fusengine/agents/prompt-optimization

<objective> Prompt Optimization analyzes and improves an existing prompt through a 5-step workflow: analyze the current prompt, identify issues, apply corrections, validate the improvement, and document the changes. Its checklist covers four dimensions -- clarity (unambiguous instructions, precise vocabulary), structure (well-delimited sections, logical order), completeness (output format, error cases, examples), and guardrails (explicit limits, forbidden behaviors, security).

It never changes the prompt's original meaning, adds unrequested features, removes existing guardrails, or lengthens the prompt without justification. See common-problems.md for before/after fix patterns and improvement-techniques.md for strengthening via CoT, few-shot, or guardrails. </objective>

Prompt Optimization

Skill for analyzing and improving existing prompts.

References

• common-problems.md - Load when: diagnosing a vague or ambiguous prompt — 5 before/after examples (vague instructions, missing context, undefined format, no error handling, weak emphasis)
• improvement-techniques.md - Load when: strengthening a prompt with Chain-of-Thought, few-shot examples, or reinforced guardrails
• scorecard-template.md - Load when: writing the before/after optimization report to hand off

Optimization Workflow

1. ANALYZE current prompt
   ↓
2. IDENTIFY issues
   ↓
3. APPLY corrections
   ↓
4. VALIDATE improvement
   ↓
5. DOCUMENT changes

Analysis Checklist

Clarity

• [ ] Unambiguous instructions?
• [ ] Clearly defined objective?
• [ ] Precise vocabulary?

Structure

• [ ] Well-delimited sections?
• [ ] Logical order?
• [ ] Clear hierarchy?

Completeness

• [ ] Output format defined?
• [ ] Error cases handled?
• [ ] Examples if needed?

Guardrails

• [ ] Explicit limits?
• [ ] Forbidden behaviors listed?
• [ ] Appropriate security?

For common problem patterns with before/after fixes, see common-problems.md. For techniques to strengthen a prompt (CoT, examples, guardrails), see improvement-techniques.md.

Forbidden

• Never change the original meaning of the prompt
• Never add unrequested features
• Never remove existing guardrails
• Never make the prompt longer without justification