prompt-variation-guide

Prompt Generation Diversity Guide

Version: 1.1.1 Created: 2026-01-08 | Updated: 2026-09-27 — Updated default notation (GPT-6 Sol · Claude Opus 5.5). Previous: 2026-07-21 model lineup update (GPT-5.6 Sol · Claude Opus 4.8/Fable 5 · Gemini 3.1) Purpose: Improving prompt generation diversity for the /auto-prompt agent

🆕 Generation Notice: The GPT XML block notation in this guide (output_verbosity_spec , etc.) follows the old 5.2/5.4 generation style — applies only when the user specifies XML style. Current GPT default = GPT-6 Sol(2026-09-22~ · previous default GPT-5.6 Sol = Markdown lean outcome-first) — the rules are skills/gpt-5.6-prompt-enhancement/references/full.md. Claude XML blocks are based on Opus 4.8·Fable 5 (current Claude default = Opus 5.5 → skills/claude-fable-5-prompt-strategies.md Part 2.7).

Goal of This Guide: Provide a method for generating prompts whose structure and expression clearly differ by model/task type/difficulty, even for the same task

See references/techniques.md for Variation Decision Matrix, Advanced Patterns, Job Adaptability Guide, /auto-prompt Integration (Part 5) See references/examples.md for complete before/after examples (3 occupations: teacher/lawyer/self-employed × 4 models × 3 difficulty levels)

Part 1: Universal Framework

1.1 4-Layer Prompt Architecture

All prompts follow the following 4-layer structure:

Layer 1: Role (역할)
   ↓
Layer 2: Context (맥락)
   ↓
Layer 3: Task (작업)
   ↓
Layer 4: Output Format (출력 형식)

Application by Difficulty:

Difficulty 1-2: Layer 1 + Layer 3 only (role + task)
Difficulty 3-5: Includes all 4 layers
Difficulty 6-7: 4 layers + model-specific XML blocks

1.2 Progressive Complexity by Difficulty

Difficulty	Format	Length	Structure	XML Block
1-2	Natural language	1-3줄	Role + Task	None
3-5	Markdown	5-15줄	4 layers + section separation	1 required
6-7	XML	20줄+	Complete XML template	2-3 required

Part 3.3: Model Personality Mapping (key selection criteria)

Model	Personality	Expression Style	Emphasis
GPT-5.6 Sol	Outcome-oriented, lean	Markdown lean outcome-first, minimize enumeration	Success Criteria + Stop Rules
Claude (Opus 4.8 · Fable 5)	Explicit, reason-oriented	Detailed, context-rich	Context + Action
Gemini 3.1	Structure-oriented, step-by-step	Hierarchical, clear order	Constraints First
Perplexity	Research-oriented, source-based	Search → Synthesize	Workflow + Sources

Part 4: Quality Validation

4.1 Differentiation Metrics

Verifying Differentiation Between Models:

✅ 동일 작업, 동일 난이도에서:
   - 구조가 다른가? (섹션 순서, 블록 종류)
   - 표현 스타일이 다른가? (간결 vs 상세)
   - 강조점이 다른가? (제약 vs 맥락 vs 워크플로우)
   - 특화 XML 블록이 포함되었는가?

Verifying Task Type Reflection:

✅ 코딩 작업 → 범위 제약 블록?
✅ 글쓰기 작업 → 문서 구조/톤 명시?
✅ 분석 작업 → 판단 기준/불확실성 처리?
✅ 데이터 처리 → JSON 스키마/필드 설명?

4.2 Common Anti-Patterns to Avoid

❌ Failure Patterns:

1. 모든 프롬프트가 동일한 섹션 순서
   → GPT/Claude/Gemini/Perplexity 모두 "역할-작업-형식" 순서

2. 모델별 차이가 문구만 바뀌고 구조 동일
   → "당신은 교사입니다" vs "당신은 변호사입니다"만 다름

3. Gemini인데 Constraints가 중간이나 끝에
   → Gemini는 **반드시** Constraints First

4. Perplexity인데 검색/출처 요청 없음
   → 워크플로우 또는 "검색해주세요" 문구 필수

5. Claude인데 맥락 설명 없음
   → Claude는 "왜 중요한지" 설명 필요

6. GPT인데 Success Criteria/Stop Rules 없음 (난이도 3+)
   → 5.6 Sol Markdown 기준 필수. 구 XML 스타일 명시 요청 시엔 output_verbosity_spec

Summary

3 Core Principles:

Differentiation by Model: Reflect the structural characteristics of GPT/Claude/Gemini/Perplexity
Based on Task Type: Classify into 4 types (coding/writing/analysis/data) regardless of occupation
Occupational Adaptability: Teacher/lawyer/self-employed examples can be converted to all occupational groups

By using this guide:

✅ Even for the same topic, the structure differs by model
✅ Automatically select the XML block matching the task type
✅ New job groups can also be handled immediately with existing patterns