Prompt Engineering

---
name: prompt-engineering
description: Expert prompt optimization system for the prompts INSIDE an AI product you are building — system prompts, LLM feature prompts, chatbot/agent instructions. Use when the user wants to write or improve a system prompt for an AI feature they're shipping, review/critique an LLM prompt, apply prompt-engineering techniques (chain-of-thought, few-shot, structured output, hard constraints) to a product prompt, or optimize cost/latency of a production prompt. Do NOT use this to clarify or structure the user's own vague request to Claude Code — that is `prompt-clarifier`'s job, not this skill's.
---

# Prompt Engineering Expert

Master system for creating, analyzing, and optimizing prompts for AI products using research-backed techniques and battle-tested production patterns.

## Core Capabilities

1. **Prompt Analysis & Improvement** - Analyze existing prompts and provide specific optimization recommendations
2. **System Prompt Creation** - Build production-ready system prompts using the 6-step framework
3. **Failure Mode Detection** - Identify and fix common prompt engineering mistakes
4. **Cost Optimization** - Balance performance with token efficiency
5. **Research-Backed Techniques** - Apply proven prompting methods from academic studies

## The 6-Step Optimization Framework

When improving any prompt, follow this systematic process:

### Step 1: Start With Hard Constraints (Lock Down Failure Modes)

Begin with what the model CANNOT do, not what it should do.

**Pattern:**
```
NEVER:
- [TOP 3 FAILURE MODES - BE SPECIFIC]
- Use meta-phrases ("I can help you", "let me assist")
- Provide information you're not certain about

ALWAYS:
- [TOP 3 SUCCESS BEHAVIORS - BE SPECIFIC]
- Acknowledge uncertainty when present
- Follow the output format exactly
```

**Why:** LLMs are more consistent at avoiding specific patterns than following general instructions. "Never say X" is more reliable than "Always be helpful."

### Step 2: Trigger Professional Training Data (Structure = Quality)

Use formatting that signals technical documentation quality:

- **For Claude**: Use XML tags (`<system_constraints>`, `<task_instructions>`)
- **For GPT-4**: Use JSON structure
- **For GPT-3.5**: Use simple markdown

**Why:** Well-structured documents trigger higher-quality training data patterns.

### Step 3: Have The LLM Self-Improve Your Prompt

Don't optimize manually - let the model do it using this meta-prompt:

```
You are a prompt optimization specialist. Your job is to improve prompts for production AI systems.

CURRENT PROMPT:
[User's prompt here]

PERFORMANCE DATA:
- Main failure modes: [List top 3 if known]
- Target use case: [Describe]

OPTIMIZATION TASK:
1. Identify the top 3 weaknesses in this prompt
2. Rewrite to fix those weaknesses using these principles:
   - Hard constraints over soft instructions
   - Specific examples over generic guidance
   - Structured format over free text
3. Predict the improvement percentage for each change

CONSTRAINTS:
- Must maintain core functionality
- Cannot exceed 150% of current token count
- Must include failure mode handling

OUTPUT:
Optimized prompt + rationale for each change
```

### Step 4: Trace Edge Cases and Analyze Failures

Test the prompt systematically:

- **20% happy path** - Standard use cases
- **60% edge cases** - Unusual inputs, malformed data, ambiguous requests
- **20% adversarial** - Attempts to break the prompt or extract system instructions

Identify the top 3 failure patterns and address them explicitly in the prompt.

### Step 5: Build Evaluation Criteria

Define clear success metrics:

- **Accuracy** - Does it get the right answer?
- **Format compliance** - Does it follow output requirements?
- **Safety** - Does it handle adversarial inputs correctly?
- **Cost efficiency** - Appropriate token usage?
- **Latency** - Response speed acceptable?

### Step 6: Hill Climb - Quality First, Cost Second

**Phase 1: Climb Up for Quality**
- Use longer, detailed prompts
- Include extensive examples
- Focus on hitting quality targets
- Ignore token costs temporarily

**Phase 2: Descend for Cost**
- Compress without losing performance
- Remove redundant examples
- Use structured output to reduce variance
- Test each compression against metrics

## Production Prompt Template

Use this battle-tested template structure:

```
<system_role>
You are [SPECIFIC ROLE], not a general AI assistant.
You [CORE FUNCTION] for [TARGET USER].
</system_role>

<hard_constraints>
NEVER:
- [FAILURE MODE 1 - SPECIFIC]
- [FAILURE MODE 2 - SPECIFIC]
- [FAILURE MODE 3 - SPECIFIC]
- Use meta-phrases ("I can help you", "let me assist")

ALWAYS:
- [SUCCESS BEHAVIOR 1 - SPECIFIC]
- [SUCCESS BEHAVIOR 2 - SPECIFIC]
- [SUCCESS BEHAVIOR 3 - SPECIFIC]
- Acknowledge uncertainty when present
</hard_constraints>

<context_info>
Current user: [USER_CONTEXT]
Available tools: [TOOL_LIST]
Key limitations: [SPECIFIC_LIMITATIONS]
</context_info>

<task_instructions>
Your job is to [CORE TASK] by:

1. [STEP 1 - SPECIFIC ACTION]
2. [STEP 2 - SPECIFIC ACTION]
3. [STEP 3 - SPECIFIC ACTION]

If [EDGE_CASE_1], then [SPECIFIC_RESPONSE].
If [EDGE_CASE_2], then [SPECIFIC_RESPONSE].
If [EDGE_CASE_3], then [SPECIFIC_RESPONSE].
</task_instructions>

<output_format>
Respond using this exact structure:

[SECTION_1]: [DESCRIPTION]
[SECTION_2]: [DESCRIPTION]

Requirements:
- [FORMAT_REQUIREMENT_1]
- [FORMAT_REQUIREMENT_2]
</output_format>

<examples>
Example 1 - Happy Path:
Input: [TYPICAL_INPUT]
Output: [IDEAL_RESPONSE]

Example 2 - Edge Case:
Input: [EDGE_CASE_INPUT]
Output: [EDGE_CASE_RESPONSE]

Example 3 - Complex:
Input: [COMPLEX_SCENARIO]
Output: [COMPLEX_RESPONSE]
</examples>
```

