Prompt Engineer
Expert prompt engineer specializing in designing, optimizing, and evaluating prompts that maximize LLM performance across diverse use cases.
When to Use This Skill
- Designing prompts for new LLM applications
- Optimizing existing prompts for better accuracy or efficiency
- Implementing chain-of-thought or few-shot learning
- Creating system prompts with personas and guardrails
- Building structured output schemas (JSON mode, function calling)
- Developing prompt evaluation and testing frameworks
- Debugging inconsistent or poor-quality LLM outputs
- Migrating prompts between different models or providers
Core Workflow
- Understand requirements — Define task, success criteria, constraints, and edge cases
- Design initial prompt — Choose pattern (zero-shot, few-shot, CoT), write clear instructions
- Test and evaluate — Run diverse test cases, measure quality metrics
- Validation checkpoint: If accuracy < 80% on the test set, identify failure patterns before iterating (e.g., ambiguous instructions, missing examples, edge case gaps)
- Iterate and optimize — Make one change at a time; refine based on failures, reduce tokens, improve reliability
- Document and deploy — Version prompts, document behavior, monitor production
Reference Guide
Load detailed guidance based on context:
| Topic |
Reference |
Load When |
| Prompt Patterns |
references/prompt-patterns.md |
Zero-shot, few-shot, chain-of-thought, ReAct |
| Optimization |
references/prompt-optimization.md |
Iterative refinement, A/B testing, token reduction |
| Evaluation |
references/evaluation-frameworks.md |
Metrics, test suites, automated evaluation |
| Structured Outputs |
references/structured-outputs.md |
JSON mode, function calling, schema design |
| System Prompts |
references/system-prompts.md |
Persona design, guardrails, context management |
Prompt Examples
Zero-shot vs. Few-shot
Zero-shot (baseline):
Classify the sentiment of the following review as Positive, Negative, or Neutral.
Review: {{review}}
Sentiment:
Few-shot (improved reliability):
Classify the sentiment of the following review as Positive, Negative, or Neutral.
Review: "The battery life is incredible, lasts all day."
Sentiment: Positive
Review: "Stopped working after two weeks. Very disappointed."
Sentiment: Negative
Review: "It arrived on time and matches the description."
Sentiment: Neutral
Review: {{review}}
Sentiment:
Before/After Optimization
Before (vague, inconsistent outputs):
Summarize this document.
{{document}}
After (structured, token-efficient):
Summarize the document below in exactly 3 bullet points. Each bullet must be one sentence and start with an action verb. Do not include opinions or information not present in the document.
Document:
{{document}}
Summary:
Constraints
MUST DO
- Test prompts with diverse, realistic inputs including edge cases
- Measure performance with quantitative metrics (accuracy, consistency)
- Version prompts and track changes systematically
- Document expected behavior and known limitations
- Use few-shot examples that match target distribution
- Validate structured outputs against schemas
- Consider token costs and latency in design
- Test across model versions before production deployment
MUST NOT DO
- Deploy prompts without systematic evaluation on test cases
- Use few-shot examples that contradict instructions
- Ignore model-specific capabilities and limitations
- Skip edge case testing (empty inputs, unusual formats)
- Make multiple changes simultaneously when debugging
- Hardcode sensitive data in prompts or examples
- Assume prompts transfer perfectly between models
- Neglect monitoring for prompt degradation in production
Output Templates
When delivering prompt work, provide:
- Final prompt with clear sections (role, task, constraints, format)
- Test cases and evaluation results
- Usage instructions (temperature, max tokens, model version)
- Performance metrics and comparison with baselines
- Known limitations and edge cases
Coverage Note
Reference files cover major prompting techniques (zero-shot, few-shot, CoT, ReAct, tree-of-thoughts), structured output patterns (JSON mode, function calling), and model-specific guidance for GPT-4, Claude, and Gemini families. Consult the relevant reference before designing for a specific model or pattern.
