Prompt Engineer

Prompt Optimizer

Optimizes system prompts by applying research-backed prompt engineering patterns. Human-in-the-loop phases: understand, plan, propose changes, receive approval, then integrate.

Purpose and Success Criteria

A well-optimized prompt achieves:

1. Behavioral clarity: Agent knows exactly what to do in common cases and edge cases
2. Appropriate scope: Complex tasks get decomposition; simple tasks don't trigger overthinking
3. Grounded changes: Every modification traces to a specific pattern with documented impact

Optimization is complete when:

• Every change has explicit pattern attribution from the reference document
• No section contradicts another section
• The prompt matches its operating context (tool-use vs. conversational, token constraints)
• Human has approved both section-level changes and full integration

Triggers

When to Use This Skill

Use when the user provides a prompt and wants it improved, refined, or reviewed for best practices.

Do NOT use for:

• Writing prompts from scratch (different skill)
• Prompts that are already working well and user just wants validation
• Non-prompt content (documentation, code, etc.)

Required Resources

Before ANY analysis, read the appropriate pattern reference(s):

Single-Turn Reference (Always Read)

Read references/prompt-engineering-single-turn.md

Contains: Technique Selection Guide table, Quick Reference principles, domain-organized techniques with citations, Anti-Patterns section.

Multi-Turn Reference (Conditional)

Read references/prompt-engineering-multi-turn.md

Read ONLY when the prompt involves:

• Multi-turn flows (iterative refinement, conversation chains)
• Multi-agent / sub-agent orchestration

Skip for:

• Static system prompts executed in a single LLM call
• Tool instructions or one-shot prompts

Workflow Reference

Read references/workflow.md

Contains: Detailed Phase 0-4 workflows, visual card template, completion checkpoint.

Process

┌─────────────────────────────────────────────────────────────────┐
│ 1. READ THE REFERENCE(S)                                        │
│    - Always: references/prompt-engineering-single-turn.md       │
│    - If multi-turn/multi-agent: also read multi-turn reference  │
├─────────────────────────────────────────────────────────────────┤
│ 2. UNDERSTAND THE PROMPT (Phase 1)                              │
│    - Operating context (single-shot? tool-use? constraints?)    │
│    - Current state (working? unclear? missing?)                 │
│    - Document specific problems with quoted prompt text         │
├─────────────────────────────────────────────────────────────────┤
│ 3. PLAN WITH VISUAL CARDS (Phase 2)                             │
│    - Present each change as a visual card with:                 │
│      SCOPE → PROBLEM → TECHNIQUE → BEFORE/AFTER                 │
│    - Quote trigger conditions from reference                    │
│    - ⚠️  WAIT FOR USER APPROVAL before proceeding               │
├─────────────────────────────────────────────────────────────────┤
│ 4. EXECUTE APPROVED CHANGES (Phase 3)                           │
│    - Apply the BEFORE → AFTER transformations                   │
├─────────────────────────────────────────────────────────────────┤
│ 5. INTEGRATE AND VERIFY QUALITY (Phase 4)                       │
│    - Check cross-section coherence                              │
│    - Final anti-pattern check                                   │
│    - Present complete optimized prompt                          │
└─────────────────────────────────────────────────────────────────┘

Triage (Phase 0)

Simple prompts (use lightweight process):

• Under 20 lines
• Single clear purpose
• No conditional logic

Complex prompts (use full process):

• Multiple sections serving different functions
• Conditional behaviors or rule hierarchies
• Tool orchestration or multi-step workflows

Core Quality Principles

1. Quote before deciding: Every technique selection must quote the reference's trigger condition.
2. Open verification questions: Ask "What behavior will this produce?" not "Is this correct?"
3. Approval happens once, upfront: The visual card format in Phase 2 shows full impact.
4. Preserve what works: Optimization means improving problems, not rewriting everything.

Completion Checkpoint

Before presenting the final prompt, verify:

• Phase 2 plan used visual card format with BEFORE/AFTER
• Phase 2 plan quoted trigger conditions from reference
• Phase 2 plan was approved by user before Phase 3
• No technique applied without matching trigger condition
• Stacking compatibility checked; no conflicts
• Anti-patterns section consulted; none introduced
• Emphasis markers used sparingly (≤3 highest-level)

Anti-Patterns

Verification

After optimization:

• Reconciliation: paste the count of visual cards presented in Phase 2 and confirm each card named its reference trigger condition (not a claim: list the card titles here)
• No section contradicts another section
• User approved Phase 2 plan before Phase 3 execution (paste the approval message or turn reference)
• Anti-patterns section consulted, none introduced
• Reconciliation: count emphasis markers (** or __) in the optimized prompt; paste the count -- must be <= 3

References

• prompt-engineering-single-turn.md - Single-turn patterns
• prompt-engineering-multi-turn.md - Multi-turn patterns
• workflow.md - Detailed phase workflows and card template
• prompting-anti-patterns.md - Anti-pattern catalog with effective replacements
• prompting-ai-detection-cliff.md - AI detection cliff and why more rules increase detectability