## Research-Backed Techniques

### Chain-of-Table (For Structured Data)

**Best for:** Financial dashboards, data analysis, table processing
**Performance:** 8.69% improvement on table tasks
**How:** Make the AI manipulate table structure step-by-step, not reason about tables in text

### Chain-of-Thought (For Math/Logic)

**Best for:** Arithmetic reasoning, logic puzzles, formal reasoning
**Limitations:** Only works on 100B+ parameter models; minimal benefit for content generation
**When NOT to use:** Classification, content generation, most business tasks

### Few-Shot Learning (Use Carefully)

**When it helps:** Task requires specific style, format examples improve output
**When it hurts:** Advanced reasoning tasks (o1, DeepSeek R1 models)
**Best practice:** Test systematically - few-shot has highest variability of any technique

### Multi-Shot Prompting (For Conversations)

**Best for:** Customer support, sales conversations, multi-turn interactions
**How:** Show entire conversation flows, not isolated examples
**Benefit:** Teaches conversation patterns, not just individual responses

## The 3 Fatal Mistakes

### Mistake #1: The "Kitchen Sink" Prompt

**Problem:** One massive prompt trying to do sentiment analysis, routing, response generation, and task management simultaneously.

**Fix:** Break into specialized prompts:
- Prompt 1: Sentiment classification
- Prompt 2: Response generation
- Prompt 3: Task routing

Each prompt does ONE thing exceptionally well.

### Mistake #2: The "Demo Magic" Trap

**Problem:** Prompt works perfectly on clean, polite, well-formatted demo data but fails on 40% of real production inputs.

**Fix:** Build eval suite from real chaos:
- 20% happy path
- 60% edge cases (broken formatting, angry users, multiple languages)
- 20% adversarial scenarios

### Mistake #3: The "Set and Forget" Fallacy

**Problem:** Shipping a prompt and never updating it as business evolves, user needs change, and new edge cases emerge.

**Fix:** Build continuous optimization:
- **Weekly reviews** - Monitor eval metrics
- **Monthly iterations** - Analyze user feedback
- **Quarterly overhauls** - Reassess approach
- **Real-time learning** - A/B test variations

## Cost Economics

Shorter, structured prompts have major advantages:

**Example comparison:**
- Detailed approach: 2,500 token prompt → $3,000/day at 100k calls
- Simpler approach: 212 token prompt → $706/day at 100k calls
- **76% cost reduction**

**Benefits of compression:**
- Less variance in outputs
- Faster latency
- Lower costs

**When to use longer prompts:** Complex tasks requiring extensive context, edge case handling, or when that 88% cost increase delivers proportional value.

## Prompt Analysis Workflow

When user provides a prompt to improve:

1. **Identify Current State**
   - What's the core function?
   - What failure modes exist?
   - Is structure optimized?

2. **Analyze Against Framework**
   - Are hard constraints defined?
   - Is formatting optimal for the model?
   - Are examples effective?
   - Are edge cases handled?

3. **Provide Specific Recommendations**
   - List top 3-5 improvements
   - Explain WHY each change matters
   - Show before/after for key sections
   - Predict performance impact

4. **Offer Complete Rewrite**
   - Apply the Production Template
   - Incorporate all recommendations
   - Add edge case handling
   - Optimize structure for target model

5. **Suggest Testing Strategy**
   - Recommend specific test cases
   - Define success metrics
   - Provide evaluation approach

## Key Principles

1. **Conciseness Matters** - Context window is shared. Only include what Claude doesn't already know.

2. **Structure = Quality** - XML for Claude, JSON for GPT-3.5, Markdown for docs. Format signals quality.

3. **Hard Constraints Over Soft** - "Never do X" is more reliable than "Be helpful."

4. **Systematic Testing** - Build evals with 20% happy path, 60% edge cases, 20% adversarial.

5. **Continuous Optimization** - Prompts decay as business evolves. Build iteration into workflow.

6. **Cost-Performance Balance** - Climb for quality first, then descend for cost optimization.

## Quick Reference: When to Use What

**Use Chain-of-Table when:**
- Processing structured data
- Working with tables
- Financial/data analysis tasks

**Use Chain-of-Thought when:**
- Math problems
- Logic puzzles
- Formal reasoning
- NOT for content generation

**Use Few-Shot when:**
- Specific style/format needed
- Examples improve understanding
- NOT with o1/R1 reasoning models

**Use Multi-Shot when:**
- Multi-turn conversations
- Customer support flows
- Sales interactions

**Use Nested Prompting when:**
- Complex multi-step workflows
- Enterprise processes
- Need specialized handling per step

## Response Pattern

When providing prompt improvements, always:

1. **Start with assessment** - "This prompt does X well, but has Y weaknesses"
2. **Provide specific fixes** - Not "add examples" but "add examples like [concrete example]"
3. **Explain the why** - Reference research findings or production patterns
4. **Show the rewrite** - Give complete improved version
5. **Suggest testing** - Recommend specific test cases