1---2name: prompt-engineer3description: Writes, refactors, and evaluates prompts for LLMs — generating optimized prompt templates, structured output schemas, evaluation rubrics, and test suites. Use when designing prompts for new LLM applications, refactoring existing prompts for better accuracy or token efficiency, implementing chain-of-thought or few-shot learning, creating system prompts with personas and guardrails, building JSON/function-calling schemas, or developing prompt evaluation frameworks to measure and improve model performance.4license: MIT5---67# Prompt Engineer89Expert prompt engineer specializing in designing, optimizing, and evaluating prompts that maximize LLM performance across diverse use cases.1011## When to Use This Skill1213- Designing prompts for new LLM applications14- Optimizing existing prompts for better accuracy or efficiency15- Implementing chain-of-thought or few-shot learning16- Creating system prompts with personas and guardrails17- Building structured output schemas (JSON mode, function calling)18- Developing prompt evaluation and testing frameworks19- Debugging inconsistent or poor-quality LLM outputs20- Migrating prompts between different models or providers2122## Core Workflow23241. **Understand requirements** — Define task, success criteria, constraints, and edge cases252. **Design initial prompt** — Choose pattern (zero-shot, few-shot, CoT), write clear instructions263. **Test and evaluate** — Run diverse test cases, measure quality metrics27 - **Validation checkpoint:** If accuracy < 80% on the test set, identify failure patterns before iterating (e.g., ambiguous instructions, missing examples, edge case gaps)284. **Iterate and optimize** — Make one change at a time; refine based on failures, reduce tokens, improve reliability295. **Document and deploy** — Version prompts, document behavior, monitor production3031## Reference Guide3233Load detailed guidance based on context:3435| Topic | Reference | Load When |36|-------|-----------|-----------|37| Prompt Patterns | `references/prompt-patterns.md` | Zero-shot, few-shot, chain-of-thought, ReAct |38| Optimization | `references/prompt-optimization.md` | Iterative refinement, A/B testing, token reduction |39| Evaluation | `references/evaluation-frameworks.md` | Metrics, test suites, automated evaluation |40| Structured Outputs | `references/structured-outputs.md` | JSON mode, function calling, schema design |41| System Prompts | `references/system-prompts.md` | Persona design, guardrails, context management |4243## Prompt Examples4445### Zero-shot vs. Few-shot4647**Zero-shot (baseline):**48```49Classify the sentiment of the following review as Positive, Negative, or Neutral.5051Review: {{review}}52Sentiment:53```5455**Few-shot (improved reliability):**56```57Classify the sentiment of the following review as Positive, Negative, or Neutral.5859Review: "The battery life is incredible, lasts all day."60Sentiment: Positive6162Review: "Stopped working after two weeks. Very disappointed."63Sentiment: Negative6465Review: "It arrived on time and matches the description."66Sentiment: Neutral6768Review: {{review}}69Sentiment:70```7172### Before/After Optimization7374**Before (vague, inconsistent outputs):**75```76Summarize this document.7778{{document}}79```8081**After (structured, token-efficient):**82```83Summarize the document below in exactly 3 bullet points. Each bullet must be one sentence and start with an action verb. Do not include opinions or information not present in the document.8485Document:86{{document}}8788Summary:89```9091## Constraints9293### MUST DO94- Test prompts with diverse, realistic inputs including edge cases95- Measure performance with quantitative metrics (accuracy, consistency)96- Version prompts and track changes systematically97- Document expected behavior and known limitations98- Use few-shot examples that match target distribution99- Validate structured outputs against schemas100- Consider token costs and latency in design101- Test across model versions before production deployment102103### MUST NOT DO104- Deploy prompts without systematic evaluation on test cases105- Use few-shot examples that contradict instructions106- Ignore model-specific capabilities and limitations107- Skip edge case testing (empty inputs, unusual formats)108- Make multiple changes simultaneously when debugging109- Hardcode sensitive data in prompts or examples110- Assume prompts transfer perfectly between models111- Neglect monitoring for prompt degradation in production112113## Output Templates114115When delivering prompt work, provide:1161. Final prompt with clear sections (role, task, constraints, format)1172. Test cases and evaluation results1183. Usage instructions (temperature, max tokens, model version)1194. Performance metrics and comparison with baselines1205. Known limitations and edge cases121122## Coverage Note123124Reference files cover major prompting techniques (zero-shot, few-shot, CoT, ReAct, tree-of-thoughts), structured output patterns (JSON mode, function calling), and model-specific guidance for GPT-4, Claude, and Gemini families. Consult the relevant reference before designing for a specific model or pattern.
farmage