Analyze and improve prompts using evidence-based research techniques.
Workflow
Step 1: Collect Prompt
Check the prompt provided by the user.
- If provided directly as text, use it as is
- If a file path is provided, read and use that file
- If not provided, ask the user
Step 2: Diagnostic Analysis
Analyze the prompt using the 7 checkpoints below. Evaluate each item as Applied/Not Applied/Not Applicable.
| # | Checkpoint | Key Question | Effect |
|---|---|---|---|
| 1 | Explicitness | Are the purpose/target/output format specific? | The single most important principle |
| 2 | Few-shot Examples | Are there input-output examples? (3 is optimal) | +50% performance improvement |
| 3 | Chain-of-Thought | Does it present thought steps for complex tasks? | +40%p (math-based) |
| 4 | XML Structuring | Are sections separated with semantic tags? | Meaningful improvement |
| 5 | Context Engineering | Is the necessary context provided directly? Is there unnecessary information? | Context > Persona |
| 6 | Constraints | Is the output scope clearly limited? | Consistent improvement |
| 7 | Hallucination Prevention | Is there an escape hatch for uncertain cases? | Accuracy improvement |
Evaluation Criteria Reference: Not all checkpoints are needed for all prompts.
- Simple tasks: 1, 4, 6 are key
- Complex reasoning: 1, 2, 3 are key
- Information extraction: 1, 5, 7 are key
Step 3: Apply Improvements
From the diagnostic results, improve the Not Applied items in order of greatest effect.
Technique Application Priority (in order of evidence-based effect):
-
Explicit and specific instructions — concretely transform ambiguous parts
- Specify purpose, target, workflow, output format
- Convert "Don't" → "Do"
- Provide reasons/motivations together (the model generalizes to similar cases)
-
Add Few-shot examples — 3 examples is optimal
<example>Structure with tags- 3 principles: relevance + diversity + clarity
- Include edge cases
-
Chain-of-Thought — apply only to complex tasks
- Basic: "Think step-by-step"
- Guide: present thought steps
- Structured:
<thinking>+<answer>Separation (most powerful)
-
XML tag structuring — semantic tag names, maintain consistency
<instructions>,<context>,<constraints>,<output_format>- Mention tag names when referencing content
-
Context Engineering — inject context directly, remove unnecessary information
- Place long documents at the top, questions at the bottom
- Beware of accuracy decline after 4,000 tokens
- Providing context is more effective than declaring a role
-
Specify constraints — limit output scope
-
Prevent hallucinations — provide an escape hatch, look for evidence first
Step 4: Output results
Output in the format below.
## 진단 결과
| 체크포인트 | 상태 | 비고 |
|-----------|------|------|
| 명시성 | ✅/⚠️/➖ | 한 줄 설명 |
| ... | ... | ... |
## 주요 개선 사항
- [적용한 기법과 이유를 bullet으로 나열]
## 개선된 프롬프트
[개선된 전체 프롬프트]
## 변경 요약
[Before/After 핵심 차이를 간결하게]
Precautions
Techniques with reduced effectiveness in 2026 — not used
| Technique | Reason |
|---|---|
| Basic role prompting ("You are an expert...") | Frontier models' basic capabilities are already sufficient |
| Manual CoT ("think step by step") | Reasoning models already perform this internally |
| Anti-laziness prompting | Causes excessive triggering in the latest models |
| Sophisticated structural tricks/templates | Models understand natural language well enough |
Criteria for applying Persona
Results of an empirical study of 162 personas:
- Creative/open-ended tasks (brainstorming, writing): effective (up to 37% improvement)
- Fact-based/accuracy tasks (classification, math): negligible effect
- Principle: "Say as this, with this material, in this format" takes priority over "speak as someone"
Prevent over-improvement
- Do not touch parts that are already well written
- Do not force unnecessary techniques onto simple prompts
- Maintain the original intent and tone