterationsCriteria define what "good" looks like:
{
"skillName": "cosmos-db-patterns",
"language": "typescript",
"correctPatterns": [
{
"code": "container.items.query()",
"description": "Use container.items.query() for parameterized queries",
"section": "query-patterns"
}
],
"incorrectPatterns": [
{
"code": "SELECT * FROM c",
"description": "Avoid SELECT * — specify required fields",
"section": "query-patterns"
}
],
"rules": [
{
"name": "error-handling",
"requiredPatterns": ["try {", "catch ("],
"forbiddenPatterns": ["// eslint-disable"]
}
]
}
Transforms evaluation findings into LLM-actionable feedback:
{
"score": 65,
"passed": false,
"findings": [
{
"severity": "error",
"rule": "pattern:query-patterns",
"message": "Incorrect pattern found: Avoid SELECT * — specify required fields"
}
],
"matchedIncorrect": ["query-patterns"]
}
## Issues Found in Generated Content
### CRITICAL ERRORS (Must Fix)
- **pattern:query-patterns**: Incorrect pattern found: Avoid SELECT * — specify required fields
- Suggestion: Review acceptance criteria for correct usage
### INCORRECT PATTERNS DETECTED
Found incorrect patterns in sections: **query-patterns**
Review the acceptance criteria for these sections and use correct patterns instead.
### SUGGESTED CORRECTIONS
Based on the acceptance criteria, consider:
- Use: Use container.items.query() for parameterized queries
# Evaluate content against criteria
node ralph-loop.cjs --evaluate --content "SELECT * FROM c" --criteria criteria.json
# Evaluate from file
node ralph-loop.cjs --evaluate --content-file output.ts --criteria criteria.json
# Extract criteria from SKILL.md
node ralph-loop.cjs --extract-criteria --skill .github/skills/cosmos-db/SKILL.md
# Build feedback from findings
node ralph-loop.cjs --feedback --findings '[{"severity":"error","rule":"syntax","message":"Missing semicolon"}]'
# Stdin mode (for extension integration)
echo '{"content":"...","criteria":{...}}' | node ralph-loop.cjs --stdin
import { ContentEvaluator, FeedbackBuilder, RalphLoopController } from './ralph-loop';
// Create evaluator with criteria
const criteria = loadSkillCriteria('cosmos-db-patterns');
const controller = new RalphLoopController(criteria, {
maxIterations: 3,
qualityThreshold: 85,
});
// Run loop with Copilot LLM
const result = await controller.run(
async (prompt) => await copilot.generate(prompt),
"Write a Cosmos DB query for user lookup",
"user-query-scenario"
);
console.log(`Final score: ${result.finalScore}`);
console.log(`Converged: ${result.converged}`);
console.log(`Iterations: ${result.iterations.length}`);
const evaluator = new ContentEvaluator(criteria);
const feedbackBuilder = new FeedbackBuilder();
const result = evaluator.evaluate(generatedCode, 'my-scenario');
const feedback = feedbackBuilder.buildFeedback(result, criteria);
if (!result.passed) {
console.log(feedback); // Give to LLM for improvement
}
Track these across iterations:
| Metric | Description |
|---|---|
score | Overall quality score (0-100) |
errorCount | Number of critical errors |
warningCount | Number of warnings |
matchedCorrect | Correct patterns found |
matchedIncorrect | Incorrect patterns found |
improvement | Score change from first to last iteration |
converged | Whether quality threshold was met |
| Don't | Do |
|---|---|
Set maxIterations too high (>5) | Usually 3-5 iterations sufficient |
Set qualityThreshold at 100 | 80-90 is realistic for most use cases |
| Include vague criteria | Be specific with patterns and rules |
| Skip feedback on failure | Always provide structured feedback |