How to use the Prompt Engineering Expert agent skill

What the prompt-engineer skill makes the agent do

An agent skill is a set of instructions that the agent reads and follows when a task matches it. This one is named prompt-engineer and is described as a prompt engineering expert for chain-of-thought, few-shot learning, evaluation, and LLM optimization. Once it is installed, the agent answers as a prompt engineering specialist.

• Structures instructions as a sequence of clear steps instead of one long instruction.
• Adds two to five concrete examples when the output format or reasoning style is unusual.
• Asks for structured output such as JSON schemas or XML tags so results are easier to parse.
• Suggests temperature and top_p settings, for example a low temperature (0.0 to 0.3) for factual tasks and a higher one (0.7 to 1.0) for creative tasks.
• Recommends automated evaluation and scoring rubrics rather than only judging outputs by feel.
• Warns about common mistakes, such as relying only on negative instructions or assuming a prompt behaves the same on every model.

How to install the prompt-engineer skill

Download the ZIP file for the prompt-engineer skill from this page. The skill is a single file, SKILL.md, so the ZIP contains that file. Unzip it into one of two places, depending on how you want to use it.

For a personal setup available across your projects, unzip it into ~/.claude/skills/. For a setup that applies only to one project, unzip it into the .claude/skills/ folder inside that project.

If you use the Claude apps instead, you can upload the skill there under Settings → Capabilities → Skills.

• Personal install: ~/.claude/skills/
• Project install: .claude/skills/ inside your project folder
• Claude apps: Settings → Capabilities → Skills, then upload the skill

How to trigger the skill with natural requests

You do not need special syntax. Describe the prompt problem in plain language and mention the topic the skill covers, such as chain-of-thought, few-shot examples, evaluation, or output format. Naming the prompt-engineer skill in your request is a clear way to point the agent at it.

Give the agent the real context: the model you target, the input it will receive, and the output you need. The skill can only recommend technique when it knows these details.

• Help me write a prompt that returns a JSON object with fields for name, date and amount.
• Use the prompt-engineer skill to design a few-shot prompt for classifying support tickets.
• How should I set temperature for a prompt that extracts facts from contracts?
• Review this prompt and suggest a rubric I can use to score its outputs.

What to expect from the output

Expect a structured answer: a role and task description, numbered steps, example input-output pairs where the format is unusual, and a suggested output format. The skill also tends to flag risks in your draft prompt, such as ambiguous wording or heavy reliance on 'do not' rules.

Treat the output as a draft to test. The skill gives general techniques. It does not measure how your prompt performs on your model or your data, so run your own tests before relying on it.

What the skill needs, and safety notes

The skill lists no required tools or dependencies. It is text guidance only: it contains no scripts, and it does not call any external service or use the network. Installing it means adding one instruction file to your agent setup.

The skill cannot verify facts or access current information. It provides no sources and no browsing, and the underlying model may still produce incorrect statements. Check any factual output yourself.

• Runs scripts: no
• Uses network: no
• Files included: SKILL.md
• Licence: Apache-2.0

Tips for getting the most from the skill

Give the agent your actual prompt and a few real inputs, not only a description. Ask for one change at a time so you can see what each change does. When you have a working prompt, ask the agent to help you write an evaluation rubric with criteria such as accuracy, completeness, relevance and format compliance.

If your task has consistent, high-volume, domain-specific requirements, ask the agent whether fine-tuning might be a better fit, since the skill notes that fine-tuning can be more cost-effective at scale.

• Name the target model, because a prompt may behave differently across models.
• Replace 'do not do X' with a description of what you want.
• Keep long inputs short enough to leave room for the model's output.
• Prefer the smallest model that gives acceptable quality, since larger models cost more per token.