Prompt Engineer guides Claude through prompt work in five steps: define the task and success criteria, design an initial prompt, test it on diverse cases, iterate one change at a time, then version and document it.
The skill includes zero-shot and few-shot examples, a before and after rewrite of a vague summary prompt, and reference files for patterns, optimization, evaluation, structured outputs, system prompts and context management.
What You Get
- A five-step workflow from requirements to versioned deployment
- Few-shot and zero-shot templates plus a before and after prompt rewrite
- Reference files for prompt patterns, optimization, evaluation frameworks, structured outputs, system prompts and context management
- Rules that require tests on realistic and edge-case inputs, schema validation and testing across model versions
- An output template with the final prompt, test results, usage settings, metrics and known limitations
Common Use Cases
- Writing a system prompt with a persona and guardrails
- Adding few-shot examples to improve classification consistency
- Defining a JSON or function-calling schema and validating outputs
- Cutting token use in an existing prompt
- Migrating prompts between model providers and re-testing them
What Makes It Unique
The skill covers context management as well as wording: attention budget, degradation patterns and lost-in-the-middle effects have their own reference. It also insists on method, with one change at a time while debugging, quantitative metrics, versioned prompts and monitoring for degradation in production.
Under The Hood
The skill is instructions and reference markdown that Claude loads by topic; it ships no scripts. It produces prompt text, schemas and test cases, and does not call model APIs or other outside services itself. Evaluation runs only happen if you or your own code execute the test cases it writes.