| name | prompt-engineer |
| version | 1.0.0 |
| description | Prompt engineering for reliable LLM outputs, including task framing, context selection, structured outputs, tool-use prompts, eval-driven iteration, prompt caching, and safe rationale handling. |
| author | skillregistry |
| license | MIT |
| agents | ["cursor","claude-code","gemini-cli"] |
| categories | ["ai-ml"] |
| tags | ["prompting","llm","evals"] |
Prompt Engineer
Design prompts as testable interfaces. Be explicit about task, constraints, inputs, output schema, and failure behavior. Do not request hidden chain-of-thought; ask for concise rationale or decision summaries when useful.
Workflow
- Define the task and success criteria.
- Separate stable instructions from variable user/data context.
- Provide only relevant context and label it clearly.
- Specify output format with JSON Schema/Zod when machine parsing matters.
- Include few-shot examples only when they reduce ambiguity.
- Run prompts against representative eval cases and revise based on failures.
Prompt Template
Role: You are a backend code reviewer.
Task: Identify correctness, security, and test coverage issues.
Rules:
- Lead with findings.
- Cite file and line.
- Do not speculate beyond evidence.
Input:
<diff>
{{diff}}
</diff>
Output:
JSON array of {severity, file, line, issue, recommendation}.
Rules
- Put static instructions and schemas before variable context to improve prompt caching.
- Prefer structured outputs for downstream automation.
- Ask for assumptions and concise reasoning summaries, not private chain-of-thought.
- Keep tool instructions separate from user content.
- Use delimiters for untrusted retrieved/context text.
- Validate model output before acting on it.
- Maintain small eval sets for every production prompt.
Verification
pnpm test
pnpm evals
Track pass/fail cases, latency, token usage, refusal/format failures, and regression examples.
Resources
Principles
- Prompts are contracts.
- Context should be relevant and labeled.
- Structured output needs validation.
- Evals are stronger than vibes.
- Do not expose or require hidden reasoning.
