Prompt Engineering

# Playbook: Prompt Engineering

## Configuration

| Setting | Value |
|---------|-------|
| **Recommended Mode** | C (Audit) or A (Full RLE) |
| **Recommended Loops** | SIEGE → FORGE → MIRROR |
| **Recommended Depth** | STANDARD |

## Protocol

### Step 1: AUDIT with SIEGE

Attack the existing prompt across categories:

**Ambiguity attacks:**
- Vague instructions that could be interpreted multiple ways
- Missing constraints that leave too much freedom
- Implicit assumptions not made explicit

**Edge case attacks:**
- Empty or minimal input
- Extremely long input
- Input in unexpected languages
- Adversarial input (prompt injection attempts)
- Input that contradicts the instructions

**Output quality attacks:**
- Does the format remain consistent across varied inputs?
- Does the tone/style hold under pressure?
- Are there hallucination-prone areas?
- Does it handle "I don't know" gracefully?

**Structure attacks:**
- Is the prompt order optimal? (instructions, context, examples, constraints)
- Are examples representative or misleading?
- Is there instruction-following degradation with length?

### Step 2: FORGE — Rewrite

Rewrite the prompt applying the audit findings:
- Clear role/persona definition
- Explicit constraints and boundaries
- Structured output format specification
- Examples that cover normal AND edge cases
- Graduated instructions (most important first)

### Step 3: MIRROR — LLM Reaction Simulation

Simulate how the target LLM would respond:
- What would a literal interpretation produce?
- Where would the LLM take shortcuts?
- Where would it over-elaborate?
- What would trigger a refusal?
- What would cause format breaking?

### Step 4: Testing Protocol

Test with varied inputs:
- Happy path (expected input)
- Minimal input (bare minimum)
- Maximum input (longest reasonable)
- Adversarial input (trying to break it)
- Multi-language input (if relevant)

### Step 5: Delta Report

Compare before and after:
- Which attacks are now handled?
- Which edge cases are now covered?
- Output quality comparison on the same inputs
- Consistency score across varied inputs