aselimc Prompt Engineering
Structured Output
import instructor
client = instructor.from_anthropic(anthropic.Anthropic())
result = client.messages.create(
model="claude-sonnet-4-6",
response_model=MyPydanticModel,
messages=[{"role": "user", "content": prompt}],
)
Patterns
- Few-shot: 3-5 examples covering edge cases
- Chain-of-thought: "Think step by step" or explicit reasoning structure
- System prompt: role + constraints + output format
- XML tags: use
<context>,<instructions>,<examples>for structure
Anti-Patterns
- Prompt injection: validate/sanitize user input before embedding in prompts
- Output validation: parse and validate LLM output, retry on failure
- Don't rely on "don't" instructions - tell the model what TO do
Rules
- Template with Jinja2 or f-strings, never string concatenation with user input
- Version control prompts alongside code
- Evaluate: consistency (same input -> same output), accuracy, latency
Key Libraries
anthropic SDK, openai SDK, instructor, outlines, guidance