Prompt patterns I actually reuse

Most prompt advice is folklore. These three survive contact with production.

1. The "rubric first" pattern

RUBRIC = """
Score 1-5 on: correctness, clarity, completeness.
Respond ONLY with JSON: {"scores": [...], "reason": "..."}
"""

def grade(answer: str) -> dict:
    # judge != generator: use a separate call for scoring
    return llm.chat([
        {"role": "system", "content": RUBRIC},
        {"role": "user", "content": answer},
    ], json_mode=True)

2. Few-shot with negative examples

Showing the model what NOT to do cuts hallucination more than extra positives.

3. Make the model spend tokens

Chain-of-verification beats chain-of-thought for factual work.
Draft, then ask it to attack its own draft, then revise.