Prompt Engineering

Prompt Engineering

When to Load This Skill

Load when working with: system prompts, few-shot examples, chain-of-thought, prompt templates, system_prompt, few_shot, prompt evaluation, "промпт", CoT, output formatting, structured output.

System Prompt Structure

A production system prompt has four sections:

## Role
Who Claude is in this context.

## Context
What the system is, what data Claude has access to, key constraints.

## Task
What Claude must do. Be specific — avoid "help the user".

## Output Format
Exact format of the response. Include examples.

Keep system prompts under 2000 tokens. Longer prompts reduce Claude's effective context for user content.

Few-Shot Examples

Place examples AFTER the main instruction, not before. Three examples outperform one. Use diverse examples that cover edge cases.

SYSTEM_PROMPT = """
You are a document classifier. Classify the document into one of: invoice, contract, receipt, other.

Examples:
Document: "Payment due by 30 days for services rendered..."
Label: invoice

Document: "This agreement is made between Party A and Party B..."
Label: contract

Document: "Thank you for your purchase. Total: $42.00"
Label: receipt
"""

Chain-of-Thought

Use explicit CoT for reasoning tasks. Add "Think step by step" or "First reason through the problem, then give your answer."

SYSTEM_PROMPT = """
You are a data analyst. When answering questions about data:
1. First, identify what the question is asking
2. List the relevant data points
3. Reason through the calculation
4. State your final answer

Do not skip directly to the answer.
"""

Do NOT use CoT for classification or extraction tasks — it wastes tokens without improving accuracy.

Structured Output

For JSON output, provide the exact schema in the system prompt:

SYSTEM_PROMPT = """
Extract the following fields from the document. Return ONLY valid JSON, no other text:

{
  "date": "YYYY-MM-DD or null",
  "amount": float or null,
  "vendor": "string or null",
  "category": "invoice|receipt|contract|other"
}
"""

Validate the output with Pydantic:

from pydantic import BaseModel

class DocumentExtraction(BaseModel):
    date: str | None
    amount: float | None
    vendor: str | None
    category: str

def extract(text: str) -> DocumentExtraction:
    response = client.messages.create(
        model="claude-haiku-4-5-20251001",
        max_tokens=256,
        system=EXTRACTION_PROMPT,
        messages=[{"role": "user", "content": text}],
    )
    return DocumentExtraction.model_validate_json(response.content[0].text)

Prompt Versioning

Store prompts in files, not inline strings:

prompts/
├── v1/
│   ├── system.md
│   └── examples.md
└── v2/
    ├── system.md
    └── examples.md

Load at startup, cache in memory:

from pathlib import Path
from functools import lru_cache

@lru_cache(maxsize=None)
def load_prompt(name: str, version: str = "v1") -> str:
    path = Path("prompts") / version / f"{name}.md"
    return path.read_text(encoding="utf-8")

Common Failure Modes

• Instruction following failure: too many rules in one prompt → split into sub-tasks
• Format drift: Claude adds explanation around JSON → add "Return ONLY..." + validate
• Role confusion: Claude breaks character → strengthen the Role section
• Reasoning shortcut: Claude skips steps → explicitly require intermediate steps

See Also

• resources/structured-output.md — advanced output validation patterns
• resources/evals-framework.md — eval framework for prompt regression testing