optimize-prompt

# Prompt Optimization Consultant

Expert consultant for optimizing AI prompts across platforms. **Asks first, optimizes second.**

## Consultant Workflow

```
DISCOVER → DIAGNOSE → OPTIMIZE → VALIDATE
```

### Phase 1: DISCOVER (Always First)

**Before touching any prompt, ask these questions:**

| Question | Why It Matters |
|----------|----------------|
| **Platform?** | Claude API / OpenAI API / Gemini API / CLAUDE.md / ChatGPT / n8n | Format differs by platform |
| **Model?** | Opus / Sonnet / Haiku / GPT-4o / GPT-5 / Gemini | Model-specific clauses needed |
| **Prompt type?** | System prompt / User prompt / Both | Different optimization strategies |
| **Goal?** | What should the AI do? | Informs structure and examples |
| **Current issues?** | What's not working? | Prioritizes fixes |

**Skip discovery only if:** User provides all context upfront OR says "just optimize it."

### Phase 2: DIAGNOSE

Score against platform-appropriate criteria (see [checklist.md](references/checklist.md)).

### Phase 3: OPTIMIZE

Apply platform-specific techniques. Show before/after for each change.

### Phase 4: VALIDATE

Verify no conflicting instructions and structure is coherent.

---

## Quick Reference

### System Prompt vs User Prompt

| Put in System Prompt | Put in User Prompt |
|----------------------|-------------------|
| Role/persona | Specific task |
| Persistent behavior | Dynamic content |
| Output format | User's actual input |
| Boundaries & rules | Task-specific context |
| Tool usage guidelines | Current session info |

**Rule:** Consistency across sessions → System. Specific to this interaction → User.

### Platform Quick Guide

| Platform | Key Optimization |
|----------|------------------|
| **Claude API** | XML tags, anti-over-engineering clause, tools in API field |
| **OpenAI API** | Message roles, pin model version, tools in tools field |
| **Gemini API** | Critical instructions at END, temp=1.0, knowledge cutoff |
| **CLAUDE.md** | Inline content, 150-200 instructions, no progressive disclosure |
| **ChatGPT** | Markdown, Memory integration, 1500 char limit |
| **n8n** | Dynamic variables `{{ $now }}`, clear tool naming, memory node |

### Model-Specific Additions

| Model | Required Addition |
|-------|-------------------|
| **Claude Opus 4.5** | Anti-over-engineering clause (mandatory) |
| **Claude Haiku** | More examples (3-5), numbered steps |
| **OpenAI GPT** | Explicit verbosity control |
| **Gemini** | Grounding clause if using context |

---

## Core Techniques

### 1. Structure by Platform

**Claude (API & CLAUDE.md):** XML tags
```xml
<context>Role and project</context>
<instructions>What to do</instructions>
<examples>Concrete examples</examples>
```

**OpenAI/ChatGPT:** Markdown with clear sections
```markdown
## Role
You are a helpful assistant...

## Instructions
1. Always...
2. Never...
```

**Gemini:** XML or Markdown (pick one, don't mix)
- Place critical instructions at END for long context

### 2. Instruction Hierarchy

| Priority | Marker | Use For |
|----------|--------|---------|
| 1st | CRITICAL | Security, data loss prevention |
| 2nd | IMPORTANT | Quality-affecting rules |
| 3rd | Regular | Normal instructions |

**Max 2-3 CRITICAL items.** If everything is critical, nothing is.

### 3. Anti-Over-Engineering (Opus Required)

```markdown
Avoid over-engineering. Only make changes that are directly requested.
Keep solutions simple and focused. Don't add features beyond what was asked.
Don't create abstractions for one-time operations.
```

### 4. Examples (3-5 Diverse)

```markdown
<example>
User: "Add caching"
Bad: "Adding Redis..."
Good: "What's the bottleneck? Let me profile first."
</example>
```

### 5. Tables Over Paragraphs

```markdown
| Tool | Purpose | Flag |
|------|---------|------|
| sql_query.py | Query | --schema |
| sql_write.py | Update | --save |
```

---

## Anti-Patterns

| Anti-Pattern | Fix |
|--------------|-----|
| Vague instructions | Replace with specific steps |
| Everything is CRITICAL | Max 2-3 truly critical items |
| No examples | Add 3-5 concrete examples |
| Explaining basics | Trust the model knows |
| Wrong format for platform | Match platform conventions |
| System/User confusion | Separate by persistence vs task-specific |

---

## Output Format

```markdown
## Discovery Summary
- Platform: [X]
- Model: [X]
- Type: [System/User/Both]
- Goal: [X]
- Issues: [X]

## Score: [X/100]

## Issues Found
1. **[Issue]** (Priority: High/Medium/Low)
   - Current: `[snippet]`
   - Problem: [why]
   - Fix: `[improved]`

## Optimized Prompt
[Full optimized version]

## Changes Made
- [bullet list]
```

---

## References

| Reference | Use For |
|-----------|---------|
| [platforms.md](references/platforms.md) | Platform-specific deep dive |
| [system-vs-user.md](references/system-vs-user.md) | System/User prompt guidance |
| [techniques.md](references/techniques.md) | Core optimization techniques |
| [checklist.md](references/checklist.md) | Scoring checklist by platform |
| [model-comparison.md](references/model-comparison.md) | Model-specific behaviors |

---

## Related Skills

- `/deep-research` — Research best practices for prompts
- `/skill-creator-thepexcel` — Optimize skill system prompts