How to use the Prompt Engineering Patterns agent skill

What does the skill make the agent do?

Agent skills are folders of instructions that an AI agent loads when a task matches them. This one is about prompt writing. When it is active, the agent can design prompt templates, add worked examples (few-shot learning), ask for step-by-step reasoning (chain-of-thought), request structured JSON output, design system prompts, and help diagnose why a prompt gives inconsistent answers.

The skill also ships a sample template library covering classification and extraction prompts, and a Python script, scripts/optimize-prompt.py, that can A/B test prompt variants against a test suite and track accuracy, latency and token use.

• Few-shot learning: choosing and writing input-output examples
• Chain-of-thought: zero-shot and few-shot reasoning prompts
• Structured outputs: JSON mode and Pydantic schema enforcement
• Prompt optimization: iterative refinement and A/B testing
• Template systems and system prompt design

How do you install the skill?

Download the skill ZIP from this page and unzip it. For personal use across projects, place the unzipped folder in ~/.claude/skills/. To limit it to one project, place it in that project's .claude/skills/ folder instead.

If you use the Claude apps, you can upload the skill instead. Open Settings, go to Capabilities, then Skills, and upload the skill there.

• Personal install: ~/.claude/skills/
• Project install: .claude/skills/ inside your project
• Claude apps: Settings → Capabilities → Skills

How do you trigger the skill with a natural request?

You do not need to name the skill. The skill description lists phrases such as "optimize a prompt", "improve prompt performance", "design a prompt template", "write better prompts", "debug prompt issues", "use chain-of-thought" and "few-shot prompting". Asking in those terms, or describing a prompt that works sometimes but not reliably, should bring the skill into play.

What are good example requests?

Requests work best when they include the task, the failure you are seeing, and the output format you need. The examples below show the kind of request the skill is written to handle.

• "Write a prompt that extracts name, date and amount from invoices as JSON, and show me how to validate it with Pydantic."
• "My sentiment prompt gives different labels for the same review. Debug it and suggest fixes."
• "Use few-shot examples to improve this classification prompt, and pick examples that cover edge cases."
• "Design a reusable template with variables for product name and tone, plus a system prompt that limits answers to our support topics."
• "Compare two versions of this prompt against my test cases and report accuracy, latency and token use."

What do you need before using it?

The skill itself is instructions and templates. Running its optimization script requires more. The script at scripts/optimize-prompt.py uses Python 3 with numpy, and expects an LLM client object with a complete() method. The Quick Start example assumes langchain-anthropic, langchain-core and pydantic, plus an Anthropic API key, which is not included.

Using the prompt guidance alone does not require running any script, but the source notes that the skill runs scripts and uses the network.

• Python 3 with numpy (for scripts/optimize-prompt.py)
• An LLM client object with a complete() method for the optimizer
• For the Quick Start: langchain-anthropic, langchain-core, pydantic
• An Anthropic API key (not included)

What should you check before running the scripts?

The optimization script executes code and calls a model endpoint. Review it before running, and do not run it on untrusted test data or against paid API keys without checking what it does. Start with a small, trusted test suite so you can see the cost and behaviour before scaling up.

Treat any sample code as a starting point. Check model names and client setup against the library you have installed.

• Read scripts/optimize-prompt.py before running it
• Use trusted, representative test data
• Be aware that paid API calls will be made

What tips help you get reliable results?

Start with a simple prompt and add patterns only when a measured problem needs them, since the patterns add overhead. Keep examples close to the real task, because mismatched examples confuse the model. Parameterise prompts instead of hardcoding values, and keep prompts under version control so you can compare them.

• Be specific: vague prompts produce inconsistent results
• Use examples that match the target task
• Validate outputs against a schema and handle malformed responses
• Test on edge cases, not only typical inputs
• Track success rate (valid, parseable outputs) alongside accuracy